Caroline Mimbs Nyce | The Atlantic

It’s Never Too Late to Learn an Instrument

2024-12-01T09:30:00-05:00

The recorder used to be an instrument people wanted to hear. As a 1946 article in The Atlantic explained, it gets mentioned lovingly in Shakespeare’s Hamlet and Milton’s Paradise Lost. One 17th-century English-navy leader wrote in his diary that it made the best sound he’d ever heard. The recorder was the instrument of kings and queens: Henry VIII had a collection of more than 70.

But by 1946, recorders were already commonly associated with terrible screeching noises, most often made by children. And today, few adults play them. In fact, they don’t really play instruments at all—certainly not recorders, but not piccolos or pianos either. A 2022 survey run by the National Endowment for the Arts found that about 11 percent of American adults play a musical instrument. Kids receive plenty of music education, but as people get older, they fall out of practice. Many stop picking up their instrument.

This is unfortunate, in part because plenty of research shows that adults could benefit from playing music. Doing so is neuroprotective. “It helps you build up larger brain networks and new pathways,” Daniel Levitin, the author of the recent book I Heard There Was a Secret Chord: Music as Medicine, told me. You build these pathways by listening to music, he told me, but physically playing an instrument strengthens motor pathways as well: “You’re building up a lot of brain capacity.” Musicians tend to have better attention than nonmusicians. Banging on a drum or tooting a horn can also relieve stress, reduce burnout, and help with anxiety and depression. For older people specifically, research has shown potential cognitive benefits along with a possible decrease in dementia risk.

So why aren’t more adults doing it? Part of the issue may indeed be that music education is associated with childhood and coursework. “When it looks like a school-based thing, it looks like something you age out of,” Mandi Schlegel, a music-education professor at the University of South Carolina, told me. (Another way of thinking about this: Have you ever done a math worksheet for fun?) Anybody can make music, but that notion is “taught out of us,” Michael Spitzer, a British musicologist and the author of The Musical Human: A History of Life on Earth, told me. After people grow out of music education in their childhood, they tend to think “that music is a special talent,” he said. When I tell people I play the flute, I always caution them that I do not play like Lizzo. Even though I’m very proud of my tone, I acknowledge that most people may not want to spend a lot of time listening to an intermediate flautist.

Of course, people are busy; they simply may not have the luxury of sitting down to study Bach once a week, much less the money to pay for an instrument or private lessons. Once you’ve gotten over those humps, finding others who have done the same is another challenge: It’s easy to go to a park or gym and pull together a game of pickup basketball, but piecing together people at the same skill level to play a concerto or even just jam in a garage is another matter. Few adults play musical instruments at all, and even fewer do so in a group. Which is a shame, because research has shown that playing music together has additional benefits: “We become more trusting; we feel more connected to others—maybe even more connected to the world at large,” Levitin said.

Still, I can attest that it’s worth the effort to pick up an instrument, even if you can’t do so in a group. Once a week, you can find me at my local music school, waiting in the lobby alongside the cool kids of Los Angeles. One night this past summer, my flute teacher, Derrick, surprised me with a recorder. It was partially a gag gift—I’d sent him a viral video of a man playing a recorder while surfing, and pledged to learn to do the same. But it was also serious: Derrick teaches the recorder as well, and I was touched by the gesture.

We spent most of that lesson laughing—I would get a few notes out, then burst out giggling mid-song. The recorder can sound positively medieval, like you’re suddenly in a Game of Thrones episode when you’re really just playing “Old MacDonald Had a Farm.” But I learned that it can also be lovely, and that it’s a whole lot of fun to play. I plan to keep learning, not because it strengthens my neuropathways per se (though I certainly don’t mind that), but because making music, even when it’s silly—perhaps especially when it’s silly—is just a whole lot of fun.

The Gateway Pundit Is Still Pushing an Alternate Reality

2024-11-02T15:06:39-04:00

The Gateway Pundit, a right-wing website with a history of spreading lies about election fraud, recently posted something out of the ordinary. It took a break from its coverage of the 2024 presidential election (sample headlines: “KAMALA IS KOLLAPSING,” “KAMALA FUNDS NAZIS”) to post a three-sentence note from the site’s founder and editor, Jim Hoft, offering some factual information about the previous presidential election.

In his brief statement, presented without any particular fanfare, Hoft writes that election officials in Georgia concluded that no widespread voter fraud took place at Atlanta’s State Farm Arena on Election Day 2020. He notes specifically that they concluded that two election workers processing votes that night, Ruby Freeman and Wandrea Moss, had not engaged “in ballot fraud or criminal misconduct.” And he explains that “a legal matter with this news organization and the two election workers has been resolved to the mutual satisfaction of the parties through a fair and reasonable settlement.”

Indeed, the blog post appeared just days after the Gateway Pundit settled a defamation lawsuit brought by Freeman and Moss, who sued the outlet for promoting false claims that they had participated in mass voter fraud. (These claims, quickly debunked, were focused on video footage of the mother-daughter pair storing ballots in their appropriate carriers—conspiracy theorists had claimed that they were instead packing them into suitcases for some wicked purpose.) The terms of the settlement were not disclosed, but after it was announced, almost 70 articles previously published on the Gateway Pundit, and cited in the lawsuit, were no longer available, according to an analysis by the Associated Press.

Even so, the site—which has promoted numerous lies and conspiracy theories in the past, and which still faces a lawsuit from Eric Coomer, a former executive at Dominion Voting Systems, for pushing false claims that he helped rig the 2020 election—shows no signs of retreat. (The Gateway Pundit has fought this lawsuit, including by filing a motion to dismiss. Although the site filed for bankruptcy in April, a judge tossed it out, concluding that the filing was in “bad faith.”) The site has continued to post with impunity, promoting on a number of occasions the conspiracy that Democrats are “openly stealing” the 2024 election with fraudulent overseas votes. A political-science professor recently told my colleague Matteo Wong that this particular claim has been one of the “dominant narratives” this year, as Donald Trump’s supporters seek ways to undermine faith in the democratic process.

This is to be expected: The Gateway Pundit has been around since 2004, and it has always been a destination for those disaffected by the “establishment media.” Comment sections—on any website, let alone those that explicitly cater to the far-right fringe—have never had a reputation for sobriety and thoughtfulness. And the Gateway Pundit’s is particularly vivid. One recent commenter described a desire to see Democratic officials “stripped naked and sprayed down with a firehose like Rambo in First Blood.” Even so, data recently shared with me by the Center for Countering Digital Hate—a nonprofit that studies disinformation and online abuse, and which reports on companies that it believes allow such content to spread—show just how nasty these communities can get. Despite the fracturing of online ecosystems in recent years—namely, the rise and fall of various social platforms and the restructuring of Google Search, both of which have resulted in an overall downturn in traffic to news sites—the Gateway Pundit has remained strikingly relevant on social media, according to the CCDH. And its user base, as seen in the comments, has regularly endorsed political violence in the past few months, despite the site’s own policies forbidding such posts.

Researchers from the CCDH recently examined the comment sections beneath 120 Gateway Pundit articles about alleged election fraud published between May and September. They found that 75 percent of those sections contained “threats or calls for violence.” One comment cited in the report reads: “Beat the hell out of any Democrat you come across today just for the hell of it.”

Another: “They could show/televise the hangings or lined up and executed by firing squad and have that be a reminder not to try to overthrow our constitution.” Overall, the researchers found more than 200 comments with violent content hosted on the Gateway Pundit.

Sites like the Gateway Pundit often attempt to justify the vitriol they host on their platforms by arguing in free-speech terms. But even free-speech absolutists can understand legitimate concerns about incitements to violence. Local election officials in Georgia and Arizona have blamed the site and its comment section for election-violence threats in the past. A 2021 Reuters report found links between the site and more than 80 “menacing” messages sent to election workers. According to Reuters, after the Gateway Pundit published a fake report about ballot fraud in Wisconsin, one election official found herself identified in the comment section, along with calls for her to be killed. “She found one post especially unnerving,” the Reuters reporters Peter Eisler and Jason Szep write. “It recommended a specific bullet for killing her—a 7.62 millimeter round for an AK-47 assault rifle.”

The CCDH researchers used data from a social-media monitoring tool called Newswhip to measure social-media engagement with election-related content from Gateway Pundit and similar sites. Although Gateway Pundit was second to Breitbart as a source for election misinformation on social media overall, the researchers found that the Gateway Pundit was actually the most popular on X, where its content was shared more than 800,000 times from the start of the year through October 2.

In response to a request for comment, John Burns, a lawyer representing Hoft and the Gateway Pundit, told me that the site relies on users reporting “offending” comments, including those expressing violence or threats. “If a few slipped through the cracks, we’ll look into it,” Burns said. He did not comment on the specifics of the CCDH report, nor the recent lawsuits against the company.

The site uses a popular third-party commenting platform called Disqus, which has taken a hands-off approach to policing far-right, racist content in the past. Disqus offers clients AI-powered, customizable moderation tools that allow them to filter out toxic or inappropriate comments from their site, or ban users. The CCDH report points out that violent comments are against Disqus’s own terms of service. “Publishers monitor and enforce their own community rules,” a Disqus spokesperson wrote in an email statement. “Only if a comment is flagged directly to the Disqus team do we review it against our terms of service. Once flagged, we aim to review within 24 hours and determine whether or not action is required based on our rules and terms of service.”

The Gateway Pundit is just one of a constellation of right-wing sites that offer readers an alternate reality. Emily Bell, the founding director of the Tow Center for Digital Journalism, told me that these sites pushed the range of what’s considered acceptable speech “quite a long way to the right,” and in some cases, away from traditional, “fact-based” media. They started to grow more popular with the rise of the social web, in which algorithmic recommendation systems and conservative influencers pushed their articles to legions of users.

The real power of these sites may come not in their broad reach, but in how they shape the opinions of a relatively small, radical subset of people. According to a paper published in Nature this summer, false and inflammatory content tends to reach “a narrow fringe” of highly motivated users. Sites like the Gateway Pundit are “influential in a very small niche,” Brendan Nyhan, a professor of government at Dartmouth and one of the authors of the paper, told me over email. As my colleague Charlie Warzel recently noted, the effect of this disinformation is not necessarily to deceive people, but rather to help this small subset of people stay anchored in their alternate reality.

I asked Pasha Dashtgard, the director of research for the Polarization and Extremism Research and Innovation Lab at American University, what exactly the relationship is between sites like Gateway Pundit and political violence. “That is such a million-dollar question,” he said. “It’s hard to tell.” By that, he means that it’s hard for researchers and law enforcement to know when online threats will translate into armed vigilantes descending on government buildings. Social-media platforms have only gotten less transparent with their data since the previous cycle, making it more difficult for researchers to suss out what’s happening on them.

“The pathway to radicalization is not linear,” Dashtgard explained. “Certainly I would want to disabuse anyone of the idea that it’s like, you go on this website and that makes you want to kill people.” People could have other risk factors that make them more likely to commit violence, such as feeling alienated or depressed, he said. These sites just represent another potential push mechanism.

And they don’t seem to be slowing down. Three hours after Hoft posted his blog post correcting the record in the case of Freeman and Moss, he posted another statement. This one was addressed to readers. “Many of you may be aware that The Gateway Pundit was in the news this week. We settled an ongoing lawsuit against us,” the post reads in part. “Despite their best efforts, we are still standing.”

The Age of AI Child Abuse Is Here

2024-10-18T13:32:30-04:00

Muah.AI is a website where people can make AI girlfriends—chatbots that will talk via text or voice and send images of themselves by request. Nearly 2 million users have registered for the service, which describes its technology as “uncensored.” And, judging by data purportedly lifted from the site, people may be using its tools in their attempts to create child-sexual-abuse material, or CSAM.

Last week, Joseph Cox, at 404 Media, was the first to report on the data set, after an anonymous hacker brought it to his attention. What Cox found was profoundly disturbing: He reviewed one prompt that included language about orgies involving “newborn babies” and “young kids.” This indicates that a user had asked Muah.AI to respond to such scenarios, although whether the program did so is unclear. Major AI platforms, including ChatGPT, employ filters and other moderation tools intended to block generation of content in response to such prompts, but less prominent services tend to have fewer scruples.

People have used AI software to generate sexually exploitative images of real individuals. Earlier this year, pornographic deepfakes of Taylor Swift circulated on X and Facebook. And child-safety advocates have warned repeatedly that generative AI is now being widely used to create sexually abusive imagery of real children, a problem that has surfaced in schools across the country.

The Muah.AI hack is one of the clearest—and most public—illustrations of the broader issue yet: For maybe the first time, the scale of the problem is being demonstrated in very clear terms.

I spoke with Troy Hunt, a well-known security consultant and the creator of the data-breach-tracking site HaveIBeenPwned.com, after seeing a thread he posted on X about the hack. Hunt had also been sent the Muah.AI data by an anonymous source: In reviewing it, he found many examples of users prompting the program for child-sexual-abuse material. When he searched the data for 13-year-old, he received more than 30,000 result s, “many alongside prompts describing sex acts.” When he tried prepubescent, he got 26,000 result s. He estimates that there are tens of thousands, if not hundreds of thousands, of prompts to create CSAM within the data set.

Hunt was surprised to find that some Muah.AI users didn’t even try to conceal their identity. In one case, he matched an email address from the breach to a LinkedIn profile belonging to a C-suite executive at a “very normal” company. “I looked at his email address, and it’s literally, like, his first name dot last name at gmail.com,” Hunt told me. “There are lots of cases where people make an attempt to obfuscate their identity, and if you can pull the right strings, you’ll figure out who they are. But this guy just didn’t even try.” Hunt said that CSAM is traditionally associated with fringe corners of the internet. “The fact that this is sitting on a mainstream website is what probably surprised me a little bit more.”

Last Friday, I reached out to Muah.AI to ask about the hack. A person who runs the company’s Discord server and goes by the name Harvard Han confirmed to me that the website had been breached by a hacker. I asked him about Hunt’s estimate that as many as hundreds of thousands of prompts to create CSAM may be in the data set. “That’s impossible,” he told me. “How is that possible? Think about it. We have 2 million users. There’s no way 5 percent is fucking pedophiles.” (It is possible, though, that a relatively small number of users are responsible for a large number of prompts.)

When I asked him whether the data Hunt has are real, he initially said, “Maybe it is possible. I am not denying.” But later in the same conversation, he said that he wasn’t sure. Han said that he had been traveling, but that his team would look into it.

The site’s staff is small, Han stressed over and over, and has limited resources to monitor what users are doing. Fewer than five people work there, he told me. But the site seems to have built a modest user base: Data provided to me from Similarweb, a traffic-analytics company, suggest that Muah.AI has averaged 1.2 million visits a month over the past year or so.

Han told me that last year, his team put a filtering system in place that automatically blocked accounts using certain words—such as teenagers and children—in their prompts. But, he told me, users complained that they were being banned unfairly. After that, the site adjusted the filter to stop automatically blocking accounts, but to still prevent images from being generated based on those keywords, he said.

At the same time, however, Han told me that his team does not check whether his company is generating child-sexual-abuse images for its users. He assumes that a lot of the requests to do so are “probably denied, denied, denied,” he said. But Han acknowledged that savvy users could likely find ways to bypass the filters.

He also offered a kind of justification for why users might be trying to generate images depicting children in the first place: Some Muah.AI users who are grieving the deaths of family members come to the service to create AI versions of their lost loved ones. When I pointed out that Hunt, the cybersecurity consultant, had seen the phrase 13-year-old used alongside sexually explicit acts, Han replied, “The problem is that we don’t have the resources to look at every prompt.” (After Cox’s article about Muah.AI, the company said in a post on its Discord that it plans to experiment with new automated methods for banning people.)

In sum, not even the people running Muah.AI know what their service is doing. At one point, Han suggested that Hunt might know more than he did about what’s in the data set. That sites like this one can operate with such little regard for the harm they may be causing raises the bigger question of whether they should exist at all, when there’s so much potential for abuse.

Meanwhile, Han took a familiar argument about censorship in the online age and stretched it to its logical extreme. “I’m American,” he told me. “I believe in freedom of speech. I believe America is different. And we believe that, hey, AI should not be trained with censorship.” He went on: “In America, we can buy a gun. And this gun can be used to protect life, your family, people that you love—or it can be used for mass shooting.”

Federal law prohibits computer-generated images of child pornography when such images feature real children. In 2002, the Supreme Court ruled that a total ban on computer-generated child pornography violated the First Amendment. How exactly existing law will apply to generative AI is an area of act ive debate. When I asked Han about federal laws regarding CSAM, Han said that Muah.AI only provides the AI processing, and compared his service to Google. He also reiterated that his company’s word filter could be blocking some images, though he is not sure.

Whatever happens to Muah.AI, these problems will certainly persist. Hunt told me he’d never even heard of the company before the breach. “And I’m sure that there are dozens and dozens more out there.” Muah.AI just happened to have its contents turned inside out by a data hack. The age of cheap AI-generated child abuse is very much here. What was once hidden in the darkest corners of the internet now seems quite easily accessible—and, equally worrisome, very difficult to stamp out.

What Does That Bark Mean?

2024-10-17T09:00:00-04:00

The first thing I ever said to my dog was, “Do you want to come home with me?” He was six pounds, and 10 weeks old. He craned his head forward and sniffed my mouth.

In the four years since, I have continued to pepper him with questions that he cannot answer. I ask him what he’s up to, if he wants to go for a walk, if he’s feeling sleepy. When he is sick, I ask him what is wrong; when another dog growls at him, I pull him aside to ask if he’s okay. He does what he can to relay his thoughts back to me: He barks; he sighs; he scratches at the door.

