Christine LyBurtus was aching and fearful of what might happen when her 13-year-old son returned home.

Noah had been sent to Children’s Hospital of Orange County for a psychiatric hold lasting up to 72 hours after he punched at walls, flipped over a table, ripped out a chunk of his mother’s hair and tried to break a car window.

“There’s nothing else to call it except a psychotic episode,” LyBurtus said.

The clock was ticking on that August day in 2022. The single mother wanted help to prevent such an episode from happening again, maybe with a different medication. Hospital staff were waiting for a psychiatric bed, possibly at another hospital with a dedicated unit for patients with autism or other developmental disabilities.

But as the hours ran out on the hold, it became clear that wasn’t happening. LyBurtus brought Noah home to their Fullerton apartment.

“When he came back home, it kind of broke my heart,” said his sister, Karissa, who is two years older. “He looked like, ‘What the heck did you guys put me into?’”

The next night, Noah was back in the ER after smashing a television and attacking his mother. This time, he was transferred to a different hospital for three weeks, prescribed medications for psychosis, and then sent to a residential facility in Garden Grove.

LyBurtus said she was told it would be a stopgap measure — just for three weeks — until she could line up more help at home. But when she phoned to ask about visiting her son, LyBurtus said she was told she couldn’t see him for a month.

Advertisement

“He lives here now,” someone told her, she said, and the staff needed time to “break him in.”

LyBurtus felt like she was being pushed to give up her son, instead of getting the help her family needed. She insisted on bringing him home.

::

Autism is a developmental condition that can shape how people think, communicate, move and process sensory information. When Noah was 3, a doctor noted he was a “very cute little boy” who played alone, rocked back and forth, and sometimes bit himself. Noah’s eye contact was “fleeting.” He could speak about 20 words, but often cried or pulled his mother’s hand to communicate.

The physician summed up his behavior as “characteristic of a DSM-IV diagnosis of autistic disorder.”

When he was in elementary school, LyBurtus stopped working full time outside the home and enrolled in a state program that paid her as his caregiver. She relies on Medi-Cal for his medical care, and much of his schooling has been in Orange County-run programs for children with moderate to severe disabilities.

Noah does not speak but sometimes uses pictures, an app on a tablet, or some sign language to communicate. When a reporter visited their home last year, Noah bobbed his head and shoulders as he listened to music on his iPad. He flapped his hands as LyBurtus made him a peanut-butter-and-banana smoothie, and then dutifully followed her instructions to chuck the peel and put the almond milk away. It was a good day, LyBurtus said with relief.

But on other days, LyBurtus said her son could be rigid; his demands, unpredictable. “Some days he’s fixated on having three pairs of pants on … Some days he wants to take seven showers. The next day, I can’t get him to take showers.”

Christine LyBurtus greets Noah as he arrives home from school.

(Allen J. Schaben / Los Angeles Times)

Advertisement

When frustrated, Noah might erupt, banging his head against walls and trying to jump out the windows of their apartment. He had kicked and bitten his mother when she tried to redirect him. In the worst instances, LyBurtus had resorted to hiding in the bathroom — her “safe room” — and urged Karissa to lock herself in the bedroom.

As Noah grew taller and stronger, LyBurtus stripped bare the walls of her apartment to try to make it safe, installed shatterproof windows and removed a knob from a closet door to prevent Noah from using it as a foothold to scale over the top of the closet door. She made sure to flag her address for the Fullerton Police Department so it knew her son was developmentally disabled.

“I’m just so grateful that my son never got shot,” LyBurtus said.

Each of the 911 calls was the start of a Sisyphean routine. Noah “has been challenging to place in [a] mental health facility due to behavioral care needs with severe autism,” a doctor wrote when he was back at Children’s Hospital of Orange County yet again.

Noah leaps into the air inside his Fullerton home. At left is Terrence Morris, one of Noah’s caregivers.

(Mel Melcon / Los Angeles Times)

Advertisement

As the family tried to get through each crisis, LyBurtus was also facing a common struggle among parents of California children with disabilities: not getting the help they were supposed to receive from the state.