But of course we have never talked to each other, not really. Some 15,000 years since humans first domesticated the wolf, scientists have learned that different barks mean different things—for instance, dogs use lower, longer barks for strangers—but our understanding of dog communication remains rather limited. (Researchers are careful to call it communication, not language, because no animal has been shown to possess the same complexity of verbal systems as humans.)

Although a bark at a squirrel is easy enough to decipher (I will eat you!), humans have more trouble knowing whether a whine is just a dog having random feelings on a Tuesday—or something far more serious. Dog owners often joke about how they’d give up years of their life just to have a chance to talk to their pet for a single hour or day. Meanwhile, hucksters posing as dog whisperers and pet psychics have happily taken their cash by claiming to be able to help them translate their dogs’ inner thoughts.

Now, amid a wave of broader interest in applications for artificial intelligence, some dog researchers are hoping that AI might provide answers. In theory, the technology is well suited for such a purpose. AI, at its core, is a pattern-recognition machine. ChatGPT is able to respond in language that seems human, because it has been trained on massive datasets of writing, which it then mimics in its responses. A similar premise applies to other generative-AI programs; large language models identify patterns in the data they’re fed, map relationships among them, and produce outputs accordingly.

Researchers are working with this same theory when it comes to dogs. They’re feeding audio or video of canines to a model, alongside text descriptions of what the dogs are doing. Then they’re seeing if the model can identify statistical patterns between the animals’ observed behavior and the noises they’re making. In effect, they’re attempting to “translate” barks.

Researchers have used similar approaches to study dog communication since at least 2006, but AI has recently gotten far better at processing huge amounts of data. Don’t expect to discuss the philosophy of Immanuel Kant with Fido over coffee anytime soon, however. It’s still early days, and researchers don’t know what kind of breakthroughs AI could deliver—if any at all. “It’s got huge potential—but the gap between the potential and the actuality hasn’t quite emerged yet,” Vanessa Woods, a dog-cognition expert at Duke University, told me.

Right now, researchers have a big problem: data. Modern chatbots are trained on large collections of text—trillions of words—that give them the illusion of language fluency. To create a model capable of translating, say, dog barks into English (if such a thing is even possible), researchers would need millions, if not billions, of neatly cataloged clips. These barks will need to be thoroughly labeled by age, breed, and situation—separating out a 10-year-old male labradoodle barking at a stranger from a six-week-old bichon frise puppy playing with its littermate.

No such catalog currently exists. This is one of the great ironies of the project: Dogs are all around us, constantly captured by phones and doorbell cameras and CCTV. You don’t need to watch Planet Earth to see the canine living in its natural habitat; the internet is filled with more clips of dogs than anyone could watch in a lifetime. And yet all of this media has never been cataloged in a serious way, at least not on the scale that would be necessary for us to better understand what their barks mean.

Perhaps the best catalog that exists is from researchers in Mexico, who have systematically recorded dogs in their homes in specific situations, getting them to bark by, say, knocking on a door or squeaking a favorite toy. A research team at the University of Michigan took some of the 20,000 recordings included in the dataset and fed it into a model trained to recognize human speech. They played barks for the model, and then had it predict what they were barking at, just based on sound. The model could predict which situation preceded the bark with about 60 percent accuracy. That’s nowhere near perfect, but still better than chance, especially considering that the model had more than a dozen bark contexts to pick from.

The same approach of using AI to decipher dog barks is happening with other animals. Perhaps the most promising work is with whale chatter, as my colleague Ross Andersen has written. Other researchers are tackling pigs, bats, chimpanzees, and dolphins. One foundation is offering up to $10 million in prize money to anyone who can “crack the code” and have a two-way conversation with an animal using generative AI.

[Read: How first contact with whale civilization could unfold]

Dogs probably won’t be the animals that help scientists win the prize. “I do not think they necessarily use words and sentences and paragraphs,” Rada Mihalcea, a co-author of the Michigan study, told me over Zoom. (Naturally, in the middle of our call, a stranger knocked on my door, causing my foster dog to bark.) As much as dog owners like myself might want something akin to Google Translate for dogs, Mihalcea’s starting with much more narrow ambitions. She hopes this line of research can “help us get an understanding into what is even there as a linguistic system—if there is such a system.”

Another research group, led by Kenny Zhu at the University of Texas at Arlington, is taking a different approach. His team is scraping massive amounts of dog videos from YouTube. But the data are extremely noisy—quite literally. The researchers have to isolate the barks from all the other sounds that happen in the background of the videos, which makes the process onerous. Zhu’s team does have preliminary findings: They had their algorithms process the sounds of six different breeds (huskies, Shiba Inus, pit bulls, German shepherds, Labradors, and Chihuahuas), and believe they’ve found 105 unique phonemes, or sound units, that span all the breeds.

Even if researchers are able to eventually get a perfect dataset, they’ll run into another problem: There’s no way to know for sure that whatever observations the AI makes is right. When training other AI models on human languages, a native speaker can verify that an output is correct, and help fine-tune the model. No dog will ever be able to verify the AI’s results. (Imagine a dog sitting in an academic research lab, nodding solemnly: Yes, that’s correct.“Ruff-ruff-ruff” means“Give me the chicken.”) The dream of AI as an intermediary between humans and dogs faces a fundamental bias: It is human researchers who are using human-made AI models and human ideas of language to better understand canines. No matter how good the technology gets, there will always be unknowns.

The focus on better understanding dogs’ verbal noises can obscure how much we already know about them. Dogs have evolve d to better communicate with humans: Their barks have changed, and their eyes have grown more expressive. Feral dogs and wolves bark less than pets, suggesting that humans are a big reason why our pups make noise. “The whole thing about dog genius is that they can communicate with us without speaking,” Woods told me. “We can also read them really clearly, which is why we’re so in love with them.”

[Read: Dogs are entering a new wave of domestication]

I know what she means. During a heat wave this summer, I decided to buy heat-resistant dog boots to protect my pup from the scorching pavement. You put them on by stretching them over your dog’s paws, and snapping them into place. The first time I put them on my dog, he stared at me. When I tried to walk him in them later that week, he thrashed in the grass and ran around chaotically. He did not want to wear the boots. And I did not need an AI to know that.

The Next Big Thing Is Still … Smart Glasses

2024-10-01T11:24:06-04:00

Last week, Mark Zuckerberg stood on a stage in California holding what appeared to be a pair of thick black eyeglasses. His baggy T-shirt displayed Latin text that seemed to compare him to Julius Caesar—aut Zuck aut nihil—and he offered a bold declaration: These are Orion, “the most advanced glasses the world has ever seen.”

Those glasses, just a prototype for now, allow users to take video calls, watch movies, and play games in so-called augmented reality, where digital imagery is overlaid on the real world. Demo videos at Meta Connect, the company’s annual conference, showed people playing Pong on the glasses, their hands functioning as paddles, as well as using the glasses to project a TV screen onto an otherwise blank wall. “A lot of people have said that this is the craziest technology they’ve ever seen,” Zuckerberg said. And although you will not be able to buy the glasses anytime soon, Meta is hawking much simpler products in the meantime: a new Quest headset and a new round of software updates to the company’s smart Ray-Bans, which have cameras and an AI audio assistant on board, but no screen in the lenses.

Orion seems like an attempt to fuse those two devices, bringing a fully immersive computerized experience into a technology that people might actually be comfortable putting on their face. And it is not, you may have noticed, the only smart-glasses product to have emerged in recent months. Amazon, Google, Apple, and Snap are all either officially working on some version of the technology or rumored to be doing so. Their implementations are each slightly different, but they point to a single idea: that the future is about integrating computing more seamlessly into everyday life.

Smartphones are no longer exciting, and the market for them has been declining for the past few years. The primary new idea there is foldable screens, which effectively allow your phone to turn into a tablet—though tablet sales have slowed too. The virtual-reality headsets that companies have spent billions developing aren’t being widely adopted.

These companies are betting big that people want to be able to check the weather without pulling out a smartphone—and that they are more willing to wear a pair of Ray-Bans with cameras than spend hours in the metaverse. And after years of false starts on the glasses front, they’re betting that AI—despite some high-profile flops—will be what finally helps them achieve this vision.

Tech companies have been working on smart frames for decades. The first real consumer smart glasses started appearing in the late 1980s and ’90s, but none broke through. At last, in 2013, Google released its infamous Glass eyewear. A thin metal frame with a camera and tiny screen above one eye, Glass could be used to check emails, take photos, and get directions. They were advanced for their time, but the general public was spooked by the idea of face-cameras constantly surveilling them. In 2015, Google abandoned the idea that Glass might ever be a consumer product, though the frames lived on as an enterprise device until last year.

Glass’s failure didn’t deter other companies from taking a swing. In 2016, Snapchat launched its first generation of Spectacles, glasses that allowed users to capture pictures and videos from cameras mounted above each eye, then post them on their account. In 2019, Amazon jumped in, teasing its Echo Frames—camera-less smart glasses with Alexa built in—which went on sale to the public the following year. Meta, then called Facebook, launched the first iteration of its collaboration with Ray-Ban in 2021, though the frames didn’t catch on.

Then there are the virtual-reality headsets, such as Meta’s Quest line. Last summer, after Apple announced the Vision Pro, my colleague Ian Bogost deemed this the “age of goggles,” pointing out that companies have been spending billions developing immersive technology, even though the exact purpose of these expensive headsets is unclear.

Consumers also seem to be wondering what that purpose is. One analyst reports that sales of the Vision Pro were so dismal that Apple scaled back production. According to The Information, the company paused work on the next model, while Meta canceled its competitor device entirely.

[Read: The age of goggles has arrived]

In some ways, this glasses moment is something of a retreat: an acknowledgment that people may be less likely to go all in on virtual reality than they are to throw on a pair of sunglasses that happens to be able to record video. These devices are supposed to look and feel more natural, while allowing for ambient-computing features, such as the ability to play music anywhere just by speaking or start a phone call without having to put in headphones.

AI is a big part of this pitch. New advances in large language models are making modern chatbots seem smarter and more conversational, and this technology is already finding its way into the glasses. Both the Meta and Amazon frames have audio assistants built in that can answer questions (How do whales breathe?) and cue up music (play “Teenage Dirtbag”). Meta’s Ray-Bans can “look” using their cameras, offering an audio description of whatever is in their field of vision. (In my experience, accuracy can be hit or miss: When I asked the audio assistant to find a book of poetry on my bookshelf, it said there wasn’t one, overlooking an anthology with the word poetry in the title, though it did identify my copy of Joseph Rodota’s The Watergate when I asked it to find a book about the Washington landmark.). At Connect, Zuckerberg said that the company plans to keep improving the AI, with a couple of big releases coming in the next few months. These updates will give the glasses the ability to do translation in real time, as well as scan QR codes and phone numbers on flyers in front of you. The AI will also, he said, be able to “remember” such things as where you parked your car. One demo showed a woman ruffling through a closet and asking the AI assistant to help her choose an outfit for a theme party.

[Read: The end of foreign-language education]

But whether AI assistants will actually be smart enough to realize all of this is still somewhat of an open question. In general, generative AI struggles to cite its sources and frequently gets things wrong, which may limit smart glasses’ overall usefulness. And, though the companies say the technology will only get better and better, that’s not entirely certain: The Wall Street Journal recently reported that, when Amazon attempted to infuse Alexa with new large language models, the assistant actually became less reliable for certain tasks.

Products such as Orion, which promise not just AI features but a full, seamless integration of the digital world into physical reality, face even steeper challenges. It’s really, really difficult to squish so many capabilities into eyewear that looks semi-normal. You need to be able to fit a battery, a camera, speakers, and processing chips all into a single device. Right now, even some of the most state-of-the-art glasses require you to be tethered to additional hardware to use them. According to The Verge’s Alex Heath, the Orion glasses require a wireless “compute puck” that can be no more than about 12 feet away from them—something Zuckerberg certainly did not mention onstage. Snap’s newest Spectacles, announced earlier this month, don’t require any extra hardware—but they have a battery life of only 45 minutes, and definitely still look big and clunky. The hardware problem has bedeviled generations of smart glasses, and there still isn’t a neat fix.

But perhaps the biggest challenge facing this generation of smart glasses is neither hardware nor software. It’s philosophical. People are stressed right now about how thoroughly technology has seeped into our everyday interactions. They feel addicted to their phones. These companies are pitching smart glasses as a salve—proposing that they could, for example, allow you to handle a text message without interrupting quality time with your toddler. “Instead of having to pull out your phone, there will just be a little hologram,” Zuckerberg said of Orion during his presentation. “And with a few subtle gestures, you can reply without getting pulled away from the moment.”

Yet committing to a world in which devices are worn on our face means committing to a world in which we might always be at least a little distracted. We could use them to quietly read our emails or scroll Instagram at a restaurant without our partner knowing. We could check our messages during a meeting while looking like we’re still paying attention. We may not need to check our phones so much, because our phones will effectively be connected to our eyeballs. Smart glasses walk a thin line between helping us be less obsessively on the internet and tethering us even more closely to it.

I spent some time this spring talking with a number of people who worked on early smart glasses. One of them was Babak Parviz, a partner at Madrona, a venture-capital firm, who previously led Google’s Glass project. We discussed the history of computers: They used to be bulky things that lived in research settings—then we got laptops, then smartphones. With Glass, the team aimed to shorten the time needed to retrieve information to seconds. “The question is, how much further do you need to take that? Do you really need to be immersed in information all the time, and have access to much faster information?” Parvis told me he’d changed his mind about what he called “information snacking,” or getting fed small bits of information throughout the day. “I think constant interruption of our regular flow by reaching out to information sources doesn’t feel very healthy to me.”

In my conversations, I asked experts whether they thought smart glasses were inevitable—and what it would take to unseat the smartphone. Some saw glasses not as a smartphone replacement at all, but as a potential addition. In general, they thought that new hardware would have to give us the ability to do something we can’t do today. Right now, companies are hoping that AI will be the thing to unlock this potential. But as with so much of the broader conversation around that technology, it’s unclear how much of this hype will actually pan out.

These devices still feel more like sketches of what could be, rather than fully realized products. The Ray-Bans and other such products can be fun and occasionally useful, but they still stumble. And although we might be closer than ever to mainstream AR glasses, they still seem a long way off.

Maybe Zuckerberg is right that Orion is the world’s most advanced pair of glasses. The question is really whether his big vision for the future is what the rest of us actually want. Glasses could be awesome. They could also be just another distraction.

Smartphones Are So Over

2024-09-17T14:09:20-04:00

Today, Snap, the parent company of Snapchat, one of the most popular social-media apps for teenage users, is announcing a new computer that you wear directly on your face. The latest in its Spectacles line of smart glasses, which the company has been working on for about a decade, shows you interactive imagery through its lenses, placing plants or imaginary pets or even a golf-putting range into the real world around you.

So-called augmented reality (or AR) is nothing new, and neither is wearable tech. Meta makes a pair of smart glasses in partnership with Ray-Ban, and claims they’re so popular that the company can’t make them fast enough. Amazon sells an Alexa-infused version of the famous Carrera frames, which make you look like a mob boss with access to an AI assistant (Alexa, where’s the best place to hide a body?). Apple launched its Vision Pro headset—which includes an AR mode, along with a fully immersive virtual-reality one—last year. And who could forget Google Glass? Consumers have sometimes been cool on the face computers, if not outright hostile toward them, but tech companies just can’t seem to quit the idea. From that perspective, it makes sense that Snap’s new Spectacles are more a demonstration of intent than an actual product: They’re targeted to developers who will apply and pay $99 a month to use them.

But this is also, arguably, what makes them interesting. In an interview last week, Snap CEO Evan Spiegel told me that he sees smart glasses as an opportunity to “reshape what a computer is, to make it something that actually keeps us grounded in the real world rather than behind a screen.” The company hasn’t accomplished this so far, of course, but the new Spectacles—and all those other smart glasses and AR headsets—are not being released into a void. They’re arriving at a moment when people are feeling pretty turned off by phones. People are angsty about how much time they spend looking down at small screens rather than engaging with the world around them. Parents are concerned that phones are driving a teen mental-health crisis. Smartphone sales have slowed, and even the latest iPhone isn’t doing great. Companies are trying to get people excited about technology again, by pitching all sorts of new hardware ideas that break the bounds of that rectangular screen, such as lapel pins or glorified walkie-talkies that work with AI assistants. I had this moment in mind as I wore the new Spectacles earlier this month, batting colorful digital blobs away while Paramore’s “Misery Business” played in the background.

[Read: Yet another iPhone, dear God]

Among all the new glasses options, the Spectacles are distinct. They are oriented less toward utility—say, asking Alexa to set a timer—and more toward fun. In doing so, they offer a very specific formulation for the future of computing: that it should be amusing and connective. “If we look at the history of computers, they’ve actually always kept us indoors, taken us away from people that we love,” Spiegel told me. Growing up, he explained, he loved computers, but he had to go to the computer lab to use them, which meant forgoing the opportunity to hang out with friends during recess. He thinks smart glasses are an opportunity to reinvent screens by integrating computers more naturally into one’s life.

But Spectacles are still far from perfect. For starters, they are notably heavy. When I tried the glasses, they got warm to the touch after use, despite Snap’s assurances that it had invested in a state-of-the-art cooling system. They support up to 45 minutes of continuous usage, which isn’t very long. They reminded me of snorkeling goggles. You absolutely could not wear them in daily life without someone asking you what exactly you’ve got on your head. Their lenses can be dimmed to look like sunglasses, or made clear so people can still see your eyes. The glasses are controlled by your hands, held out in front of you. You pinch your index finger and your thumb together to “click.” (The onboarding process involves practicing by popping bubbles floating a few feet from your face.)

Mostly, they’re fun. Snapchat is famously popular with young people, and the glasses feel like a piece of hardware designed for this audience—closer to a Nintendo Switch than a Google Glass. In one game developed in partnership with Lego, you can project virtual bricks onto your kitchen table and move them around to build different creations. Ask it for an additional small blue brick, and one appears before you. In collaboration with Niantic, the company behind Pokémon Go, Snap is also launching a game called Peridot Beyond that lets you care for virtual pets.

Perhaps most important, at least when it comes to Spiegel’s bigger vision, the Spectacles can sync together, so that multiple people can see the same digital creations at once. In one experience, called Imagine Together, users can shout words to create cartoons that then appear in little bubbles on the screen. “Imagine a fox!” you might say, and then a small fox appears, floating in a bubble in midair between you and your friend.

[Read: The Apple Vision Pro is spectacular and sad]

Spiegel, who has four children, dreams that someday he’ll see his kids playing together in augmented reality. I asked him what he might say to parents who would be nervous about their children adding an additional level of computing into their daily life. (Parents are already plenty concerned about screen time as is, without the screens being barely an inch from their teens’ faces.) What would he say to the parents who just want their kids to go outside? Spiegel countered that he is a go-outside-and-play parent himself—but argued that the glasses could make playing together outside more fun.

At times, I found the Spectacles genuinely amusing, in a way the current Meta and Alexa glasses aren’t. And yet, they still don’t feel essential. Any device that’s hoping to disrupt the smartphone will have to be extremely good. Whether smart glasses are indeed the future of computing will depend on whether someone can make a pair that’s useful in day-to-day life. Spectacles aren’t there yet; they’re more novelty than utility. But the philosophical argument they make is a provocative one, even if it’s just that right now—an argument. Like the imaginary pet I saw while wearing them, it technically exists, but just barely.

This Fire Is Too Close to L.A. for Comfort

2024-09-10T16:13:33-04:00

From downtown Los Angeles all the way out to the edge of the Line Fire is sprawl that turns into more sprawl. It’s just block after block after block of homes and businesses and people living their life, until on one side of the street is a suburban neighborhood, and on the other side, a 26,000-acre wildfire. Some 65,000 buildings are threatened, and more than 10,000 people have been ordered to evacuate.

In recent years, fires have begun spilling into places dominated by people. Americans who live on the edges of major cities have long been much safer from the threat of fire than those who live in the middle of a forest. But wildfires in the West are growing so big, and so quickly, that cities are becoming vulnerable too.

Cities used to burn all the time. My predecessors at The Atlantic covered urban blazes in Portland, Maine (1866), and in the magazine’s hometown of Boston (1872). Chicago famously burned in 1871. These began as urban fires, started by human error or other mishaps—legend blames the Great Chicago Fire on a cow knocking over a lantern in a barn—but natural disasters could set them off too: San Francisco went up in flames in 1906, in the aftermath of a devastating earthquake. In the 20th century, people started to get serious about fire prevention. They developed thorough fire codes, the kind we’re used to today: sprinkler systems, fire exits, evacuation signs. Catastrophic urban fires became old horror stories.

Then, in the 21st century, the wildfires got big—so big that they started roaring into more densely populated areas. In 2017, a fire hit Santa Rosa, in California’s wine country, and flattened more than 5,000 structures. Then, in 2018, a fire tornado tore into the fringes of Redding (population 91,000), in Northern California. In 2021, a late-December fire in Colorado blew into the Boulder suburbs, destroying about 1,000 homes. Then, last year, on Maui, the deadliest fire recorded in modern American history destroyed the town of Lahaina in a matter of hours.

“I see it as like watching polio come back, or some plague that we fixed,” Stephen J. Pyne, a professor emeritus at Arizona State University and the author of The Pyrocene, told me. Pyne was part of the team behind a 2023 paper arguing that, although public perception of these disasters is that they were “wildfires that involved houses,” they really were “urban fires initiated by wildfires.” Essentially, even if a fire starts as a wildfire, when it reaches an urban area, it can change so much in the way it spreads that it’s a different beast. A wildfire moves among trees, but an urban fire moves among buildings.

Modern communities aren’t built to prepare for this kind of spillover. “The problem is fundamentally that we have built cities and towns without all the pyric hygiene that used to come with the cities,” Pyne explained. For decades, no one had to think about this problem, so no one did, even as cities grew and sprawl became a default. “Everybody thought it was done,” he said.

When fire scientists talk about urban fire, they don’t necessarily mean a fire unfolding in the center of a major city. They also mean suburbs and smaller cities—anywhere that has homes close together. Fighting a fire deep in a forest requires a very different strategy than fighting a fire in a neighborhood. Wildland firefighters try to prioritize life and property, but their job is to wrangle blazes into control. That could mean letting some areas burn if they’re not densely inhabited. But for urban firefighters, as Pyne pointed out, “every fire is an existential threat to life and property.” When an urban spillover fire occurs, firefighters have to deal with both types of fire at once. It’s no wonder that these types of fires are among some of the costliest and most destructive in recent history.

Part of what’s causing so many of these spillover events are embers. Giant fires can emit sparks that, when blown by the wind, can travel up to five miles ahead of the fire. Pyne compared them to a blizzard, or a swarm of locusts. They can burrow through a rooftop vent into a home’s attic, igniting some forgotten box of old clothes. Then the whole house catches fire. The solution is, essentially, to fortify the homes on the outskirts of communities. Homes can be built with special, more fire-resistant materials, and homeowners can clear their property of highly flammable items close to their house. These are standard precautions, sometimes even required by law for people who live on the edges of forests. But now cities and homeowners have more reason to weigh taking these precautions miles into the built environment. Barring extreme wildfire conditions, a fire probably wouldn’t burn all the way to the skyscrapers of Los Angeles, but one could burn thousands of homes on the fringe of the city.

The Line Fire isn’t even the only fire burning around Los Angeles right now; it’s just the biggest. Thankfully, it appears to be moving north and east, away from the suburbs and deeper into the forest, and firefighters have been able to contain the part of the fire that brushes up against the most densely populated area, Rick Carhart, a public-information officer with Cal Fire, told me. (Some mountain towns, including Big Bear Lake, are still under threat.) The wind occasionally changes directions in a way that’s unpredictable—but unless they make a catastrophic shift, the L.A. suburbs seem safe.

Still, the whole thing is just a bit too close for comfort. Major fires are burning across the West right now, in Oregon and Nevada and Idaho and Montana. Some of those fires are bigger than the Line Fire; so much area is burning right now that the country’s firefighting resources are strained. Fires keep getting larger and unrulier, thanks in part to climate change—but also because, over the past century, Americans suppressed many natural fires rather than letting them burn through. Now the forests are loaded with potential fuel, and big fires keep happening.

These fires mean fighting that much harder to keep them in the wildland. Fires are a natural part of many forests’ ecology; it’s not unusual for forests to burn. But humans, perhaps a bit arrogantly, thought they had bumped the problem off their streets and into the woods forever. Now it’s creeping back in, and flames keep brushing up against our communities, forcing us to rethink who is at risk.

The Dome Is Watching You

2024-08-29T16:33:39-04:00

On a recent Wednesday night in Los Angeles, I was ready to buy a hot dog with my face.

I was at the Intuit Dome, a $2 billion entertainment complex that opened earlier this month. Soon, it will be the home of the L.A. Clippers, but I was there to watch Olivia Rodrigo, queen of teen angst, perform a sold-out show. The arena was filled with people wearing purple cowboy hats and the same silver sequin miniskirt, all of us ready to scream-sing for two hours straight. But first, we needed food.

[Read: Sphere and loathing in Las Vegas]

Feeding yourself—or, really, doing much of anything—at the Dome requires the use of an official app. When you register, it asks for your name, phone number, email address, and zip code. If you want, you can also add your credit-card information and upload a selfie as part of the “Game Face ID” program. That last part, though optional, is a key feature of the venue: Facial-recognition cameras are absolutely everywhere. They’re embedded in large, basketball-shaped devices with circular screens. Some of them are planted in walls, while others stand alone atop black poles. They are the keepers of the Dome. If they recognize you, they will grant you prompt entry to the venue, club suites, and concession stands.

Creeping surveillance is a well-documented phenomenon at major venues: Many arenas throughout the country have used some form of facial recognition for years, typically under the premise that it makes the overall experience more convenient for customers. But the Dome is one of the first to package all of this in earnest, to create the ultimate smartphone-powered, face-recognizing, fully digitized stadium-going experience. It is a preview of a new generation of tech-supercharged event venues, a teaser for a world where you can’t even buy chicken tenders at a basketball game without first setting up an account.

[Read: Why go with an evil-looking orb?]

But on the night of the Rodrigo concert, I wasn’t thinking about any of this: I just wanted my hot dog. My boyfriend and I had made the conscious decision not to upload selfies before the event—I try to use facial recognition sparingly, for privacy reasons—but a long wait and technical difficulties left me feeling like I would have given up my Social Security number for some sustenance. After eight minutes in line, we finally approached the cameras. They weren’t working very well. Employees posted at each concession entrance had to manually help guests navigate the system, one by one. It took three minutes of tapping our phones and letting the cameras scan our faces to get the gate to open. (Even if you don’t enroll in the facial-recognition feature, the unit attempts to find a match when you approach.) Once inside, we quickly picked up our food from among the boxes neatly laid out for us, and left. An elaborate system that uses computer vision and yet more cameras—I counted more than 20 mounted on the ceiling—recognizes the selected items and automatically charges consumers accordingly. There’s no need to interact with another person or swipe a credit card—and certainly no need to fuss around with cash, which is, in fact, not accepted at the arena. Later, I found the receipt in my app: $26.40 for two hot dogs and a churro. (They were pretty good.)

There are, of course, a few caveats: If you do not opt in to the facial-recognition system, you can use the app’s “Identity Pass”—a sort of digital ID card that can be added to your Apple or Google Wallet—to gain entry to the concession stand. You can also choose to use a physical card or Apple or Google Pay to tap in and pay anonymously. Children as well as people with assistance needs may also forgo the app in favor of tap-to-enter wristbands. Yet there is no question that convenience is a powerful motivator for people to enroll in the facial-recognition system. A few days after Rodrigo’s performance, I returned to tour the Dome with George Hanna, the chief technology and digital officer for the Clippers. He told me that, overall, about 50 percent of guests have opted in to the Game Face ID program at the start of an event—but that, by the end, the number grows to 70 to 75 percent of attendees.

The system, he said, stores just the single selfie, which the camera compares to the person standing in front of it. Hanna told me there is no ambient facial collection, and that faces are only scanned by the devices in the context of a “transaction”: walking into the arena, trying to get into a club. He added that users can delete their selfie at any time, in which case the image is cleared from the Dome’s system immediately. People who aren’t comfortable with the system simply don’t have to opt in, he said.

[Read: The tech that’s radically reimagining the public sphere]

People have good reason to be suspicious about all of this. Last year, a lawyer chaperoning her daughter’s Girl Scout troop to Radio City Music Hall was denied entry to a Rockettes show after a facial-recognition system flagged her: She was on an “attorney exclusion list” that had been instituted to prevent firms involved in litigation against MSG Entertainment from entering venues owned by the company. The case made national headlines and angered privacy advocates, who saw it as a warning of the technology’s abuse potential. (In a statement to NPR at the time, MSG Entertainment said, in part, “While we understand this policy is disappointing to some, we cannot ignore the fact that litigation creates an inherently adversarial environment.”) Just last week, a group of privacy orgs protested against the use of facial recognition at a Major League Baseball game at New York’s Citi Field. In an open letter, Fight for the Future, one such group, argued that the technology is invasive and unnecessary, and that it should not be normalized.

On my second trip to the Dome, I decided to try facial recognition for myself. Hanna said that the system was working “light-years” better than it did on opening night. I uploaded a selfie to the app, and the orb on a stick let me inside in less than a minute. I was also able to get into the self-service concession area no problem.

This time I was able to use my face to buy a box of churros. As we wandered the stadium’s curving halls, I ate them, and asked Hanna a question that had been bugging me: How many cameras are in the Dome? “A lot,” he said. I let out a nervous laugh. “More than 10,000?” I asked. Fewer than that, he said, but demurred on giving an exact number. He wasn’t trying to be cagey, he explained. He just didn’t know.

Silicon Valley Is Coming Out in Force Against an AI-Safety Bill

2024-08-21T19:18:00-04:00

Since the start of the AI boom, the attention on this technology has focused on not just its world-changing potential, but also fears of how it could go wrong. A set of so-called AI doomers have suggested that artificial intelligence could grow powerful enough to spur nuclear war or enable large-scale cyberattacks. Even top leaders in the AI industry have said that the technology is so dangerous, it needs to be heavily regulated.

A high-profile bill in California is now attempting to do that. The proposed law, Senate Bill 1047, introduced by State Senator Scott Wiener in February, hopes to stave off the worst possible effects of AI by requiring companies to take certain safety precautions. Wiener objects to any characterization of it as a doomer bill. “AI has the potential to make the world a better place,” he told me yesterday. “But as with any powerful technology, it brings benefits and also risks.”

S.B. 1047 subjects any AI model that costs more than $100 million to train to a number of safety regulations. Under the proposed law, the companies that make such models would have to submit a plan describing their protocols for managing the risk and agree to annual third-party audits, and they would have to be able to turn the technology off at any time—essentially instituting a kill-switch. AI companies could face fines if their technology causes “critical harm.”

The bill, which is set to be voted on in the coming days, has encountered intense resistance. Tech companies including Meta, Google, and OpenAI have raised concerns. Opponents argu e that the bill will stifle innovation, hold developers liable for users’ abuses, and drive the AI business out of California. Last week, eight Democratic members of Congress wrote a letter to Governor Gavin Newsom, noting that, although it is “somewhat unusual” for them to weigh in on state legislation, they felt compelled to do so. In the letter, the members worry that the bill overly focuses on the most dire effects of AI, and “creates unnecessary risks for California’s economy with very little public safety benefit.” They urged Newsom to veto it, should it pass. To top it all off, Nancy Pelosi weighed in separately on Friday, calling the bill “well-intentioned but ill informed.”

In part, the debate over the bill gets at a core question with AI. Will this technology end the world, or have people just been watching too much sci-fi? At the center of it all is Wiener. Because so many AI companies are based in California, the bill, if passed, could have major implications nationwide. I caught up with the state senator yesterday to discuss what he describes as his “hardball politics” of this bill—and whether he actually believes that AI is capable of going rogue and firing off nuclear weapons.

Our conversation has been condensed and edited for clarity.

Caroline Mimbs Nyce: How did this bill get so controversial?

Scott Wiener: Any time you’re trying to regulate any industry in any way, even in a light-touch way—which, this legislation is light-touch—you’re going to get pushback. And particularly with the tech industry. This is an industry that has gotten very, very accustomed to not being regulated in the public interest. And I say this as someone who has been a supporter of the technology industry in San Francisco for many years; I’m not in any way anti-tech. But we also have to be mindful of public interest.

It’s not surprising at all that there was pushback. And I respect the pushback. That’s democracy. I don’t respect some of the fearmongering and misinformation that Andreessen Horowitz and others have been spreading around. [Editor’s note: Andreessen Horowitz, also known as a16z, did not respond to a request for comment.]

Nyce: What in particular is grinding your gears?

Wiener: People were telling start-up founders that S.B. 1047 was going to send them to prison if their model caused any unanticipated harm, which was completely false and made up. Putting aside the fact that the bill does not apply to start-ups—you have to spend more than $100 million training the model for the bill even to apply to you—the bill is not going to send anyone to prison. There have been some inaccurate statements around open sourcing.

These are just a couple of examples. It’s just a lot of inaccuracies, exaggerations, and, at times, misrepresentations about the bill. Listen: I’m not naive. I come out of San Francisco politics. I’m used to hardball politics. And this is hardball politics.

Nyce: You’ve also gotten some pushback from politicians at the national level. What did you make of the letter from the eight members of Congress?

Wiener: As much as I respect the signers of the letter, I respectfully and strongly disagree with them.

In an ideal world, all of this should be handled at the federal level. All of it. When I authored California’s net-neutrality law in 2018, I was very clear that I would be happy to close up shop if Congress were to pass a strong net-neutrality law. We passed that law in California, and here we are six years later; Congress has yet to enact a net-neutrality law.

If Congress goes ahead and is able to pass a strong federal AI-safety law, that’s fantastic. But I’m not holding my breath, given the track record.

Nyce: Let’s walk through a few of the popular critiques of this bill. The first one is that it takes a doomer perspective. Do you really believe that AI could be involved in the “creation and use” of nuclear weapons?

Wiener: Just to be clear, this is not a doomer bill. The opposition claims that the bill is focused on “science-fiction risks.” They’re trying to say that anyone who supports this bill is a doomer and is crazy. This bill is not about the Terminator risk. This bill is about huge harms that are quite tangible.

If we’re talking about an AI model shutting down the electric grid or disrupting the banking system in a major way—and making it much easier for bad actors to do those things—these are major harms. We know that there are people who are trying to do that today, and sometimes succeeding, in limited ways. Imagine if it becomes profoundly easier and more efficient.

In terms of chemical, biological, radiological, nuclear weapons, we’re not talking about what you can learn on Google. We’re talking about if it’s going to be much, much easier and more efficient to do that with an AI.

Nyce: The next critique of your bill is around harm—that it doesn’t address the real harms of AI, such as job losses and biased systems.