LyBurtus was getting assistance through a local regional center, one of the nonprofit agencies contracted by the California Department of Developmental Services. She said she’d been authorized to receive 40 hours weekly of respite care — meant to relieve families of children with disabilities for short periods — but was sometimes receiving only 12 to 16 hours.

She was also supposed to have two workers at a time, LyBurtus said, but caregivers were so scarce that she was scheduling one at a time in order to cover as many hours as she could.

In the meantime, Noah wasn’t sleeping and she was going through so much laundry detergent and quarters that her grocery budget was drained. At one point, she wanted to go to a food bank, but there would be no one to watch him.

“I could not be anymore tired and frustrated!!!!” she wrote to her regional center coordinator. “Is the only way Noah is going to get help [is] if I abandoned him and surrender him to the State!?!?”

Christine LyBurtus said she’s struggled to find the right care for Noah.

(Allen J. Schaben / Los Angeles Times)

::

Across the country, surging numbers of young people have landed in emergency rooms in the throes of a mental health crisis amid a shortage of needed care. Children in need of psychiatric care are routinely held in emergency departments for hours or even days. Even amid COVID, as people tried to avoid emergency rooms, mental health-related visits continued to rise among teens in 2021 and 2022.

Among those hit hardest by the crisis are autistic youth, who turn up in emergency rooms at higher rates than other kids — and are much more likely to do so for psychiatric issues. Many have overlapping conditions such as anxiety, and researchers have also found they face a higher risk of abuse and trauma.

“We’re a misunderstood, marginalized population of people” at higher risk of suicide, Lisa Morgan, founder of the Autism and Suicide Prevention Workgroup, said at a national meeting.

Yet the available assistance is “not designed for us.”

According to the National Autism Indicators Report, more than half of parents of autistic youth who were surveyed had trouble getting the mental health services their autistic kids needed, with 22% saying it was “very difficult” or “impossible.” A report commissioned by L.A. County found autistic youth were especially likely to languish in ERs amid few options for ongoing psychiatric treatment.

Karissa interacts with her brother, Noah, as he watches a video after school.

(Allen J. Schaben / Los Angeles Times)

Advertisement

In decades past, many psychiatrists were unwilling to diagnose mental health disorders in autistic people, believing “it was either part of the autism or for other reasons it was undiagnosable,” said Jessica Rast, an assistant research professor affiliated with the A.J. Drexel Autism Institute. Much more is now known about both autism and mental health treatment, but experts say the two fields aren’t consistently linked in practice.

Mental health providers may focus on an autism diagnosis for a prospective patient and say, “‘Well, that’s not in our wheelhouse. We’re treating things like depression or anxiety,’” said Brenna Maddox, assistant professor of psychiatry at the University of North Carolina School of Medicine.

Yet patients or their families “weren’t asking for autism treatment. They were asking for depression or anxiety or other mental health treatment,” Maddox said.

In the meantime, the system that serves children with developmental disabilities has faltered.

“Never have I seen that we can’t staff the needed things on so many cases,” Larry Landauer, executive director of the Regional Center of Orange County, said last year. Statewide, “there’s thousands and thousands of cases that are struggling.”

“If I’m a respite worker and I get called on to provide help to families … who am I going to select?” Landauer asked. “The [person] that watches TV and plays on his iPad and I just sit and monitor him? Or do I take someone that is significantly behaviorally challenged — that pulls my hair, that scares me all the time, that tries to run out the door? … Those are the ones getting left out.”

::

The fall and winter of 2022 were so trying that LyBurtus eventually took matters into her own hands. Noah bit his mother and smashed a bathroom window and tried to climb out before the Fullerton Fire Department arrived. Weeks later, LyBurtus had to dial 911 again after he bit his sister’s finger badly enough to draw blood.

Caregiver Terrence Morris, left, keeps a watchful eye on Noah.

(Mel Melcon / Los Angeles Times)

Advertisement

He ended up in a hold at Children’s Hospital of Orange County, which searched for another facility that might help him, but “all placement options declined patient placement,” according to his medical records.

Noah was again sent home with his mother, but the next day, he was back at Children’s Hospital of Orange County after slamming his head against a tile floor.