Wiener: It’s classic whataboutism. There are various risks from AI: deepfakes, algorithmic discrimination, job loss, misinformation. These are all harms that we should address and that we should try to prevent from happening. We have bills that are moving forward to do that. But in addition, we should try to get ahead of these catastrophic risks to reduce the probability that they will happen.

Nyce: This is one of the first major AI-regulation bills to garner national attention. I would be curious what your experience has been—and what you’ve learned.

Wiener: I have definitely learned a lot about the AI factions, for lack of a better term—the effective altruists and effective accelerationists. It’s like the Jets and the Sharks.

As is human nature, the two sides caricature each other and try to demonize each other. The effective accelerationists will classify the effective altruists as insane doomers. Some of the effective altruists will classify all of the effective accelerationists as extreme libertarians. Of course, as is the case with human existence, and human opinions, it’s a spectrum.

Nyce: You don’t sound too frustrated, all things considered.

Wiener: This legislative process—even though I get frustrated with some of the inaccurate statements that are made about the bill—this has actually been, in many ways, a very thoughtful process, with a lot of people with really thoughtful views, whether I agree or disagree with them. I’m honored to be part of a legislative process where so many people care, because the issue is actually important.

When the opposition refers to the risks of AI as “science fiction,” well, we know that’s not true, because if they really thought the risk was science fiction, they would not be opposing the bill. They wouldn’t care, right? Because it would all be made up. But it’s not made-up science fiction. It’s real.

Why Does AI Art Look Like That?

2024-08-16T14:09:33-04:00

This week, X launched an AI-image generator, allowing paying subscribers of Elon Musk’s social platform to make their own art. So—naturally—some users appear to have immediately made images of Donald Trump flying a plane toward the World Trade Center; Mickey Mouse wielding an assault rifle, and another of him enjoying a cigarette and some beer on the beach; and so on. Some of the images that people have created using the tool are deeply unsettling; others are just strange, or even kind of funny. They depict wildly different scenarios and characters. But somehow they all kind of look alike, bearing unmistakable hallmarks of AI art that have cropped up in recent years thanks to products such as Midjourney and DALL-E.

Two years into the generative-AI boom, these programs’ creations seem more technically advanced—the Trump image looks better than, say, a similarly distasteful one of SpongeBob SquarePants that Microsoft’s Bing Image Creator generated last October—but they are stuck with a distinct aesthetic. The colors are bright and saturated, the people are beautiful, and the lighting is dramatic. Much of the imagery appears blurred or airbrushed, carefully smoothed like frosting on a wedding cake. At times, the visuals look exaggerated. (And yes, there are frequently errors, such as extra fingers.) A user can get around this algorithmic monotony by using more specific prompts—for example, by typing a picture of a dog riding a horse in the style of Andy Warhol rather than just a picture of a dog riding a horse. But when a person fails to specify, these tools seem to default to an odd blend of cartoon and dreamscape.

These programs are becoming more common. Google just announced a new AI-image-making app called Pixel Studio that will allow people to make such art on their Pixel phone. The app will come preinstalled on all of the company’s latest devices. Apple will launch Image Playground as part of its Apple Intelligence suite of AI tools later this year. OpenAI now allows ChatGPT users to generate two free images a day from DALL-E 3, its newest text-to-image model. (Previously, a user needed a paid premium plan to access the tool.) And so I wanted to understand: Why does so much AI art look the same?

[Read: AI has a hotness problem]

The AI companies themselves aren’t particularly forthcoming. X sent back a form email in response to a request for comment about its new product and the images its users are creating. Four firms behind popular image generators—OpenAI, Google, Stability AI, and Midjourney—either did not respond or did not provide comment. A Microsoft spokesperson directed me toward some of its prompting guides and referred any technical questions to OpenAI, because Microsoft uses a version of DALL-E in products such as Bing Image Creator.

So I turned to outside experts, who gave me four possible explanations. The first focuses on the data that models are trained on. Text-to-image generators rely on extensive libraries of photos paired with text descriptions, which they then use to create their own original imagery. The tools may inadvertently pick up on any biases in their data sets—whether that’s racial or gender bias, or something as simple as bright colors and good lighting. The internet is filled with decades of filtered and artificially brightened photos, as well as a ton of ethereal illustrations. “We see a lot of fantasy-style art and stock photography, which then trickles into the models themselves,” Ziv Epstein, a scientist at the Stanford Institute for Human-Centered AI, told me. There are also only so many good data sets available for people to use to build image models, Phillip Isola, a professor at the MIT Computer Science & Artificial Intelligence Laboratory, told me, meaning the models might overlap in what they’re trained on. (One popular one, CelebA, features 200,000 labeled photos of celebrities. Another, LAION 5B, is an open-source option featuring 5.8 billion pairs of photos and text.)

The second explanation has to do with the technology itself. Most modern models use a technique called diffusion: During training, models are taught to add “noise” to existing images, which are paired with text descriptions. “Think of it as TV static,” Apolinário Passos, a machine-learning art engineer at Hugging Face, a company that makes its own open-source models, told me. The model then is trained to remove this noise, over and over, for tens of thousands, if not millions, of images. The process repeats itself, and the model learns how to de-noise an image. Eventually, it’s able to take this static and create an original image from it. All it needs is a text prompt.

[Read: Generative art is stupid]

Many companies use this technique. “These models are, I think, all technically quite alike,” Isola said, noting that recent tools are based on the transformer model. Perhaps this technology is biased toward a specific look. Take an example from the not-so-distant past: Five years ago, he explained, image generators tended to create really blurry outputs. Researchers realized that it was the result of a mathematical fluke; the models were essentially averaging all the images they were trained on. Averaging, it turns out, “looks like blur.” It’s possible that, today, something similarly technical is happening with this generation of image models that leads them to plop out the same kind of dramatic, highly stylized imagery—but researchers haven’t quite figured it out yet. Additionally, “most models have an ‘aesthetic’ filter on both the input and output that reject images that don't meet a certain aesthetic criteria,” Hany Farid, a professor at the UC Berkeley School of Information, told me over email. “This type of filtering on the input and output is almost certainly a big part of why AI-generated images all have a certain ethereal quality.”

The third theory revolves around the humans who use these tools. Some of these sophisticated models incorporate human feedback; they learn as they go. This could be by taking in a signal, such as which photos are downloaded. Others, Isola explained, have trainers manually rate which photos they like and which ones they don’t. Perhaps this feedback is making its way into the model. If people are downloading art that tends to have really dramatic sunsets and absurdly beautiful oceanscapes, then the tools might be learning that that’s what humans want, and then giving them more of that. Alexandru Costin, a vice president of generative AI at Adobe, and Zeke Koch, a vice president of product management for Adobe Firefly (the company’s AI-image tool) told me in an email that user feedback can indeed be a factor for some AI models—a process called “reinforcement learning from human feedback,” or RLHF. They also pointed to training data as well as assessments performed by human evaluators as influencing factors. “Art generated by AI models sometimes have a distinct look (especially when created using simple prompts),” they said in a statement. “That’s generally caused by a combination of the images used to train the image output and the tastes of those who train or evaluate the images.”

The fourth theory has to do with the creators of these tools. Although representatives for Adobe told me that their company does not do anything to encourage a specific aesthetic, it is possible that other AI makers have picked up on human preference and coded that in—essentially putting their thumb on the scale, telling the models to make more dreamy beach scenes and fairylike women. This could be intentional: If such imagery has a market, maybe companies would begin to converge around it. Or it could be unintentional; companies do lots of manual work in their models to combat bias, for example, and various tweaks favoring one kind of imagery over another could inadvertently result in a particular look.

More than one of these explanations could be true. In fact, that’s probably what’s happening: Experts told me that, most likely, the style we see is caused by multiple factors at once. Ironically, all of these explanations suggest that the uncanny scenes we associate with AI-generated imagery are actually a reflection of our own human preferences, taken to an extreme. No surprise, then, that Facebook is filled with AI-generated slop imagery that earns creators money, that Etsy recently asked users to label products made with AI following a surge of junk listings, and that the arts-and-craft store Michaels recently got caught selling a canvas featuring an image that was partially generated by AI (the company pulled the product, calling this an “unacceptable error.”).

[Read: AI-generated junk is flooding Etsy]

AI imagery is poised to seep even further into everyday life. For now, such art is usually visually distinct enough that people can tell it was made by a machine. But that may change. The technology could get better. Passos told me he sees “an attempt to diverge from” the current aesthetic “on newer models.” Indeed, someday computer-generated art may shed its weird, cartoonish look, and start to slip past us unnoticed. Perhaps then we’ll miss the corny style that was once a dead giveaway.

California’s Fire-Insurance Crisis Just Got Real

2024-08-12T12:02:07-04:00

Susie Lawing moved to Cohasset, a small community located in the forested canyons above the city of Chico, California, in the 1970s. After she and her husband divorced, Lawing stayed, presiding over 26 acres of lush family compound. Loved ones built homes of their own on the property, and they began hosting weddings and retreats. Lawing started to grow her own food.

All of that is now gone, she told me. Two weeks ago, the Park Fire ripped through the property. Lawing, now 81, lost everything. She did not have insurance. Lawing lives modestly on Social Security benefits, supplemented by renting out her home and selling essential oils, and simply could not afford the $12,000 a year—$1,000 a month—home-insurance policy she was quoted for a state-backed policy, the last resort for many homeowners. Paying that would have doubled her monthly expenses. “There was no way I could afford that,” she told me. “What do you do? You just let it go.”

Now the family faces the prospect of rebuilding without a safety net. Lawing’s daughter has set up a GoFundMe page for her. (Her grandson Myles Lawing also has a GoFundMe set up for his dad, who had an uninsured home on the property.) Others on the crowdfunding site are raising money for families who’d lost their homes to fire once before, when just six years ago, the deadly Camp Fire raged through the town of Paradise, just 30 minutes down the road. Some Paradise families, such as the Bakers, chose to resettle in Cohasset, only to have their new home burn.

This is the reality of California’s new age of fire. Wildfires have gotten more ferocious in recent years, thanks in part to warming temperatures: Park is the fourth largest in the state’s recorded history. As homes in high-risk areas become harder to insure, premiums are rising, and some insurers are leaving the state altogether. The safety net that people once depended on has developed holes, and now people are falling through.

“People need insurance. It’s essential for their recovery,” Carolyn Kousky, the associate vice president for economics and policy at Environmental Defense Fund, told me. But if state-funded insurance is people’s only option, she said, the question becomes “How much are we going to subsidize that?” As climate change brings about bigger fires and stronger hurricanes and more intense floods, the country is being forced to decide what homes to save and whom to leave on their own.

California’s insurance crisis first started around 2017. In that year and the ones that followed, a series of costly fires erased decades of profits, and forced insurance companies to reconsider their rates and their presence in the state. Premiums began rising, and in the past two years, major national companies including State Farm, Farmers, and Allstate, as well as smaller firms, have pulled back, declining to renew tens of thousands of policies. Coming on top of rising inflation and building costs, wildfires have made the cost of doing business just too high, insurers argue. For those living in areas where no private company will take on the risk, California offers a last-resort option called FAIR. From 2019 to 2024, as insurance companies retreated, the number of California FAIR plans has more than doubled. But FAIR plans are also getting more expensive. Many Californians are underinsured—and some are opting to go without insurance at all.

The people living in the Park Fire burn area are struggling with these exact dynamics. Counting how many people are uninsured in a given area is difficult. But from 2015 to 2021, insurance companies issued almost 7,000 nonrenewal notices in Cohasset’s zip code, which has about 13,000 properties total, according to state data, analyzed and provided to me by First Street Foundation, a nonprofit that models climate risk. That means insurance companies potentially pulled policies for more than half of the homes in the area. And these data are only through 2021, before the exodus of insurers began in earnest.

Cohasset is located in an extremely risky part of the state; First Street Foundation’s models put it at severe fire risk, and predict that 100 percent of the structures in the area will be threatened in the next 30 years. People might balk at high insurance rates, but those prices are a warning of disaster to come. “As brutal as it is, when insurance companies stop offering insurance, it’s a signal that the risk is uninsurable—that those losses are coming,” Abrahm Lustgarten, a reporter and the author of On the Move, a book about climate migration, told me. Blunting these signals with policies such as state-subsidized insurance plans may create incentives to stay when families should really consider leaving.

But moving isn’t easy. It means leaving a life behind, perhaps generations’ worth of local memories. It means uprooting oneself from the community you grew up in, and maybe even saying goodbye to loved ones. For some, this is too difficult. Others are just overly optimistic about the risk—psychology and behavioral-economics research suggest that people have a hard time processing such risk, Kousky pointed out.

Others just can’t afford to go elsewhere. Leaving a place might mean leaving a job, or a business, or a garden that helps you save on groceries. California is an extremely expensive place to live. Moving from the edge of the forest to a city would be safer, but infeasible for some people. Sky-high costs of living have pushed people farther and farther out in search of cheaper housing—and directly in the path of fire. Lustgarten’s reporting suggests that Americans are less likely to pick up and run in terror from disaster, he told me, and more likely to uproot when the cost of staying becomes unrealistic, whether because of a disaster like Park or the rising cost of air-conditioning in a hot area like Phoenix. At first, such migration can be incremental—moving from one town to another nearby, as the people who moved from Paradise to Cohasset did, which didn’t put them beyond risk. People may have to experience loss multiple times before they truly uproot themselves.

The Park Fire is still burning, slinking through the Sierra Nevada and threatening thousands of homes and buildings in the small mountain towns that dot the region. Already, some 600 structures have burned. FEMA provides some individual assistance in the aftermath of a disaster. But the agency has warned over and over that the funding it can offer is no substitute for insurance. Many fire victims are now turning to crowdfunding resources such as GoFundMe to try to blunt catastrophic financial losses: In the past five years, the number of wildfire fundraisers on GoFundMe has tripled, a representative for the company told me in an email.

Lawing’s daughter Jessica Adams told me that she probably wouldn’t be grieving the loss of her family’s compound so hard if they’d had insurance. They still would have been devastated—but at least they’d know they had money to rebuild. Her mom wants to move back to the property—to get out of the city where she’s taken refuge and back up into the hills where the birds and frogs sing. They’re considering building some kind of yurt or tiny house. But they’re facing a long road back to any kind of stability. “I don’t know what the answer is. But I sure wish there was more support,” Adams told me. Her voice began to wobble, then break. “It really would have been nice if my mom had insurance. And she couldn’t get it.”

In the coming decades, as climate change makes disasters more likely, Americans will need to decide how to approach situations like this. The solutions don’t have to mean clearing whole areas of people altogether. Kousky said that, in the case of floods, she’s seen proposals to offer lower premiums only to the people who really need it—while forcing more affluent families to pay the full price to live in these zones. She told me that she hasn’t seen that policy suggested for wildfire insurance yet. The reality, though, is that people will continue to live in places like Cohasset, even if it means taking the risk that a fire could burn through their life and leave them scrambling for a way to recover.

The Worst Feature Apple Ever Made

2024-08-04T07:00:00-04:00

One Saturday last month, I had a perfect day. I woke early, drove up the California coastline, and surfed for a couple of hours with friends. Then I met up with another friend nearby, went kayaking, and ate a late lunch. After that—sun-worn and salty—I drove home, washed off my gear, walked the dog, and ate pizza on my couch.

A big part of what made the day so perfect was all the time spent outside—away from work deadlines, chores, and screens. Yet despite my best effort to escape, I still logged six hours of screen time, more than usual. Two hours and 52 minutes of that was spent on Google Maps. An hour came from texting. I also spent 45 minutes on Safari (shopping for a dress for an upcoming wedding), 24 minutes on Spotify (listening to music), and 10 minutes on Venmo (paying some friends for recent meals). None of this was a bad use of my phone—it wasn’t like I was doomscrolling. Still, when I saw the total number that evening, after checking my iPhone’s Screen Time tool, I couldn’t help but feel a reflexive jolt of guilt. Six hours?

Screen Time is a curious thing: an Apple feature designed to help people be more mindful about using their Apple gadget. First launched in 2018, Screen Time provides daily and weekly reports on how long you’re spending on your iPhone or iPad, broken down by app. After opting in to Screen Time, you’re likely to encounter what I call the Sunday-morning guilt trip, a weekly recap delivered as a push notification. “Your screen time was up 20 percent last week,” it might say, “for an average of 4 hours, 15 minutes a day.” Screen Time also lets you set limits on specific apps—say, restricting TikTok use to just 20 minutes a day.

Apple has championed Screen Time as a way for people to “take control” of their phone usage in this age of screen anxiety—an attempt to reassure customers that Apple is working in their favor. At no point does Screen Time ever outright tell you to consider putting down your phone, but the implication is clear: Ideally, you want your weekly screen-time numbers to be trending down, not up. People find themselves reaching for their phone at every idle moment, potentially wasting hours watching cat videos on Instagram. Of late, concerns about phone addiction have only escalated. A recent best-selling book by the NYU sociologist Jonathan Haidt blames phones, in part, for creating an “anxious generation,” and last month, the surgeon general called for social-media apps to have a tobacco-esque warning label.

[Read: Instagram is not a cigarette]

The problem is that Screen Time—the Apple tool, and the broader fixation—doesn’t seem to help. The main issue is that it flattens phone usage into a single number. “We treat screen time as this unitary experience,” Nicholas Allen, a psychologist at the University of Oregon and the director of its Center for Digital Mental Health, told me. “And of course, it’s an incredibly diverse experience. It can be everything from finding out useful information, to being bullied, to catching up on the news, to watching pornography, to connecting with a friend.”

When it comes to the health consequences of phones, so much depends on context. How someone uses an app matters, as well which app. One person might use Instagram to message with friends, whereas another could just scroll their feed aimlessly, feeling worse about themselves. “If I just say, ‘How much time do you spend on social media?,’ I don’t get the nuance,” David Bickham, the research director at the Digital Wellness Lab at Boston Children’s Hospital, told me. Scrolling through your camera roll is fun if you’re looking at vacation photos; it’s maybe not so great if you’re obsessing over pictures of your ex.

Much of the concern about screen time is about one specific kind: social media. Parents in particular worry that a recent spike in adolescent anxiety and depression is the result of too much scrolling Instagram or TikTok and not enough hanging out in person. (Famously, Facebook’s own leaked internal research found that Instagram can harm teen girls’ body image.) But the research focusing on teens specifically is hotly contested. One study found that the relationship between digital tech use and teen mental health is “negative but small”—too small to guide public policy. The effects on adults are murky too: One meta-analysis of more than 200 studies on well-being and social-media use—studies that spanned countries and age groups—found only small correlations, which varied based on demographics, location, and the type of use.

Instead of fixating on time, experts I spoke with recommend reflecting on how certain applications make you feel. “Really, the best thing is to get people to reflect and be aware of, Oh my God, I’m doomscrolling here,” Allen said. The one exception both Allen and Bickham made was sleep: No matter what you’re doing on your phone, if it’s interrupting your sleep, you’re better off putting down the device and snoozing.

Screen Time is just a tool, of course. It’s up to people themselves to moderate their phone usage. But it is an imperfect tool. Screen Time can be used to put a time limit on an app, but it is too easy to bypass. When a time limit is reached, a tool issues a pop-up warning—but then offers to add time to the limit, or to suspend it indefinitely. Getting back on the app takes just a few taps (and maybe entering a password). Over email, an Apple spokesperson did not answer my question about whether Apple has any evidence that Screen Time actually helps people cut back on phone usage.

Apple is in a weird spot. The company that makes smartphones and oversees the App Store doesn’t exactly have a good reason to tell you to stop tapping. Screen Time is just one especially popular tool in a whole anti-smartphone ecosystem—technology to fix the problem of using technology too much. Google also has its own set of screen-time-reduction tools for Android, called Digital Wellbeing, the design of which is similar to Apple’s.

While reporting this story, I tried five other screen-time apps: Opal, ClearSpace, OffScreen, ScreenZen, and Freedom. In addition to apps, there are dumb phones that have only basic functionality, and boxes you can lock your phone in . A company called Brick makes a physical device—a gray square—that, when scanned, blocks unwanted apps. You can hide the device or place it across the room, so that you have to walk over to regain access. YouTubers make videos about how to redesign your iPhone home screen to minimize distraction.

Some of these tools seem to work better than Screen Time. They block you from being able to open a distracting app outright, or force you to wait five seconds or take a deep breath before launching whatever it is you tapped on. But there are no easy answers here. A lot of the concerns around phones have focused on teens, nuance that sometimes gets lost: “Do not confuse the conversations about phones being bad for 15-year-olds with phones being bad for grown adults,” Katie Notopoulos wrote in Business Insider this spring.

Screen Time and the whole ecosystem of tools like it reinforce the vague sense that everyone should be using their phone less, even if we’re not exactly sure why. The problem with the smartphone is also its greatest achievement: The device squishes an enormous amount of capability into the palm of your hand. So much of it is necessary. So much of it is a waste. People do have good reasons to cut down on phone usage. Smartphones can distract us, overwhelm us, spoil our mood, and even mess with our posture and eyesight. But the tortured relationship that people have with their screens doesn’t get better if you simply remind people that they have a tortured relationship with their screens. No one needs to be made to feel guilty for using Google Maps or streaming a YouTube exercise class or texting their parents a picture of their dog.

The truth is, the perfect day can involve using your phone a lot. And that’s okay.

Google Wins the Gold Medal for Worst Olympic Ad

2024-07-30T14:42:55-04:00

Google is running a new commercial during the Olympics. It’s about a cute little girl—she’s a runner, and she loves Team USA’s Sydney McLaughlin-Levrone, a world-record-holding track star who won two Olympic gold medals in 2021. The little girl wants to write her a letter. So Dad fires up an AI chatbot.

“Help my daughter write a letter telling Sydney McLaughlin-Levrone how inspiring she is,” he asks Google’s Gemini. He instructs it to add a line about how his daughter plans to break McLaughlin-Levrone’s world record one day (and to be sure to include the phrase sorry, not sorry.) The ad never shows the final letter in full, just pans over snippets of it. The whole thing is supposed to be endearing, demonstrating to viewers how AI can help forge human connection and facilitate creativity.

But come on: Nothing about this ad makes sense.

Isn’t what makes a letter like this cute the fact that it is written by a child? Shouldn’t a young person get to explore their feelings and then authentically relay them? And what about McLaughlin-Levrone? Will she be able to tell that the letter was written by AI? How would she feel about that? Would she send back her own AI-composed message, thanking the child for taking the time to write to her?

[Read: The AI industry is stuck on one very specific way to use a chatbot]

The whole thing is bleak. It takes the feel-good cliché of a child getting to interact with their idol and squishes a multimillion-dollar large language model between them. Google is pitching a world in which even the most personal interactions are mediated by computers. The company may make bold claims about AI’s capabilities to radically advance civilization. But it can’t escape the reality that it’s co-opting the hopeful aesthetics of the Olympics, which are meant to celebrate human accomplishments, in order to promote a digital technology that can be used to undermine human labor.

The reaction so far has not been positive. The author Will Leitch said the ad “takes a little chunk out of my soul every time I see it.” The professor and media personality Shelly Palmer wrote that it makes him “want to scream.” On YouTube, where Google posted the ad four days ago, comments are turned off—a step that the company does not typically take on its videos, and one that suggests concern about a backlash.

When I reached out to Google to ask about that backlash, a spokesperson told me, “We believe that AI can be a great tool for enhancing human creativity, but can never replace it. Our goal was to create an authentic story celebrating Team USA.” The ad, which the spokesperson said features a real father and daughter, “aims to show how the Gemini app can provide a starting point, thought starter, or early draft for someone looking for ideas for their writing.”

Google’s marketing team has a tough job right now. The company has aggressively pivoted to a technology that may be dazzling, but that many people remain skeptical of. AI will revolutionize everything, boosters say, but it’s still unclear exactly how. Wall Street is starting to wonder whether investments in the technology will actually pay off. To the extent that generative AI is present in everyday life, it’s not always on the best terms: The technology has arguably degraded once-reliable search engines, plundered human creativity, and taken jobs.

All of which is to say, the reality is far from the sunshine and jump ropes of the “Dear Sydney” ad. Perhaps that’s not unusual: For years, Big Tech’s marketing has relied on sweet montages of regular people using their tools to skirt the very real problems presented by their products. The likes of Meta, TikTok, and Apple may be able to get away with this framing, because their products do connect people at the end of the day, but generative AI is more about humans talking to a computer instead of one another. (Apple found itself in a similar situation earlier this year with an iPad ad that, accidentally or not, evoked AI’s ability to crush art with a machine; Apple quickly apologized and halted plans to run the commercial on TV.)

[Read: Watch Apple trash-compact human culture]

Google appears to have misread the moment. The Olympics are supposed to be about humans accomplishing amazing feats in the physical world. While the ad was running this weekend, the American surfer Caroline Marks scored a close-to-perfect 9.43 points out of 10 after dropping into the barrel of a giant wave in Tahiti. The 17-year-old Canadian swimming prodigy Summer McIntosh won her first gold medal for the women’s 400-meter individual medley. And the legendary gymnast Simone Biles continued to defy the laws of physics despite an injury. These athletes are indeed inspiring. We don’t need a chatbot to tell them so.

California’s Fire Luck Just Ran Out

2024-07-27T10:33:36-04:00

For two years, the fire gods cut California a break. The winter rains came down heavy and brought the state’s yearslong drought to an end. Plants started growing again. Grasses were green. The poppies bloomed larger than normal. For awhile, living here meant seeing the place’s better nature—going outside and exploring the mountains and lakes and vineyards without thinking of breathing in toxic smoke plumes. The apocalyptic scenes of 2020 and 2021 receded like a bad dream; any worries about fire were a problem of the past, or the future.

Then the heat came, and the green faded. Plants died. People who know where to look started to see the warning signs. Now when David Acuna, a battalion chief at Cal Fire, walks around his local area, he sees layers of grass: standing grass, but also the remnants of previous years’ grasses. “They are just waiting to burn,” he told me yesterday. Wildfire is cyclical, and wet years can set up future ones for worse fires. Even when the landscape is lush and healthy, California is operating on borrowed time.

This week, fire came roaring back. California’s first major fire in three years is burning. The Park Fire, located near the city of Chico in Northern California, started Wednesday and grew quickly, tripling in size in a single day. By this morning, the blaze, which started when a man allegedly rolled a burning car into a gully, had spread across more than 300,000 acres and was zero percent contained. Already it is one of the 10 largest recorded fires in California history, and it is moving extremely fast. “We had our fire grow by 120,000 acres in a single day,” Acuna said. “That is not normal.”

Fire is a natural part of California’s ecosystem, and can help clear space for new plant life. But in the past 10 years, the combination of dry fuels, hot temperatures, and winds have made for more explosive fire growth, according to Dan Macon, a UC Cooperative Extension livestock and natural-resources adviser who monitors the grass conditions in the area just south of where the Park Fire is. “When I was a kid, a big fire was 5,000 acres,” he told me. Abnormally hot weather, in particular, may be helping feed bigger and more violent fires. One paper tried to isolate the role of climate change in California’s wildfires over the past 50 years and found that human-caused warming was responsible for almost all of the increase in acreage burnt.

Those exact dynamics seem to be driving the current fire. California’s two consecutive wet springs, in 2023 and 2024, left the state with a lot of extra vegetation—or, as wildfire experts call it, fuel. Extreme heat early this summer dried all that fuel out: One heat wave around the Fourth of July drove temperatures up to, or past, 110 degrees in parts of the state. Conditions are bad right now, and fire activity has picked up accordingly. The state’s five-year average for acres burned by this time of year is about 117,000 acres, Acuna said. This year, some 467,000 acres, more than three times what’s normal, have already been scorched. Matthew Shameson, a meteorologist at the U.S. Forest Service, told me he and his colleagues expect above-average fire activity to continue for much of the state through September.

None of that means that this particular fire, at this particular time, was inevitable. Last year could’ve been a bad one—Acuna, with Cal Fire, told me he’d braced for that—but it ended up being relatively quiet. California got lucky. And even the biggest fires can start by chance: The majority of wildfires in the U.S. are caused by humans, as is the case with the Park Fire, though in many cases the spark is less dramatic—a runaway camp fire or a misplaced cigarette butt. (The man who allegedly started this blaze is under arrest.) The second biggest cause is lightning.

The odds that California—and the rest of the West—get any lucky breaks this year seem low. It’s only July. National firefighting resources are already strained, and “we’ve still got a lot of dry, hot weather ahead of us,” Macon pointed out. People living in the West know to expect fire, even if we try to forget it during green seasons and years of reprieve. But the breaks always end. The Park Fire is eerily close to the site of the Camp Fire, which killed 85 people in 2018. Parts of Paradise, a town that’s still recovering from that fire, are under evacuation warning.

Just this week, two other fires burned through Canada’s Jasper National Park, where people flock to bathe in impressive forests and cliffsides, to feel humbled by the marvels around them. Living in this part of the world means living amid beauty. And it means forever waiting for the moment when all that beauty goes up in smoke.

A New Development in the Debate About Instagram and Teens

2024-07-17T10:00:10-04:00

The teens are on Instagram. That much is obvious. A majority of teens say they use the app, including 8 percent who say they use it “almost constantly,” according to the Pew Research Center. And yet a lot is still unknown about what such extensive use might do to kids. Many people believe that it and other social-media apps are contributing to a teen mental-health crisis.

Now, after years of contentious relationships with academic researchers, Meta is opening a small pilot program that would allow a handful of them to access Instagram data for up to about six months in order to study the app’s effect on the well-being of teens and young adults. The company will announce today that it is seeking proposals that focus on certain research areas—investigating whether social-media use is associated with different effects in different regions of the world, for example—and that it plans to accept up to seven submissions. Once approved, researchers will be able to access relevant data from study participants—how many accounts they follow, for example, or how much they use Instagram and when. Meta has said that certain types of data will be off-limits, such as user-demographic information and the content of media published by users; a full list of eligible data is forthcoming, and it is as yet unclear whether internal information related to ads that are served to users or Instagram’s content-sorting algorithm, for example, might be provided. The program is being run in partnership with the Center for Open Science, or COS, a nonprofit. Researchers, not Meta, will be responsible for recruiting the teens, and will be required to get parental consent and take privacy precautions. Meta shared details about the initiative exclusively with The Atlantic ahead of the announcement.

The project cracks open the door for greater insights into social media’s effects—yet some researchers are nevertheless regarding it with trepidation. Like many online platforms, Instagram is essentially a black box, which has made it difficult for outsiders to draw direct links between the app and its possible effects on mental-health. “We consider ourselves to be in a very difficult and unusual situation, which is [that] the social-media companies have treasure troves of data that no academic researcher will ever amass on their own,” Holden Thorp, the editor in chief of Science, which published studies about the 2020 election in collaboration with Meta, told me. “So you have potentially a resource that could answer questions that you can’t answer any other way.”

[Read: No one knows exactly what social media is doing to teens]

Part of the reason this feels particularly fraught is that leaks from within Meta have indicated that the company has conducted its own research into the harms of its products. In 2021, documents released by the whistleblower Frances Haugen showed that the company’s own research has repeatedly found that Instagram can harm teenagers, especially teenage girls. “Almost no one outside of Facebook knows what happens inside of Facebook,” Haugen said in congressional testimony that year. (Meta was previously known as Facebook, which it owns; the company rebranded just a few weeks after Haugen’s appearance.) Later in her testimony, she said that “there is a broad swath of research that supports the idea that usage of social media amplifies the risk” of mental-health issues such as depression. Before that, Facebook became notorious among researchers for restricting their ability to study the site, including one high-profile incident in 2021, in which it kicked a group of researchers from New York University off the platform.

All of which underscores the value of independent research: The stakes are high, but the actual data are limited. Existing experimental research has produced mixed results, in part because of the issues around access. In the meantime, the idea that social media is harmful has calcified. Last month, the U.S. surgeon general proposed putting a cigarette-style warning label on social sites—to serve as a reminder to parents that they haven’t been proved safe. Cities and school districts across the country are busy passing rules and legislation to restrict the use of devices in the classroom.

[Read: Get phones out of schools now]

It is against this backdrop that Meta has decided to loosen its grip, however slightly. “As this topic has heated up, we have felt like we needed to find a way to share data in a responsible way, in a privacy-preserving way,” Curtiss Cobb, a vice president of research at Meta, told me. “It’s reasonable for people to have these questions. If we have the data that can illuminate it, and it can be shared in a responsible way, it’s in all of our interests to do that.”

Outside experts I talked with had mixed opinions on the project. Thorp pointed out that Meta has ultimate control over the data that are handed over. Candice Odgers, a psychologist at UC Irvine who studies the effects of technology on adolescent mental health and has written on the subject for The Atlantic, said the pilot program is a decent, if limited, first step. “Scientifically, I think this is a critical step in the right direction as it offers a potentially open and transparent way of testing how social media may be impacting adolescents’ well-being and lives,” she told me. “It can help to ensure that science is conducted in the light of day, by having researchers preregister their findings and openly share their code, data, and results for others to replicate.” Researchers have long called for more data sharing from Meta, Odgers noted. “This announcement represents one step forward, although they can, and should, certainly do more.”

Notably, Meta has been a complicated research partner for similar projects in the past. The political-partisanship studies published in Science came from a kindred program, though its design was slightly different; Meta served a bigger role as a research partner. As The Wall Street Journal reported, the company and researchers ended up disagreeing on the work’s conclusions before the studies were even published. The studies were ultimately inconclusive about Facebook’s ability to drive partisanship in U.S. elections, though Meta positioned them as adding “to a growing body of research showing there is little evidence that key features of Meta’s platforms alone” cause partisanship or change in political attitudes.

Cobb told me that Meta has eliminated some of the problems with the 2020 election project by introducing a technique called “registered reports.” This, he said, will avoid some later back-and-forth over interpretations of the results that cropped up last time: Would-be researchers will be required to get their processes peer-reviewed upfront, and the results will be published regardless of outcome. Cobb also noted that Meta won’t be a research collaborator on the work, as it was in 2020. “We’re just going to be providing the data,” he explained. (The company is funding this research through a grant to the COS.)

Meta, for its part, has also framed the project as one that could later be built upon if it’s successful. Perhaps it’s best understood as a baby step forward in the direction of data transparency—and a much needed one at that.

Hot AI Jesus Is Huge on Facebook

2024-06-28T07:00:00-04:00

Jesus is punching the devil on Facebook.

The two are in a boxing ring. Jesus is wearing a pair of white boxing shorts with his name embroidered on the waistband. He is ripped beyond belief; not only does he have six-pack abs, but every muscle on his body is bulging. Jesus is hitting the devil directly on the chin, a knockout blow. “Nunca te arrepentiras de darle me gusta a esta foto”—“You will never regret liking this photo”—the caption reads in part, followed by a bunch of spam hashtags. The post has more than 600,000 likes.