LyBurtus, frantic and bruised, made call after call and finally used her credit card to pay for an ambulance to take him to UCLA Resnick Neuropsychiatric Hospital, where he was admitted.

Week by week, psychiatrists there said Noah seemed to be making some strides as they adjusted his alphabet soup of medications. But hospital staff struggled to understand what would set him off.

Once, while playing cards, Noah suddenly started knocking the cards off the table and struck another patient in the face. Another day, he appeared suddenly to be frightened after using the bathroom, and then charged at a computer plugged in nearby.

But there were also days when he danced to a Michael Jackson song, or played Giant Jenga outside on the deck. One day, a doctor wrote, “He made eye contact for a few seconds. I waved to him, and he looked at his hand, as though he was wondering what to do with it in return.”

Christine LyBurtus washes her son’s face. When Noah was 3, a doctor noted he was a “very cute little boy” who played alone, rocked back and forth, and sometimes bit himself.

(Mel Melcon / Los Angeles Times)

LyBurtus was straining to find more help at home so UCLA held off on discharging him, but at the end of January 2023 Noah was sent home. With no changes in medication planned, “and the strong possibility that Noah grew tired of the inpatient setting, the ward no longer was deemed therapeutic or necessary,” a doctor wrote.

Less than a month later, he was back in the emergency room at Children’s Hospital of Orange County after biting and attacking his mother.

A psychiatrist at the pediatric hospital wrote that because he had limited ability to communicate, another round of psychiatric hospitalization would do little unless it was specialized for “individuals with neurodevelopmental needs.” When the 72-hour hold at children’s hospital ran out, LyBurtus asked for an ambulance to take Noah home, fearful of driving him herself.

In May, the month Noah turned 14, LyBurtus heard the regional center had found a place for Noah: a four-bed facility in Rio Linda, a tiny town near Sacramento that she’d never heard of. He could live there for more than a year, she was told, and then hopefully return home with the right support.

Christine LyBurtus shows photographs to Noah.

(Mel Melcon / Los Angeles Times)

Advertisement

But LyBurtus fretted about what she would do if something happened to him so far away. She felt, she said, like she had failed her child. Months passed as they waited for a spot there; LyBurtus said she was told they were trying to hire the needed staff.

“I don’t want him to go,” she said, “but I don’t want to continue going on the way that we’re going on.”

Then in August, LyBurtus was told the regional center had found a spot at a facility much closer to home: the state-run South STAR facility in Costa Mesa, about 20 miles from their apartment. Noah would occupy one of only 15 STAR beds across the state for developmentally disabled adolescents in “acute crisis.”

On a bright September morning, LyBurtus pulled up at an unassuming gray house with a “Home Sweet Home” sign by the door. The three teens living there were gone for the morning while an administrator and South STAR program director Kim Hamilton-Royse showed LyBurtus around the house.

Minutes into the tour, LyBurtus found herself crying. Hamilton-Royse stopped her explanation of the daily schedule. “I know this is super hard for you,” she said gently.

But LyBurtus brightened at the sight of the sensory room outfitted with crash pads and a mesmerizing, colorful cylinder of bubbling water. Hamilton-Royse pointed out a vibrating chair and added that they had a projector that would fill the room with illuminated stars.

LyBurtus took photos on her smartphone to show Noah. “You’re not going to be able to get him out of here,” she said.

As they rounded the rest of the house — bedrooms with dressers secured to the wall, a living room with paintings of sailboats, a fish tank — Hamilton-Royse asked if LyBurtus felt any better.

Christine LyBurtus reacts while boxing up items for Noah’s move.

(Mel Melcon / Los Angeles Times)

Advertisement

“I do,” she said. “I just hope that he can behave.”

Hamilton-Royse reassured her that South STAR had never kicked anyone out. “And we’ve had some really challenging folks,” she said.

“I promise you we’ll take very good care of him.”

As she returned to her car, LyBurtus took a deep breath. “It’s hard not to feel like I’m betraying him,” she said, her voice shaking. “But I can’t keep living like this, you know?”