In another image, Jesus has icy-blue eyes. A bloody cross adorns his forehead. He looks like the actor Jared Leto. This one has more than 240,000 likes. It is just one of hundreds of variations posted by a single page; in another, he wears a large Coachella-esque flower crown.

Hot AI Jesus hath risen. The son of God, as rendered by modern artificial intelligence, is chiseled and has startlingly good hair. (He is not to be confused with Shrimp Jesus, another AI-generated variant.) These depictions of Christ are at times extremely popular on Facebook and Instagram. Jesus, hot or not, is a significant motif in this era of online AI junk; he is to AI Facebook spam as water lilies are to Monet, and dancers to Degas. Spend enough time scrolling the AI wastelands of social media, and you will likely encounter him, in all his glory. He raises a number of questions about social media, religion, and art, the most basic one being: Why on earth does AI present the son of God as such a smoke show?

[Read: Generative art is stupid]

That one’s actually the easiest to answer. As I’ve written before, AI image tools tend to create good-looking people by default. “Ask [AI to generate] anybody,” Hany Farid, a professor at the UC Berkeley School of Information, told me. “Ask for a professor, an engineer, a plumber, an electrician, a firefighter, a police officer, a nurse, anything.” The resulting images are usually gorgeous. That’s likely because the data sets these tools are trained on are biased toward hotness: Photographs of celebrities are widely available and are thus overrepresented in digital-image libraries. To the extent that generative AI may be trained on social-media posts, well: People tend to post flattering photographs of themselves online. But Farid doesn’t think that these are the only explanations. He told me that there may be an algorithmic feedback loop at play, that individual users of generative-AI tools tend to select the most visually appealing outputs, reinforcing these options as “correct.” Or maybe tech companies have intentionally designed image-generating products that make hot people, because people tend to like photos of hot people. In any case, the bias is real: Adobe previously told me that it had noticed in its model this drift toward hotness (and works to de-bias it accordingly); OpenAI has similarly acknowledged that its DALL-E tool has this issue.

(Jesus isn’t the only religious figure who is showing up on Instagram looking like an influencer: One account dedicated to “creating unique visuals of Catholic saints” serves up images of what looks like the cast of some yet-to-be-announced Game of Thrones spin-off featuring only really, really ridiculously good-looking people. Like most AI religious creations I encountered, the saints are almost all white, despite long-running debates about the whitewashing of such figures.)

Hot Jesus appears to be catnip for users on Facebook, where he is routinely posted to generate engagement. Many of these posts are accompanied by a demanding caption. “Why don’t pictures like this ever trend?” they ask over and over, almost threateningly. The faithful are challenged to comment “Amen.” And many accounts do. But not all of these comments are necessarily left by real people. Jason Koebler, a journalist at and co-founder of the technology-news site 404 Media (and one of the world’s foremost chroniclers of bizarre Facebook AI art), tried an experiment: He messaged about 300 accounts who’d commented on AI-generated-art posts, and netted only four replies, suggesting that at least some of this engagement may be from bots. Typically, the more an image is engaged with on social media, the more likely a platform’s algorithm is to show it to more people; popularity begets popularity. Koebler suspects that the photos are propelled by bots, which are programmed to react to the images; the engagement makes the images more likely to be shown to more Facebook users, presumably including some substantial number of actual humans. The mysterious people running these AI-junk Facebook pages must have some financial incentive to create this spam, though it’s unclear precisely how they’re profiting from them. Josh Goldstein, a fellow at Georgetown who co-wrote a research paper about these types of pages, told me that he and his co-author suspect that these spammers build big audiences and then leverage those eyeballs to generate revenue, perhaps by posting links to ad-laden junk websites.

[Read: Jesus of the small screen]

When I reached out to Meta to ask whether Hot Jesus violates their content policies—or whether the company has any insight into how much of the engagement with the images is real—it did not respond. The company allows but does not require users to disclose when images are made with AI; however, images may automatically be labeled as such if the company’s systems detect that they were AI-generated. Meta, more broadly, isn’t anti–AI art, it’s building AI art tools within Instagram and its Meta AI chatbot. (Meta’s AI, however, refuses to generate Jesus images: “I can’t generate images of religious figures,” it explained to my editor this morning.) “I just don’t know how [Hot Jesus] would violate” the company’s policies, Brian Fishman, a former policy director at Meta who has since co-founded a trust-and-safety platform called Cinder, told me. He explained that “these kind of allegorical images aren’t exactly misinformation, even if folks find them distasteful.”

Is Hot AI Jesus distasteful? The images build on a history of American evangelical tradition, David Morgan, a professor of religious studies at Duke and the author of The Forge of Vision: A Visual History of Modern Christianity, told me. Billy Sunday, a famous athlete turned preacher in the early 1900s, would shadowbox the devil onstage—presaging the AI-generated image of Jesus knocking Satan out. As far back as the ’60s, Jesus has been depicted as aggressively muscular; hypermasculine representations also emerged in the ’90s and 2000s. I asked Morgan whether he saw Hot Jesus as offensive, and he told me he’d given up judging when Mel Gibson’s The Passion of the Christ came out. He would ask evangelicals how they felt about the brutality of that film, which some critics likened to pornography. “They see it through a very thick set of theological glasses,” he explained, “that transform the violence—transform the machismo—into a kind of triumphal declaration of American manhood.”

[From the December 1923 issue: Adventures in Christianity]

Like Morgan, Kristin Kobes Du Mez, a historian and the author of Jesus and John Wayne, points to 20th-century fears that Jesus had gotten too soft and feminine; Warner Sallman’s Head of Christ, in which Jesus looks especially gentle, fanned these anxieties. People started saying, “We need a more rugged, masculine Jesus,” Du Mez told me. “And that’s when you had the turn toward the more kind of warrior motif.” Hot AI Jesus almost feels like an amalgamation of both traditions: warrior Jesus and beautiful Jesus. Perhaps these AI tools are picking up on both themes within their data sets, and supercharging them.

Jesus, in some ways, is always a reflection of the culture of the day. So it’s only natural that current depictions of him would adopt the heavy-handed, airbrushed style of AI image generators. The only remaining question is how long he’ll stick around for: Koebler, the reporter, told me he’s already seen some AI-art trends come and go. “Once that type of content falls out of favor, it seems like these fan pages stop making it,” he said. “The one thing that has persisted for months and months is Jesus.” That Hot Jesus has so far demonstrated his staying power in the bowels of zombie-AI Facebook proves, Koebler said, that he is a popular guy, and that there is still money to be made off of him doing virtual battle with the devil. Amen.

Instagram Is Not a Cigarette

2024-06-18T12:43:17-04:00

Many teens and adults use the word addictive when describing social-media sites, as if the apps themselves are laced with nicotine. The U.S. surgeon general, Vivek Murthy, wants to drive that point home as glaringly as possible: In an op-ed published by The New York Times yesterday, he writes that the country should start labeling such sites as if they’re cigarettes.

Murthy proposes putting an official surgeon’s-general warning—the same type found on tobacco and alcohol products—on social-media websites to “regularly remind parents and adolescents that social media has not been proved safe.” Such a warning would require formal congressional approval. To make his case, Murthy cites a 2019 study that found that adolescents who spend more than three hours a day on social media may be at higher risk for certain mental-health problems; he also pointed to research in which teens reported that social media made them feel worse about their body. “The moral test of any society is how well it protects its children,” he writes. “Why is it that we have failed to respond to the harms of social media when they are no less urgent or widespread than those posed by unsafe cars, planes or food?”

It’s a radical idea, and one with a real basis in science: There is strong evidence that tobacco warnings work, David Hammond, a professor in the school of public-health sciences at Canada’s University of Waterloo, told me. Although no intervention is perfect, such labels reduce tobacco use by reaching the right audience at the moment of consumption, Hammond said, and they are particularly effective at deterring young people. But social media is not tobacco. Some platforms have no doubt caused real harm to many children, but research into the effects of social media on young people has been a mixed bag; even the studies cited by Murthy are not as straightforward as presented in the op-ed. A warning label on a pack of cigarettes is attention-grabbing and succinct: No one wants cancer or heart disease. Social media does not boil down as easily.

[Read: No one knows exactly what social media is doing to teens]

What would a social-media warning look like? Murthy doesn’t go into further detail in his article, and nothing would be decided until Congress authorized the label. (It’s unclear how likely it is to pass, but there has been bipartisan interest in the topic, broadly speaking; earlier this year, at a congressional hearing on kid safety on the internet, members from both parties expressed frustration with Big Tech CEOs.) It could be a persistent pop-up that a user has to click out of each time they open an app. Or it could be something that shows up only once, in the footer, when a person creates an account. Or it could be a banner that never goes away. To be effective, Hammond told me, the message must be “salient”—it should be noticeable and presented frequently.

Design may be the easy part. The actual warning text within a social app might be hard to settle on, because an absolute, causal link has not yet been shown between, say, Instagram and the onset of depression; by contrast, we know that smoking causes cancer, and why it does so. “One of the reasons that we have such a wide range of opinions is that the work still isn’t quite conclusive,” David S. Bickham of the Digital Wellness Lab at Boston Children’s Hospital, whose research on body image was cited in Murthy’s op-ed, told me. One major meta-analysis (a study of studies) found that the effect of digital technology on adolescent well-being was “negative but small”—“too small to warrant policy change.” (That paper has since been critiqued by researchers including Jean Twenge and Jonathan Haidt, who have contributed writing about teen smartphone use to The Atlantic; they argue that the study’s methodology resulted in an “underestimation” of the problem. The authors of the original study then “rejected” these critiques by providing additional analysis. And so this goes.) The very fact that there is so much debate doesn’t make for neat public-health recommendations.

In the absence of a firm conclusion, you can imagine a label that would use hedged language—“This app may have a negative effect on teens’ mental health depending on how it’s used,” for example—though such a diluted label may not be useful. I asked Devorah Heitner, the author of Growing Up in Public: Coming of Age in a Digital World, what she would recommend. For starters, she said, any warning should include a line about how lack of sleep harms kids (a problem to which late-night social-media use may contribute). She also suggested that the warning might address young people directly: “If I were going to put something on a label, it would be, like, ‘Hey, this can intensify any feelings you might already be having, so just be thoughtful about: Is this actually making me feel good? If it’s making me feel bad, I should probably put it away.”

[Read: End the phone-based childhood now]

If Murthy’s label does become a reality, another challenge will be figuring out what constitutes social media in the first place. We tend to think of the social web as a specific set of apps, including Facebook, Instagram, Snapchat, and TikTok. But plenty of sites with social components may fall into this category. Murthy papers over this challenge somewhat in his op-ed. When he writes, “Adolescents who spend more than three hours a day on social media face double the risk of anxiety and depression symptoms,” he is referring to a study that asked teens only whether they use “social networks like Facebook, Google Plus, YouTube, MySpace, Linkedin, Twitter, Tumblr, Instagram, Pinterest, or Snapchat.” These platforms do not all have a lot in common, and the study does not draw any definitive conclusions about why using such platforms might be associated with an increased risk of mental-health problems. Murthy’s proposal doesn’t make clear which sites would be required to declare that they are associated with negative health outcomes. Would Roblox or Fortnite qualify? Or a newspaper with a particularly vibrant comments section?

Practical concerns aside, experts I spoke with also worried that the label puts the onus on kids and their parents rather than on the technology companies that make these sites. This is something Murthy acknowledges in his essay, noting that labeling alone won’t make social media safe for kids. “I don’t want the labels to let the social-media companies off the hook, right? Like, Oh, well, we labeled our harmful thing,” Heitner said. In other words, a warning alone may not solve whatever problems social apps might be causing.

[Read: The panic over smartphones doesn’t help teens]

Murthy’s proposal comes at a time when parents seem especially desperate to keep teens safe online. Haidt’s latest book about smartphones and kids, The Anxious Generation, has been on the New York Times best-seller list for weeks. Haidt told me over email that he applauds the surgeon general for calling for such labels: “We as a country are generally careful about the consumer products and medications that harm small numbers of children. Yet we have done nothing, absolutely nothing, ever, to protect children from the main consumer product they use every day.”

People are frightened. But fear isn’t always the best way to help young people. “The science simply does not support this action and issuing advisories based on fear will only weaken our trust in the institutions that wield them in this way,” Candice L. Odgers, a psychology professor at UC Irvine who studies how adolescents use digital technology (and recently wrote her own article on social-media panic for The Atlantic), told me over email. “It is time to have a real conversation about adolescent mental health in this country versus simply scapegoating social media.”

A Generation of AI Guinea Pigs

2024-06-14T15:33:16-04:00

This spring, the Los Angeles Unified School District—the second-largest public school district in the United States—introduced students and parents to a new “educational friend” named Ed. A learning platform that includes a chatbot represented by a small illustration of a smiling sun, Ed is being tested in 100 schools within the district and is accessible at all hours through a website. It can answer questions about a child’s courses, grades, and attendance, and point users to optional activities.

As Superintendent Alberto M. Carvalho put it to me, “AI is here to stay. If you don’t master it, it will master you.” Carvalho says he wants to empower teachers and students to learn to use AI safely. Rather than “keep these assets permanently locked away,” the district has opted to “sensitize our students and the adults around them to the benefits, but also the challenges, the risks.” Ed is just one manifestation of that philosophy; the school district also has a mandatory Digital Citizenship in the Age of AI course for students ages 13 and up.

Ed is, according to three first graders I spoke with this week at Alta Loma Elementary School, very good. They especially like it when Ed awards them gold stars for completing exercises. But even as they use the program, they don’t quite understand it. When I asked them if they know what AI is, they demurred. One asked me if it was a supersmart robot.

Children are once again serving as beta testers for a new generation of digital tech, just as they did in the early days of social media. Different age groups will experience AI in different ways—the smallest children may hear bedtime stories generated via ChatGPT by their parents, while older teens may run into chatbots on the apps they use every day—but this is now the reality. A confusing, sometimes inspiring, and frequently problematic technology is here and rewiring online life.

Kids can encounter AI in plenty of places. Companies such as Google, Apple, and Meta are interweaving generative-AI models into products such as Google Search, iOS, and Instagram. Snapchat—an app that has been used by 60 percent of all American teens and comparatively few older adults—offers a chatbot called My AI, an iteration of ChatGPT that had purportedly been used by more than 150 million people as of last June. Chromebooks, the relatively inexpensive laptops used by tens of millions of K–12 students in schools nationwide, are getting AI upgrades. Get-rich-quick hustlers are already using AI to make and post synthetic videos for kids on YouTube, which they can then monetize.

Whatever AI is actually good for, kids will probably be the ones to figure it out. They will also be the ones to experience some of its worst effects. “It is kind of a social fact of nature that kids will be more experimental and drive a lot of the innovation” in how new tech is used culturally, Mizuko Ito, a longtime researcher of kids and technology at UC Irvine, told me. “It’s also a social fact of nature that grown-ups will kind of panic and judge and try to limit.”

That may be understandable. Parents and educators have worried about kids leaning on these tools for schoolwork. Those who use OpenAI’s ChatGPT say that they are three times more likely to use it for schoolwork than search engines like Google, according to one poll. If chatbots can write entire papers in seconds, what’s the point of a take-home essay? How will today’s kids learn how to write? Still another is bad information via bot: AI chatbots can spit out biased responses, or factually incorrect material. Privacy is also an issue; these models need lots and lots of data to work, and already, children’s data have reportedly been used without consent. (The Atlantic has a corporate partnership with OpenAI. The editorial division of The Atlantic operates independently from the business division.)

[Read: This is what it looks like when AI eats the world]

And AI enables new forms of adolescent cruelty. In March, five students were expelled from a Beverly Hills middle school after fake nude photos of their classmates made with generative AI began circulating. (Carvalho told me that L.A. has not seen “anything remotely close to that” incident within his district of more than 540,000 kids.) The New York Times has reported that students using AI to create such media of their classmates has, in fact, become an “epidemic” in schools across the country. In April, top AI companies (including Google, Meta, and OpenAI) committed to new standards to prevent sexual harms against children, including responsibly sourcing their training material to avoid data that could contain child sexual abuse material.

Kids, of course, are not a monolith. Different ages will experience AI differently, and every child is unique. Participants in a recent survey from Common Sense that sought to capture perspectives on generative AI from “teens and young adults”—all of whom were ages 14 to 22—expressed mixed feelings: About 40 percent said they believe that AI will bring both good and bad into their lives in the next decade. The optimistic respondents believe that it will assist them with work, school, and community, as well as supercharge their creativity, while the pessimistic ones are worried about losing jobs to AI, copyright violations, misinformation, and—yes—the technology “taking over the world.”

But I’ve wondered especially about the youngest kids who may encounter AI without any real concept of what it is. For them, the line between what media are real and what aren’t is already blurry. When it comes to smart speakers, for example, “really young kids might think, Oh, there’s a little person in that box talking to me,” Heather Kirkorian, the director of the Cognitive Development and Media Lab at the University of Wisconsin at Madison, told me. Even more humanlike AI could further blur the lines for them, says Ying Xu, an education professor at University of Michigan—to the point where some might start talking to other humans the way talk to Alexa: rudely and bossily (well, more rudely and bossily).

[Read: Chatbots will change how we talk to people]

Older children and teens are able to think more concretely, but they may struggle to separate reality from deepfakes, Kirkorian pointed out. Even adults are struggling with the AI-generated stuff—for middle- and high-school kids, that task is still more challenging. “It’s going to be even harder for kids to learn that,” Kirkorian explained, citing the need for more media and digital literacy. Teens in particular may be vulnerable to some of AI’s worst effects, given that they’re possibly some of the biggest users of AI overall.

More than a decade on, adults are still trying to unravel what smartphones and social media did—and are doing—to young people. If anything, anxiety about their effect on childhood and mental health has only grown. The introduction of AI means today’s parents are dealing with multiple waves of tech backlash all at once. (They’re already worried about screen time, cyberbullying, and whatever else—and here comes ChatGPT.) With any new technology, experts generally advise that parents talk with their children about it, and become a trusted partner in their exploration of it. Kids, as experts, can also help us figure out the path forward. “There’s a lot of work happening on AI governance. It’s really great. But where are the children?” Steven Vosloo, a UNICEF policy specialist who helped develop the organization’s AI guidelines, told me over video call. Vosloo argued that kids deserve to be consulted as rules are made about AI. UNICEF has created its own list of nine requirements for “child-centered AI.”

Ito noted one thing that feels distinct from previous moments of technological anxiety: “There’s more anticipatory dread than what I’ve seen in earlier waves of technology.” Young people led the way with phones and social media, leaving adults stuck playing regulatory catch-up in the years that followed. “I think, with AI, it’s almost like the opposite,” she said. “Not much has happened. Everybody’s already panicked.”

Trump’s Big New Megaphone

2024-06-03T19:45:00-04:00

Donald Trump has officially joined TikTok. His first video, posted on Saturday night—his only post so far—is a montage showing the former president making the rounds at a UFC fight in New Jersey. He waves to fans and takes pictures with them while Kid Rock’s “American Bad Ass” plays in the background.

Trump—who has appeared on WrestleMania, perfected his image on reality television, and commanded the world’s attention through a demagogic Twitter account—is made for this. His account already has more than 4 million followers; the account belonging to the Biden campaign, created in February, trails far behind with about 349,000. It is a sign that social media may once again be used for political warfare by a man who poses an existential threat to American democracy; what might be shocking will instead be played for laughs and engagement on a platform that Trump attempted to ban as president.

In 2020, Trump signed an executive order that would have banned TikTok if it didn’t find a U.S. buyer—over concerns that the Communist Party of China might be able to use the app to access “Americans’ personal and proprietary information.” (President Joe Biden eventually revoked that ban, which was never enacted; he later signed a bill that would similarly ban the app if it is not divested from its Chinese parent company.) In March, Trump reportedly softened his stance on a TikTok ban and said that Facebook is an “enemy of the people.”

This change in attitude should surprise no one. Trump’s political career is and always has been a circus; all that seems to matter is whether people are watching, even if they’re filled with disgust. TikTok, which is defined by endlessly scrolling short-form, edited video, will play to his strengths. In the real world, when he speaks at campaign rallies, he is meandering, often devolving into gibberish. Here, all of that can be massaged away in favor of dramatic supercuts, just like the one he posted over the weekend.

Trump is, in a sense, already all over TikTok; MAGA fans are constantly sharing political commentary and remixes of his speeches. By setting up his own account, Trump now has the opportunity to steer the narrative himself even more directly. His first TikTok post came just two days after a jury in New York convicted him on all 34 felony counts against him, which the post makes no mention of.

Trump’s videos, if he continues to post on the platform, will reach a crucial voting population. Polls have shown that Biden is struggling with voters ages 18 to 29—precisely TikTok’s sweet spot. Nearly two-thirds of Americans in that age group say they use the platform, many of them for news. An internal TikTok report found nearly twice as many pro-Trump posts on the platform as pro-Biden posts since November, according to Puck.

These days, it’s perfectly normal for a presidential candidate to join social media—they need to be wherever their constituents are. But remember: This is a candidate who was banned from Twitter and Facebook as a sitting president after his calls to violence around the January 6 riot, though his accounts were later reinstated. He will almost surely test TikTok’s moderation rules. (The company did not immediately respond to a request for comment.) Meanwhile, he will slot himself right in alongside the rest of the app’s standard viral fare, as if all of this is normal, even if it’s very much not.

Google Is Playing a Dangerous Game With AI Search

2024-05-24T16:30:00-04:00

Updated at 11:38 a.m. ET on June 21, 2024

Doctors often have a piece of advice for the rest of us: Don’t Google it. The search giant tends to be the first stop for people hoping to answer every health-related question: Why is my scab oozing? What is this pink bump on my arm? Search for symptoms, and you might click through to WebMD and other sites that can provide an overwhelming possibility of reasons for what’s ailing you. The experience of freaking out about what you find online is so common that researchers have a word for it: cyberchondria.

Google has introduced a new feature that effectively allows it to play doctor itself. Although the search giant has long included snippets of text at the top of its search results, now generative AI is taking things a step further. As of last week, the company is rolling out its “AI overview” feature to everyone in the United States, one of the biggest design changes in recent years. Many Google searches will return an AI-generated answer right underneath the search bar, above any links to outside websites. This includes questions about health. When I searched Can you die from too much caffeine?, Google’s AI overview spit out a four-paragraph answer, citing five sources.

But this is still a chatbot. In just a week, Google users have pointed out all kinds of inaccuracies with the new AI tool. It has reportedly asserted that dogs have played in the NFL and that President Andrew Johnson had 14 degrees from the University of Wisconsin at Madison. Health answers have been no exception; a number of flagrantly wrong or outright weird responses have surfaced. Rocks are safe to eat. Running with scissors can be fine. These search fails can be funny when they are harmless. But when more serious health questions get the AI treatment, Google is playing a risky game.

Google’s AI overviews don’t trigger for every search, and that’s by design. “What laptop should I buy?” is a lower-stakes query than “Do I have cancer?” of course. Even before the introduction of AI search results, Google has said that it treats health queries with special care to surface the most reputable results at the top of the page. “AI overviews are rooted in Google Search’s core quality and safety systems,” a Google spokesperson told me in an email, “and we have an even higher bar for quality in the cases where we do show an AI overview on a health query.” The spokesperson also said that Google tries to show the overview only when the system is most confident in the answer. Otherwise it will just show a regular search result.

When I tested the new tool on more than 100 health-related queries this week, an AI overview popped up for most of them, even the sensitive questions. For real-life inspiration, I used Google’s Trend s, which gave me a sense of what people actually tend to search for on a given health topic. Google’s search bot advised me on how to lose weight, how to get diagnosed with ADHD, what to do if someone’s eyeball is popping out of its socket, whether menstrual-cycle tracking works to prevent pregnancy, how to know if I’m having an allergic reaction, what the weird bump on the back of my arm is, how to know if I’m dying. (Some of the AI responses I found have since changed, or no longer show up.)

Not all the advice seemed bad, to be clear. Signs of a heart attack pulled up an AI overview that basically got it right—chest pain, shortness of breath, lightheadedness—and cited sources such as the Mayo Clinic and the CDC. But health is a sensitive area for a technology giant to be operating what is still an experiment: At the bottom of some AI responses is small text saying that the tool is “for informational purposes only … For medical advice or diagnosis, consult a professional. Generative AI is experimental.” Many health questions contain the potential for real-world harm, if answered even just partially incorrectly. AI responses that stoke anxiety about an illness you don’t have are one thing, but what about results that, say, miss the signs of an allergic reaction?

Even if Google says it is limiting its AI-overviews tool in certain areas, some searches might still slip through the cracks. At times, it would refuse to answer a question, presumably for safety reasons, and then answer a similar version of the same question. For example, Is Ozempic safe? did not unfurl an AI response, but Should I take Ozempic? did. When it came to cancer, the tool was similarly finicky: It would not tell me the symptoms of breast cancer, but when I asked about symptoms of lung and prostate cancer, it obliged. When I tried again later, it reversed course and listed out breast-cancer symptoms for me, too.

Some searches would not result in an AI overview, no matter how I phrased the queries. The tool did not appear for any queries containing the word COVID. It also shut me down when I asked about drugs—fentanyl, cocaine, weed—and sometimes nudged me toward calling a suicide and crisis hotline. This risk with generative AI isn’t just about Google spitting out blatantly wrong, eye-roll-worthy answers. As the AI research scientist Margaret Mitchell tweeted, “This isn't about ‘gotchas,’ this is about pointing out clearly foreseeable harms.” Most people, I hope, should know not to eat rocks. The bigger concern is smaller sourcing and reasoning errors—especially when someone is Googling for an immediate answer, and might be more likely to read nothing more than the AI overview. For instance, it told me that pregnant women could eat sushi as long as it doesn’t contain raw fish. Which is technically true, but basically all sushi has raw fish. When I asked about ADHD, it cited AccreditedSchoolsOnline.org, an irrelevant website about school quality.

When I Googled How effective is chemotherapy?, the AI overview said that the one-year survival rate is 52 percent. That statistic comes from a real scientific paper, but it’s specifically about head and neck cancers, and the survival rate for patients not receiving chemotherapy was far lower. The AI overview confidently bolded and highlighted the stat as if it applied to all cancers.

In certain instances, a search bot might genuinely be helpful. Wading through a huge list of Google search results can be a pain, especially compared with a chatbot response that sums it up for you. The tool might also get better with time. Still, it may never be perfect. At Google’s size, content moderation is incredibly challenging even without generative AI. One Google executive told me last year that 15 percent of daily searches are ones the company has never seen before. Now Google Search is stuck with the same problems that other chatbots have: Companies can create rules about what they should and shouldn’t respond to, but they can’t always be enforced with precision. “Jailbreaking” ChatGPT with creative prompts has become a game in itself. There are so many ways to phrase any given Google search—so many ways to ask questions about your body, your life, your world.

If these AI overviews are seemingly inconsistent for health advice, a space that Google is committed to going above and beyond in, what about all the rest of our searches?

This article originally stated that Google's AI Overview feature told users that chicken is safe to eat at 102 degrees Fahrenheit. This statement was based on a doctored social-media post and has been removed.

The Algorithmic Radicalization of Taylor Swift

2024-05-18T06:00:00-04:00

It is nighttime in Paris. We are more than a year into Taylor Swift’s Eras Tour, and tonight, her fans are once again trying to figure out what her clothes mean.

The star is in a glittering yellow-and-red two-piece set, a possible reference to the colors of the Kansas City Chiefs, the football team Swift’s boyfriend, Travis Kelce, plays on. This is also the 87th performance in the tour, and—aha!—Kelce wears jersey number 87. The hundreds of thousands of fans watching along through bootlegged livestreams on TikTok and YouTube have solved another mystery.

This is the beginning of the European leg of Eras, which will stretch on and on until Swift returns to North America this fall and plays the final show of the tour on December 8 (that is, assuming she doesn’t extend it, as she has multiple times already). You’d think people would have lost interest by now. But Taylor Swift has kept fans’ attention by tapping into an algorithmic machine unlike anyone has before her.

Swift is savvy, and leverages social-media culture to her advantage. Over her 18-year career, she has trained her fandom to inspect everything she does for Easter eggs; she knows that even a small reveal can send people into a frenzy. She likes to leave clues about upcoming music in her outfits, in music videos, even in commercials she films with brands. She knows people are interested in her personal life—her romances, her feuds—and capitalizes on that, leaving them hints in her liner notes or in song titles.

In response, fans analyze dates and look for numbers that add up to 13, her favorite number. They create spreadsheets of every single outfit she’s worn on tour, methodically tracking each surprise song she’s played. They chat nonstop across platforms, swapping elaborate theories to try to decode when the next album is coming or whom each song is about. For more than half a decade, they’ve been convinced that there’s a lost album called Karma, which was shelved in the mid-2010s amid Swift’s feud with Kanye West (now known as Ye) and Kim Kardashian. According to one theory, the orange outfits she’s been wearing in Paris are a sign that she’ll release music from Karma. It’s like QAnon, if QAnon involved a lot of DIY rhinestone boots.

[Read: The real Taylor Swift would never]

Swifties don’t storm the Capitol, but they will flood Kardashian’s Instagram with snake emoji in response to Swift talking about the pain their fight brought her, just as they will fight Ticketmaster when the company botches her concert-ticket rollout. Their thinking is often conspiratorial. In one recent TikTok, a fan argued that Swift would be releasing something on May 3, according to this logic: A recent screenshot of a music-video still posted to Swift’s team’s Instagram included the letter-and-number combination 14.3V—Swift’s latest music video was for “Fortnight,” and a fortnight is two weeks; two weeks is 14 days. One plus four equals five. The three rounds it out: Something’s happening on the 3rd. The V is actually the Roman numeral for five. (May 3 came and went without a release.)

Extreme cliques might be one side effect of our digital culture. “Our algorithms and media are designed to produce fandoms around consumption goods,” Petter Törnberg, a professor of computational social science at the University of Amsterdam, told me over email. “There is hence a fundamental similarity between Swifties, Apple-fans and MAGA Republicans: our current era has the tendency of turning our preferences into identities, and shaping a form of postmodern tribes around both consumption goods and political leaders.” (See also: fans of Beyoncé and BTS.)

In other words: Social platforms can have a radicalizing effect on fandoms. When we study algorithmic radicalization, we tend to do so in the context of politics, but the same systems might also calcify our beliefs about cultural products. Yet we still have a fairly limited understanding of how all of this works. “The very best studies we have are still really struggling to detect effects, because there’s so many challenges when you try to study this stuff,” Chris Bail, the founding director of the Polarization Lab at Duke University, told me.

No one single algorithm powers this fandom. It operates across platforms; in a single day, a Swift fan might stream her music on Spotify, watch her music videos on YouTube, and consume posts about her on TikTok. All of these sites have distinct recommendation systems. Companies also tend to keep these systems a secret, making them hard to research.

But we can say this: Algorithms tend to reinforce what’s already popular, because attention attracts more attention. Growth begets growth, as Törnberg put it. In this way, Swift also demonstrates how platforms that supposedly target content based on an individual’s interests can, in fact, end up clustering around one monolithic force. “It just seems like, Oh, that’s sort of weird, I thought everybody was supposed to have their own algorithmic niche now,” Nick Seaver, the author of Computing Taste: Algorithms and the Makers of Music Recommendation, told me. “And instead—I mean, maybe in addition to that—we also all have Taylor Swift.”

[Read: Nobody knows what’s happening online anymore]

Our modern Swiftocracy is a reminder that we are still subject to strange algorithmic forces, even as the web is supposedly fractured. Yet the consequences of this can be as hard to decode as an Easter egg dropped by Swift. On her final show in Paris, she opted for a “berry”-red dress for the Folklore section of her set. It may be a sign of something to come. Or not.

The End of the ‘Photoshop Fail’

2024-05-09T17:23:43-04:00

In 2017, Rihanna posted a photo of herself on Instagram in which she appeared to have an extra thumb. It was, in retrospect, the thumb-shaped canary in the coal mine. Although far from the first celebrity “Photoshop fail,” it just so happened to predict the era of faux-finger drama we now live in: AI image generators are universally, horrifically bad at rendering human hands. Today, an extra finger is a telltale sign of digital manipulation.

Flaws aside, faking it has never been easier. Advances in generative AI mean that anyone can spin up a faux picture of the pope wearing a chic white puffer, no design skills required. New AI image creators such as Midjourney and Stable Diffusion use sophisticated technology to let users conjure entire worlds from just a few words. Instagram is rolling out AI-editing features; with a couple of taps, an everyday user can place their dog at the foot of the Eiffel Tower. We are living in the world Adobe Photoshop first teased 34 years ago—but it is no longer defined by the enterprise software.

Yet there is a reason Photoshop is the Kleenex of photo manipulation, the rare piece of software so powerful that it became a verb. To Google is to search, and to Photoshop is to sneakily slim a waistline, or to make it seem like something happened that didn’t. The software has been blamed for all kinds of societal ills, such as spreading misinformation and reinforcing unattainable body standards. Our entire modern cultural understanding of photographic trickery is closely entwined with the history of this application. It should be uniquely well suited for this moment.

[Read: AI is about to Photoshop your memories]

From the get-go, Photoshop had ties to unreality. The software was developed in the late 1980s by a pair of brothers, one of whom worked in Hollywood special effects. In early demonstrations, one of the co-creators, John Knoll, would pull up a picture of his then-girlfriend sitting topless in the sand in Bora Bora, looking out over perfect blue waters. He’d select and clone her, creating a second topless woman. When it formally launched in 1990 after being acquired by Adobe, Photoshop quickly differentiated itself from previous photo-editing software. As Walter Scheirer recounts in his book A History of Fake Things on the Internet, the key innovation was that it allowed creatives to make and share “plug-ins” to augment the software, such as Aldus Gallery Effects, a popular set that came with tools including “Splatter” and “Neon Glow.” These helped foster an online community around the software.

Photoshop grew alongside the rise of the consumer internet; people traded presets and add-ons on sites such as DeviantArt. Digital photography was becoming more and more popular, and a cheaper version of the software, Photoshop Elements, offered the average consumer the ability to tweak their digital photos. In the early 2000s, publications began using Photoshop as a verb; by 2008, usage was so widespread that Merriam-Webster had added the word to its dictionary. At this point, Photoshop was at the apex of its cultural power: It popped up in discussions of misinformation and cultural conversations about whether the technology had gone too far. The blog Photoshop Disasters documented hundreds of hiccups among celebrities and advertisements. “Indeed, in a world where so many images of the beautiful and famous are enhanced, ordinary people sometimes believe they need to prettify pictures of themselves just to keep pace,” Alex Williams wrote in The New York Times in 2008.

In the 2010s, smartphones and social media moved photo editing onto apps such as Instagram and later, Facetune. Many professionals continue to use Photoshop, and Adobe still makes a ton of money, but the application is less central in our culture: People have a lot of competing options on their smartphones. Adobe, meanwhile, has continued to innovate. Long before today’s generative AI, Photoshop introduced algorithmic functions that transformed the work of image editing: “Content-Aware” features, for example, react to what is depicted in an image, allowing a user to, say, seamlessly remove a person from a scenic forest view or add additional fluffy cumuli to a cloudy sky.