1

Advertisement

2

3

1. Christine Lyburtus tours a residential care facility in Costa Mesa, about 20 minutes from her home. (Irfan Khan / Los Angeles Times) 2. At the South STAR facility, LyBurtus was told, Noah would occupy one of only 15 STAR beds across the state for developmentally disabled adolescents in “acute crisis.” (Irfan Khan / Los Angeles Times) 3. “I just hope that he can behave,” LyBurtus said of son Noah. (Irfan Khan / Los Angeles Times)

Three days later, Noah went back to the Children’s Hospital of Orange County on another psychiatric hold. He came home, then was back in the emergency department a week and a half later.

::

The October night before Noah left home, LyBurtus had brought home sushi for him, one of his favorite foods. He fell asleep around 6:30 p.m, and woke up again at 1 a.m. LyBurtus gave him his medication and as he drifted back to sleep, his mother held him, enjoying the peace.

When he woke up in the morning, she could tell he knew something was up. His clothes had been packed. She’d already shown him photos of the Costa Mesa home and told him, “This is where you’re going. I’m still your mom. I’m still going to go and see you.”

Noah embraces his mother shortly before he was picked up and driven to a residential care facility in Costa Mesa.

(Mel Melcon / Los Angeles Times)

When the black SUV arrived, LyBurtus offered Oreos to coax him into the unfamiliar car. She followed the SUV in her car, staying far enough behind to avoid having Noah see her when he arrived. LyBurtus had been told it would ease the transition.

Back at home, she sank into the bathtub, utterly spent. “I’m going to have to just go with trusting this process as much as I can,” she said, “because I don’t have another choice right now.”

The next day, she met with the South STAR staff to tell them more about Noah. What he likes to eat. What triggers him. His favorite things to do. The Costa Mesa home called whenever staff had physically restrained Noah, but when a weekend passed without a call, she felt some relief.

Lyburtus smiled at the photos and videos sent home: putting together an elaborate stacking toy, washing dishes. It felt like things were going well, LyBurtus said. The staff had scaled back the amount of psychiatric medication he was taking.

But more than a month later, when she first went to visit Noah, he excitedly took her to the front door, as if to say, “Let’s go,” she recalled. She gently told him she was just visiting.

Christine LyBurtus is comforted by caregivers Schahara Zad, left, and Terrence Morris after Noah moved into his residential care facility.

(Mel Melcon / Los Angeles Times)

Advertisement

He led her to the side door instead. She steered him away again. They stepped into the courtyard, and Noah immediately went to the gate to exit.

LyBurtus fell into a funk. As she worried about Noah, she was also figuring out how to make ends meet. With Noah in the Costa Mesa home, Lyburtus was no longer being paid more than $4,000 a month as his caregiver, her sole source of income for years. She tried a number of jobs but ultimately found the work that suited her: caregiving for an elderly woman and children with disabilities.

Her second and third visits with Noah were easier. She snapped photos — Mother and son nestled together on the couch. Noah touching her forehead.

The STAR program runs up to 13 months. As time passed, the regional center had started talking to her about where Noah would go next. LyBurtus was startled.

Wasn’t the plan for him to come home, she asked?

Christine LyBurtus, left, is briefed by Kim Hamilton-Royse while touring a residential care facility for her son.

(Irfan Khan/Los Angeles Times)

That was still on the table, LyBurtus said she was told. But if he wasn’t ready, they didn’t want to wait until the last minute to find somewhere else for Noah, who turned 15 in May.

LyBurtus wanted to block out the idea of him going to another facility.

“I never want to live the way we were living again,” she said.

“But is that worse than him being hours away? I don’t know.”

Every time someone uses ChatGPT to write an essay, create an image or advise them on planning their day, the environment pays a price.

A query on the chatbot that uses artificial intelligence is estimated to require at least 10 times more electricity than a standard search on Google.

If all Google searches similarly used generative AI, they might consume as much electricity as a country the size of Ireland, calculates Alex de Vries, the founder of Digiconomist, a website that aims to expose the unintended consequences of digital trends.

Yet someone using ChatGPT or another artificial intelligence application has no way of knowing how much power their questions will consume as they are processed in the tech companies’ enormous data centers.

De Vries said the skyrocketing energy demand of AI technologies will no doubt require the world to burn more climate-warming oil, gas and coal.

“Even if we manage to feed AI with renewables, we have to realize those are limited in supply, so we’ll be using more fossil fuels elsewhere,” he said. “The ultimate outcome of this is more carbon emissions.”

AI is also thirsty for water. ChatGPT gulps roughly a 16-ounce bottle in as few as 10 queries, calculates Shaolei Ren, associate professor of electrical and computer engineering at UC Riverside, and his colleagues.

The increasing consumption of energy and water by AI has raised concerns in California and around the globe. Experts have detailed how it could stall the transition to green energy, while increasing consumer’s electric bills and the risk of blackouts.

To try to prevent those consequences, De Vries, Ren and other experts are calling on the tech companies to disclose to users how much power and water their queries will consume.

“I think the first step is to have more transparency,” Ren said. The AI developers, he said, “tend to be secretive about their energy usage and their water consumption.”

Ren said users should be told on the websites where they are asked to type in their queries how much energy and water their requests will require. He said this would be similar to how Google now tells people searching for airline flights how much carbon emissions the trip will generate.

“If we had that knowledge,” he said, “then we could make more informed decisions.”

Data centers — enormous warehouses of computer servers that support the internet — have long been big power users. But the specialized computer chips required for generative AI use far more electricity because they are designed to read through vast amounts of data.

The new chips also generate so much heat that even more power and water is needed to keep them cool.

Even though the benefits and risks of AI aren’t yet fully known, companies are increasingly incorporating the technology into existing products.

In May, for example, Google announced that it was adding what it called “AI Overviews” to its search engine. Whenever someone now types a question into Google search, the company’s AI generates an answer from the search results, which is highlighted at the top.

Not all of Google’s AI-generated answers have been correct, including when it told a user to add Elmer’s glue to pizza sauce to keep cheese from sliding off the crust.

But searchers who don’t want those AI-generated answers or want to avoid the extra use of power and water can’t turn off the feature.

“Right now, the user doesn’t have the option to opt out,” Ren said.

Google did not respond to questions from The Times.

OpenAI, the company that created ChatGPT, responded with a prepared statement, but declined to answer specific questions, such as how much power and water the chatbot used.

“AI can be energy-intensive and that’s why we are constantly working to improve efficiencies,” OpenAI said. “We carefully consider the best use of our computing power and support our partners’ efforts to achieve their sustainability goals. We also believe that AI can play a key role in accelerating scientific progress in the discovery of climate solutions.”

Three years ago, Google vowed to reach “net-zero” — where its emissions of greenhouse gases would be equal to what it removed — by 2030.

The company isn’t making progress toward that goal. In 2023, its total carbon emissions increased by 13%, the company disclosed in a July report. Since 2019, its emissions are up 48%.

“As we further integrate AI into our products, reducing emissions may be challenging due to increasing energy demands from the greater intensity of AI compute,” the company said in the report.

Google added that it expects its emissions to continue to rise before dropping sometime in the future. It did not say when that may be.

The company also disclosed that its data centers consumed 6.1 billion gallons of water in 2023 — 17% more than the year before.

“We’re committed to developing AI responsibly by working to address its environmental footprint,” the report said.

De Vries said he was disappointed Google had not disclosed in the report how much AI was adding to its power needs. The company said in the report that such a “distinction between AI and other workloads” would “not be meaningful.”

By not separately reporting the power use of AI, he said, it is impossible to calculate just how much more electricity Google search was now using with the addition of AI Overviews.

“While capable of delivering the required info,” he said, “they are now withholding it.”

The internet is becoming awash in words and images generated by artificial intelligence.

Sam Altman, OpenAI’s chief executive, wrote in February that the company generated about 100 billion words per day — a million novels’ worth of text, every day, an unknown share of which finds its way onto the internet.

A.I.-generated text may show up as a restaurant review, a dating profile or a social media post. And it may show up as a news article, too: NewsGuard, a group that tracks online misinformation, recently identified over a thousand websites that churn out error-prone A.I.-generated news articles.