Photoshop was well positioned for the AI arms race, though it lagged behind image-generation tools such as DALL-E. By the time ChatGPT snapped the world to attention, Adobe was already moving toward putting smaller generative-AI-image features into its products. After that, the company seemed to pick up the pace: The next spring, it launched a full text-to-image model, Firefly, in beta form.

[Read: Kate Middleton and the end of shared reality]

When I asked Ashley Still, the senior vice president and general manager of the Creative Product Group at Adobe, about whether Photoshop is threatened in this new age of fake imagery, she told me “not at all,” and that the program has been consistently growing. Subscriptions, she said, were up 30 percent last year. “What’s happening is, more and more images are being created—more and more people are being creative. And that increases the market for us, too,” she explained. (When asked, representatives for the company did not provide revenue numbers for Photoshop specifically, instead directing me to Adobe’s overall earnings, which are at a record high.)

The company has tried to be “thoughtful about how generative AI can help creatives and make their work more productive,” Still told me. Indeed, the company’s previous algorithmic features saved people a lot of time. (“No one is proselytizing about the dangers of the content-aware tool, when it’s essentially the same technology,” Jordan Wannemacher, a freelance graphic designer, told me.) Whereas other image generators have been drawn into copyright lawsuits, Adobe advertises Firefly as safe for commercial use.

And there’s the challenge: Photoshop, the most professionalized fake-image program, now has to walk a middle path. It needs to offer generative-AI tools to stay cutting-edge, but it also cannot risk alienating its clientele by taking off the guardrails. Adobe is a technology company that serves artists, and in the art world, AI is deeply polarizing.

Still, people have always manipulated photos and created outright fakes, even long before the age of personal computers. “We’re sort of always yearning for authenticity, and it’s sort of also forever eluding us in some ways,” Michael Serazio, the author of The Authenticity Industries: Keeping It “Real” in Media, Culture, and Politics, told me. For the first generation of internet users, Photoshop was the primary battlefield for debates around fake imagery. Now AI is the focus of our anxieties. Surely there will be another thumb-related celebrity scandal in the future—but this time, it will almost certainly be the fault of a bot.

I Witnessed the Future of AI, and It’s a Broken Toy

2024-04-27T06:00:00-04:00

This story was supposed to have a different beginning. You were supposed to hear about how, earlier this week, I attended a splashy launch party for a new AI gadget—the Rabbit R1—in New York City, and then, standing on a windy curb outside the venue, pressed a button on the device to summon an Uber home. Instead, after maybe an hour of getting it set up and fidgeting with it, the connection failed.

The R1 is a bright-orange chunk of a device, with a camera, a mic, and a small screen. Press and hold its single button, ask it a question or give it a command using your voice, and the cute bouncing rabbit on screen will perk up its ears, then talk back to you. It’s theoretically like communicating with ChatGPT through a walkie-talkie. You could ask it to identify a given flower through its camera or play a song based on half-remembered lyrics; you could ask it for an Uber, but it might get hung up on the last step and leave you stranded in Queens.

When I finally got back to my hotel room, I turned on the R1’s camera and held up a cold slice of pizza. “What am I looking at?” I asked. “You are looking at a slice of pizza,” the voice told me. (Correct!) “It looks appetizing and freshly baked.” (Well, no.) I decided to try something else. “What are top 10 …” I stumbled, letting go of the button. I tried again: “What are the top 10 best use cases for AI for a normal person?” The device, perhaps confused by our previous interaction, started listing out pizza toppings beginning with the No. 2. “2. Sausage. 3. Mushrooms. 4. Extra Cheese.”

Until now, consumer AI has largely been defined by software: chatbots such as ChatGPT or the iPhone’s souped-up autocorrect. Now we are experiencing a thingification: Companies are launching and manufacturing actual bits of metal and plastic that are entirely dedicated to AI features. These devices are distinguished from previous AI gadgets, such as the Amazon Echo, in that they incorporate the more advanced generative-AI technology that has recently been in vogue, allowing users more natural interactions. There are pins and pendants and a whole new round of smart glasses.

[Read: Alexa, should we trust you?]

Yet for all its promise, this new era is not going very well. Take Humane, a Rabbit competitor that launched a wearable “AI Pin” earlier this month. That device has been positioned as a smartphone replacement, with a price to match: It costs $699 and requires a $24 monthly subscription fee. Reviewers brutalized the pin, saying it is slow, overheats, and struggles to answer basic queries. “I’m hard-pressed to name a single thing it’s genuinely good at,” The Verge wrote.

By comparison, the R1 is satisfyingly small in its ambition and (relatively) affordable in its price ($199, no subscription). The device itself is fun and retro-chic: Jesse Lyu, Rabbit’s founder and CEO, reportedly bought every member of his team a Tamagotchi for inspiration. And, in fairness, the R1 does some interesting things. Onstage, Lyu showed how the device can interpret a handwritten table and convert it into a working digital spreadsheet. It managed to speak a summary of a handwritten page when I asked, though only with about 65 percent accuracy. I was able to use the gadget to order an acai bowl on DoorDash, although it couldn’t handle any customizations. (I wanted peanut butter.) And I never got Uber to work. (Though at one point, the device told me the request had failed when it in fact hadn’t, leaving me on the hook for a $9 ride I didn’t even take.)

One of the big selling points of the R1 is that it supposedly runs something called a large action model, or LAM—a spin on the phrase large language model, which is the technology powering recent chatbots. Whereas ChatGPT can answer questions and draft you a mediocre essay, the R1 can, in theory, complete actions that you might take on different apps (Venmo-ing your friend $20, for example). Rabbit has said the device will be able to learn any app, if you teach it. Lyu compared the technology to a Tesla: When on autopilot, a Tesla car can in theory recognize a stop sign not because engineers tell it how a stop sign looks but because it has been trained on countless hours of footage to recognize the sign’s physical attributes. Likewise, R1 will be able to accomplish tasks on your phone without having to be taught each app.

The problem is, none of this is actually real. At least not yet. As with so many AI products, the R1 is fueled more by hype than by a persuasive use case. (So many of its functions could, after all, be done on a smartphone.) Back in February, Lyu said the Rabbit was training its model on 800 apps. This week, it launched with the ability to use just four: Spotify, DoorDash, Uber, and Midjourney (a popular AI art generator). The company says LAM is in “very early stages.”

[Read: Phones will never be fun again]

Onstage before an audience of reporters and Rabbit fanboys on Tuesday night, Lyu seemed nervous at times, at one point encouraging people to laugh in order to ease his nerves. Prior to the event, a user had posted on GitHub accusing Rabbit of misrepresenting its technology. “For those with a technical background, it’s painfully clear that there’s no artificial intelligence or large action model in sight,” the anonymous post, which has since been deleted, read. On X, Lyu characterized the post as “all false claims.” Lyu promised to fix any bugs that might crop up in R1 devices. Before demoing DoorDash onstage, he admitted that the feature doesn’t yet work as fast as they’d like it to: “But I want to show you, and I want to be frank with you guys.”

Yet Lyu also breathlessly announced a number of new initiatives, including a high-concept system that would allow people to someday merge the physical and the digital, so people could point at various smart items in their home and control them through Rabbit’s AI. (Never mind that the R1 has launched without many of its promised features.) Toward the end of the presentation, the words Be Humble appeared on the giant screen behind him. “We are a really, really humble team,” Lyu told the crowd. Those words were still displayed when, a few moments later, the curtains on either side of the stage dramatically dropped to reveal conveyor belts loaded with boxes of R1s. Music started blasting, and people started lining up to snatch theirs.

The R1 is a reminder of the disconnect, for better and for worse, between a Silicon Valley culture that often prioritizes speed over quality and high consumer expectations about the products they use. And to be fair, expectations are high at least in part because of the extraordinary products that have emerged from that same competitive and iterative culture over the years.

As the party wound down, news of the first bug arrived: There was no way to change the time zone on the devices, many of which were programmed by default to the West Coast. Turns out the future is stuck three hours behind.

The New Empress of Self-Help Is a TikTok Star

2024-04-17T15:38:15-04:00

In 2006, Oprah Winfrey couldn’t stop talking about The Secret. She devoted multiple episodes of her talk show to the franchise, which started as a kind of DVD seminar and later became a best-selling book. Its author, Rhonda Byrne, claimed to have stumbled upon an ancient principle, one that can teach anyone to manifest anything they want: money, health, better relationships. Winfrey retroactively credited its core philosophy for bringing her success, and her endorsement helped bring the book international fame: It has now sold more than 35 million copies. But in the era of endless scrolling, an author doesn’t necessarily need Winfrey’s stamp of approval. They just need TikTok.

Keila Shaheen figured this out last year, when her self-published book The Shadow Work Journal began to dominate the app’s feeds. A slim volume, the book purports to help people unpack their “shadow” self—the repressed unconscious—through various activities. In video after video, TikTok users show themselves filling out its exercises and talk about the journal as if it has magical powers. They learn about Carl Jung’s model of the psyche. They circle terms related to their trauma. They heal their inner child! If you use a new coupon on TikTok Shop, the app’s new built-in store, you too can heal, for just a couple of bucks! they say. (Many of those posting earn a commission from each sale, but pay that no mind.)

The journal has sold more than 600,000 copies on TikTok alone, and more than 1 million copies in total, a feat usually accomplished by the Prince Harrys and Colleen Hoovers of the world. Shaheen, a 25-year-old writer with a marketing background, is the new breakout star of the self-help genre. She even outsold Winfrey’s latest book.

[Read: The 24-year-old who outsold Oprah this week]

Her story began in an untraditional way: Here is a young author, plucked from obscurity by a powerful app’s algorithm during a conveniently timed e-commerce push and turned into a best-selling phenom. Yet her next chapter is following an expected arc. She has signed a multi-book deal with Simon & Schuster to bring an updated version of The Shadow Work Journal to new audiences. Specifically, she is working with the brand-new imprint Primero Sueño Press, which will launch her book as its “flagship,” Shaheen told me, in addition to releasing a new Spanish translation later this year. The self-help queen of TikTok is officially going mainstream.

Shaheen’s popularity makes a lot of sense. We live in the age of therapy-speak; talking about one’s mental health isn’t as stigmatized as it once was. And yet a lot of people are still struggling. Teenagers—many of whom say they use the app “almost constantly”—are experiencing hopelessness and sadness at record highs. TikTok is known for authenticity, at least when compared with the picture-perfect posts on Instagram—it is supposed to be messier, more real. The kind of place where you’d talk about your struggles while in your sweatpants.

The Shadow Work Journal isn’t the only such success on the platform. One of Shaheen’s other books, The Lucky Girl Journal—which teaches readers how to manifest their own good fortune, rather than leaving things up to chance—has sold more than 25,000 copies on the app’s store. Don’t Believe Everything You Think, a self-published volume by Joseph Nguyen, a mental-health content creator with little notoriety outside social media, has sold about 60,000 copies on TikTok, and is currently in the top 10 most sold books on Amazon.

It’s boom times for self-help on social media. Kathleen Schmidt, who helped publicize The Secret and now runs a public-relations company (and writes the Substack newsletter Publishing Confidential), first heard about Nguyen’s book when her 16-year-old daughter asked for a copy. “I can see why it has caught on,” she told me. “It’s very simplistic, and it gives you big promises, like You’ll stop suffering, you’ll understand how to let go of anxiety, and all that.” A lot of self-help books, she explained, are too complicated or ask the reader to do too much; the more successful books tend to be accessible. If The Secret were published today, she argued, “it probably would have gone viral on TikTok and would have had somewhat of the same effect—but without Oprah.”

[Read: TikTok is doing something very un-TikTok]

With all of this in mind, I asked Shaheen why she’d made the decision to go a more traditional route. It was over Zoom, during a meet and greet set up by her publisher (and attended, as far as I could tell, by just me, one other writer, and some folks from her team). “I think I was just at a time and place where I couldn’t control what was going on,” she said, of all the attention last year, “and it was very overwhelming for me.” She realized that if she “wanted to continue helping people and grow the impact of this journal,” then she “would need help from a traditional publishing company.” She said she’d entertained offers from various publishers before settling on Primero Sueño Press, which will take over the production of her books. And anyway, her books will still be available for purchase on TikTok Shop.

One publishing house she hadn’t heard from is 8th Note Press—which is owned by TikTok’s parent company, ByteDance. It appears to have acquired three titles so far, and published its first book last month; a representative for TikTok told me that it has seen significant growth on TikTok Shop and success for a variety of books and book sellers, but did not comment on Shaheen’s decision to sign with a traditional publisher.

Perhaps ByteDance has a little too much on its plate to prioritize courting authors. TikTok still faces the threat of a national ban in the United States. When defending itself in ads or before Congress, the app likes to tout how many small-business owners it supports—people like Shaheen. For some businesses, that’s definitely true. But with all the uncertainty about the platform’s future, a big, traditional publishing house can offer two things that never feel especially present on social media: stability and security. After all, Simon & Schuster has a pretty good track record. It publishes a little book called The Secret.

You Don’t Have to Type Anymore

2024-04-08T15:35:00-04:00

As a little girl, I often found myself in my family’s basement, doing battle with a dragon. I wasn’t gaming or playing pretend: My dragon was a piece of enterprise voice-dictation software called Dragon Naturally Speaking, launched in 1997 (and purchased by my dad, an early adopter).

As a kid, I was enchanted by the idea of a computer that could type for you. The premise was simple: Wear a headset, pull up the software, and speak. Your words would fill a document on-screen without your hands having to bear the indignity of actually typing. But no matter how much I tried to enunciate, no matter how slowly I spoke, the program simply did not register my tiny, high-pitched voice. The page would stay mostly blank, occasionally transcribing the wrong words. Eventually, I’d get frustrated, give up, and go play with something else.

Much has changed in the intervening decades. Voice recognition—the computer-science term for the ability of a machine to accurately transcribe what is being said—is improving rapidly thanks in part to recent advances in AI. Today, I’m a voice-texting wizard, often dictating obnoxiously long paragraphs on my iPhone to friends and family while walking my dog or driving. I find myself speaking into my phone’s text box all the time now, simply because I feel like it. Apple updated its dictation software last year, and it’s great. So are many other programs. The dream of accurate speech-to-text—long held not just in my parents’ basement but by people all over the world—is coming together. The dragon has nearly been slain.

“All of these things that we’ve been working on are suddenly working,” Mark Hasegawa-Johnson, a professor of electrical and computer engineering at the University of Illinois at Urbana-Champaign, told me. Scientists have been researching speech-recognition tools since at least the mid-20th century; early examples include the IBM Shoebox, a rudimentary computer housed within a wooden box that could measure sounds from a microphone and associate them with 16 different preprogrammed words. By the end of the 1980s, voice-dictation models could process thousands of words. And by the late ’90s, as the personal-computing boom was in full swing, dictation software was beginning to reach consumers. These programs were joined in the 2010s by digital assistants such as Siri, but even these more advanced tools were far from perfect.

“For a long time, we were making gradual, incremental progress, and then suddenly things started to get better much faster,” Hasegawa-Johnson said. Experts pointed me to a few different factors that helped accelerate this technology over the past decade. First, researchers had more digitized speech to work with. Large open-source data sets were compiled, including LibriSpeech, which contains 1,000 hours of recorded speech from public-domain audiobooks. Consumers also started regularly using voice tools such as Alexa and Siri, which likely gave private companies more data to train on. Data are key to quality: The more speech data that a model has access to, the better it can recognize what’s being said—“water,” say, not “daughter” or “squatter.” Models were once trained on just a few thousand hours of speech; now they are trained on a lifetime’s worth.

The models themselves also got more sophisticated as part of larger, industry-wide advancements in machine learning and AI. The rise of end-to-end neural networks—networks that could directly pair audio with words rather than trying to transcribe by breaking them down into syllables—has also accelerated models’ accuracy. And hardware has improved to allow more units of processing power on our personal devices, which allows bigger and fancier models to run in the palm of your hand.

Of course, the tools are not yet perfect. For starters, their quality can depend on who is speaking: Voice-recognition models have been found to have higher error rates for Black speakers compared with white speakers, and they also sometimes struggle to understand people with dysarthric, or irregular, speech, such as those with Parkinson’s disease. (Hasegawa-Johnson, who compiles stats related to these issues, is the principal researcher at the Speech Accessibility Project, which aims to train models on more dysarthric speech to improve their outputs.)

The future of voice dictation will also be further complicated by the rise of generative AI. Large language models of the sort that power ChatGPT can also be used with audio, which would allow a program to better predict which word should come next in a sequence. For example, when transcribing, such an audio tool might reason that, based on the context, a person is likely saying that their dog—not their frog—needs to go for its morning walk.

Yet like their text counterparts, voice-recognition tools that use large language models can “hallucinate,” transcribing words that were never actually spoken. A team of scholars recently documented violent and unsavory hallucinations, as well as those that perpetuate harmful stereotypes, coming from OpenAI’s new audio model, Whisper. (In response to a request for comment about this research, a spokesperson for OpenAI said, in part, “We continually conduct research on how we can improve the accuracy of our models, including how we can reduce hallucinations.”)

So goes the AI boom: The technology is both creating impressive new things and introducing new problems. In voice dictation, the chasm between two once-distinct mediums, audio and text, is closing, leaving us to appreciate the marvel available in our hands—and to proceed with caution.