In reality, with no foolproof methods to detect this kind of content, much will simply remain undetected.

All this A.I.-generated information can make it harder for us to know what’s real. And it also poses a problem for A.I. companies. As they trawl the web for new data to train their next models on — an increasingly challenging task — they’re likely to ingest some of their own A.I.-generated content, creating an unintentional feedback loop in which what was once the output from one A.I. becomes the input for another.

In the long run, this cycle may pose a threat to A.I. itself. Research has shown that when generative A.I. is trained on a lot of its own output, it can get a lot worse.

Here’s a simple illustration of what happens when an A.I. system is trained on its own output, over and over again:

While this is a simplified example, it illustrates a problem on the horizon.

Imagine a medical-advice chatbot that lists fewer diseases that match your symptoms, because it was trained on a narrower spectrum of medical knowledge generated by previous chatbots. Or an A.I. history tutor that ingests A.I.-generated propaganda and can no longer separate fact from fiction.

Just as a copy of a copy can drift away from the original, when generative A.I. is trained on its own content, its output can also drift away from reality, growing further apart from the original data that it was intended to imitate.

In a paper published last month in the journal Nature, a group of researchers in Britain and Canada showed how this process results in a narrower range of A.I. output over time — an early stage of what they called “model collapse.”

The eroding digits we just saw show this collapse. When untethered from human input, the A.I. output dropped in quality (the digits became blurry) and in diversity (they grew similar).

If only some of the training data were A.I.-generated, the decline would be slower or more subtle. But it would still occur, researchers say, unless the synthetic data was complemented with a lot of new, real data.

Degenerative A.I.

In one example, the researchers trained a large language model on its own sentences over and over again, asking it to complete the same prompt after each round.

When they asked the A.I. to complete a sentence that started with “To cook a turkey for Thanksgiving, you…,” at first, it responded like this:

“The model becomes poisoned with its own projection of reality,” the researchers wrote of this phenomenon.

This problem isn’t just confined to text. Another team of researchers at Rice University studied what would happen when the kinds of A.I. that generate images are repeatedly trained on their own output — a problem that could already be occurring as A.I.-generated images flood the web.

They found that glitches and image artifacts started to build up in the A.I.’s output, eventually producing distorted images with wrinkled patterns and mangled fingers.

“You’re kind of drifting into parts of the space that are like a no-fly zone,” said Richard Baraniuk, a professor who led the research on A.I. image models.

The researchers found that the only way to stave off this problem was to ensure that the A.I. was also trained on a sufficient supply of new, real data.

While selfies are certainly not in short supply on the internet, there could be categories of images where A.I. output outnumbers genuine data, they said.

For example, A.I.-generated images in the style of van Gogh could outnumber actual photographs of van Gogh paintings in A.I.’s training data, and this may lead to errors and distortions down the road. (Early signs of this problem will be hard to detect because the leading A.I. models are closed to outside scrutiny, the researchers said.)

Why collapse happens

All of these problems arise because A.I.-generated data is often a poor substitute for the real thing.

This is sometimes easy to see, like when chatbots state absurd facts or when A.I.-generated hands have too many fingers.

But the differences that lead to model collapse aren’t necessarily obvious — and they can be difficult to detect.

When generative A.I. is “trained” on vast amounts of data, what’s really happening under the hood is that it is assembling a statistical distribution — a set of probabilities that predicts the next word in a sentence, or the pixels in a picture.

For example, when we trained an A.I. to imitate handwritten digits, its output could be arranged into a statistical distribution that looks like this:

The peak of this bell-shaped curve represents the most probable A.I. output — in this case, the most typical A.I.-generated digits. The tail ends describe output that is less common.

Notice that when the model was trained on human data, it had a healthy spread of possible outputs, which you can see in the width of the curve above.

But after it was trained on its own output, this is what happened to the curve:

It gets taller and narrower. As a result, the model becomes more and more likely to produce a smaller range of output, and the output can drift away from the original data.

Meanwhile, the tail ends of the curve — which contain the rare, unusual or surprising outcomes — fade away.

This is a telltale sign of model collapse: Rare data becomes even rarer.

If this process went unchecked, the curve would eventually become a spike:

This was when all of the digits became identical, and the model completely collapsed.

Why it matters

This doesn’t mean generative A.I. will grind to a halt anytime soon.

The companies that make these tools are aware of these problems, and they will notice if their A.I. systems start to deteriorate in quality.

But it may slow things down. As existing sources of data dry up or become contaminated with A.I. “slop,” researchers say it makes it harder for newcomers to compete.

A.I.-generated words and images are already beginning to flood social media and the wider web. They’re even hiding in some of the data sets used to train A.I., the Rice researchers found.

“The web is becoming increasingly a dangerous place to look for your data,” said Sina Alemohammad, a graduate student at Rice who studied how A.I. contamination affects image models.

Big players will be affected, too. Computer scientists at N.Y.U. found that when there is a lot of A.I.-generated content in the training data, it takes more computing power to train A.I. — which translates into more energy and more money.

“Models won’t scale anymore as they should be scaling,” said ​​Julia Kempe, the N.Y.U. professor who led this work.

The leading A.I. models already cost tens to hundreds of millions of dollars to train, and they consume staggering amounts of energy, so this can be a sizable problem.

‘A hidden danger’

Finally, there’s another threat posed by even the early stages of collapse: an erosion of diversity.

And it’s an outcome that could become more likely as companies try to avoid the glitches and “hallucinations” that often occur with A.I. data.

This is easiest to see when the data matches a form of diversity that we can visually recognize — people’s faces:

This drop in diversity is “a hidden danger,” Mr. Alemohammad said. “You might just ignore it and then you don’t understand it until it’s too late.”

Just as with the digits, the changes are clearest when most of the data is A.I.-generated. With a more realistic mix of real and synthetic data, the decline would be more gradual.

But the problem is relevant to the real world, the researchers said, and will inevitably occur unless A.I. companies go out of their way to avoid their own output.

Related research shows that when A.I. language models are trained on their own words, their vocabulary shrinks and their sentences become less varied in their grammatical structure — a loss of “linguistic diversity.”

And studies have found that this process can amplify biases in the data and is more likely to erase data pertaining to minorities.

Ways out

Perhaps the biggest takeaway of this research is that high-quality, diverse data is valuable and hard for computers to emulate.

One solution, then, is for A.I. companies to pay for this data instead of scooping it up from the internet, ensuring both human origin and high quality.

OpenAI and Google have made deals with some publishers or websites to use their data to improve A.I. (The New York Times sued OpenAI and Microsoft last year, alleging copyright infringement. OpenAI and Microsoft say their use of the content is considered fair use under copyright law.)

Better ways to detect A.I. output would also help mitigate these problems.

Google and OpenAI are working on A.I. “watermarking” tools, which introduce hidden patterns that can be used to identify A.I.-generated images and text.

But watermarking text is challenging, researchers say, because these watermarks can’t always be reliably detected and can easily be subverted (they may not survive being translated into another language, for example).

A.I. slop is not the only reason that companies may need to be wary of synthetic data. Another problem is that there are only so many words on the internet.

Some experts estimate that the largest A.I. models have been trained on a few percent of the available pool of text on the internet. They project that these models may run out of public data to sustain their current pace of growth within a decade.

“These models are so enormous that the entire internet of images or conversations is somehow close to being not enough,” Professor Baraniuk said.

To meet their growing data needs, some companies are considering using today’s A.I. models to generate data to train tomorrow’s models. But researchers say this can lead to unintended consequences (such as the drop in quality or diversity that we saw above).

There are certain contexts where synthetic data can help A.I.s learn — for example, when output from a larger A.I. model is used to train a smaller one, or when the correct answer can be verified, like the solution to a math problem or the best strategies in games like chess or Go.

And new research suggests that when humans curate synthetic data (for example, by ranking A.I. answers and choosing the best one), it can alleviate some of the problems of collapse.

Companies are already spending a lot on curating data, Professor Kempe said, and she believes this will become even more important as they learn about the problems of synthetic data.

But for now, there’s no replacement for the real thing.