Connect with us

Science

When A.I.’s Output Is a Threat to A.I. Itself

Published

on

When A.I.’s Output Is a Threat to A.I. Itself

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.

Advertisement

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:

Advertisement

This is part of a data set of 60,000 handwritten digits.

When we trained an A.I. to mimic those digits, its output looked like this.

This new set was made by an A.I. trained on the previous A.I.-generated digits. What happens if this process continues?

Advertisement

After 20 generations of training new A.I.s on their predecessors’ output, the digits blur and start to erode.

After 30 generations, they converge into a single shape.

Advertisement

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.

Advertisement

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).

How an A.I. that draws digits “collapses” after being trained on its own output

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.

Advertisement

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:

Even at the outset, the A.I. “hallucinates.” But when the researchers further trained it on its own sentences, it got a lot worse…

An example of text generated by an A.I. model.

Advertisement

After two generations, it started simply printing long lists.

An example of text generated by an A.I. model after being trained on its own sentences for 2 generations.

And after four generations, it began to repeat phrases incoherently.

Advertisement

An example of text generated by an A.I. model after being trained on its own sentences for 4 generations.

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

Advertisement

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.

When A.I. image models are trained on their own output, they can produce distorted images, mangled fingers or strange patterns.

A.I.-generated images by Sina Alemohammad and others.

Advertisement

“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

Advertisement

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:

Advertisement

Distribution of A.I.-generated data

Examples of
initial A.I. output:

Advertisement

The distribution shown here is simplified for clarity.

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:

Advertisement

Distribution of A.I.-generated data when trained on its own output

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:

Advertisement

Distribution of A.I.-generated data when trained on its own output

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.

Advertisement

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.

Advertisement

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:

Advertisement

This set of A.I. faces was created by the same Rice researchers who produced the distorted faces above. This time, they tweaked the model to avoid visual glitches.

A grid of A.I.-generated faces showing variations in their poses, expressions, ages and races.

This is the output after they trained a new A.I. on the previous set of faces. At first glance, it may seem like the model changes worked: The glitches are gone.

Advertisement

After one generation of training on A.I. output, the A.I.-generated faces appear more similar.

After two generations …

After two generations of training on A.I. output, the A.I.-generated faces are less diverse than the original image.

Advertisement

After three generations …

After three generations of training on A.I. output, the A.I.-generated faces grow more similar.

After four generations, the faces all appeared to converge.

After four generations of training on A.I. output, the A.I.-generated faces appear almost identical.

Advertisement

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.

Advertisement

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.

Advertisement

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).

Advertisement

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.

Advertisement

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.

About the data

To produce the images of A.I.-generated digits, we followed a procedure outlined by researchers. We first trained a type of a neural network known as a variational autoencoder using a standard data set of 60,000 handwritten digits.

Advertisement

We then trained a new neural network using only the A.I.-generated digits produced by the previous neural network, and repeated this process in a loop 30 times.

To create the statistical distributions of A.I. output, we used each generation’s neural network to create 10,000 drawings of digits. We then used the first neural network (the one that was trained on the original handwritten digits) to encode these drawings as a set of numbers, known as a “latent space” encoding. This allowed us to quantitatively compare the output of different generations of neural networks. For simplicity, we used the average value of this latent space encoding to generate the statistical distributions shown in the article.

Continue Reading
Advertisement
Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Science

Why Cameras Are Popping Up in Eldercare Facilities

Published

on

Why Cameras Are Popping Up in Eldercare Facilities

The assisted-living facility in Edina, Minn., where Jean H. Peters and her siblings moved their mother in 2011, looked lovely. “But then you start uncovering things,” Ms. Peters said.

Her mother, Jackie Hourigan, widowed and developing memory problems at 82, too often was still in bed when her children came to see her in mid-morning.

“She wasn’t being toileted, so her pants would be soaked,” said Ms. Peters, 69, a retired nurse-practitioner in Bloomington, Minn. “They didn’t give her water. They didn’t get her up for meals.” She dwindled to 94 pounds.

Most ominously, Ms. Peters said, “we noticed bruises on her arm that we couldn’t account for.” Complaints to administrators — in person, by phone and by email — brought “tons of excuses.”

So Ms. Peters bought an inexpensive camera at Best Buy. She and her sisters installed it atop the refrigerator in her mother’s apartment, worrying that the facility might evict her if the staff noticed it.

Advertisement

Monitoring from an app on their phones, the family saw Ms. Hourigan going hours without being changed. They saw and heard an aide loudly berating her and handling her roughly as she helped her dress.

They watched as another aide awakened her for breakfast and left the room even though Ms. Hourigan was unable to open the heavy apartment door and go to the dining room. “It was traumatic to learn that we were right,” Ms. Peters said.

In 2016, after filing a police report and a lawsuit, and after her mother’s death, Ms. Peters helped found Elder Voice Advocates, which lobbied for a state law permitting cameras in residents’ rooms in nursing homes and assisted-living facilities. Minnesota passed it in 2019.

Though they remain a contentious subject, cameras in care facilities are gaining ground. By 2020, eight states had joined Minnesota in enacting laws allowing them, according to the National Consumer Voice for Quality Long-Term Care: Illinois, Kansas, Louisiana, Missouri, New Mexico, Oklahoma, Texas and Washington.

The legislative pace has picked up since, with nine more states enacting laws: Connecticut, North Dakota, South Dakota, Nevada, Ohio, Rhode Island, Utah, Virginia and Wyoming. Legislation is pending in several others.

Advertisement

California and Maryland have adopted guidelines, not laws. The state governments in New Jersey and Wisconsin will lend cameras to families concerned about loved ones’ safety.

But bills have also gone down to defeat, most recently in Arizona. In March, for the second year, a camera bill passed the House of Representatives overwhelmingly but failed to get a floor vote in the State Senate.

“My temperature is a little high right now,” said State Representative Quang Nguyen, a Republican who is the bill’s primary sponsor and plans to reintroduce it. He blamed opposition from industry groups, which in Arizona included LeadingAge, which represents nonprofit aging services providers, for the bill’s failure to pass.

The American Health Care Association, whose members are mostly for-profit long-term care providers, doesn’t take a national position on cameras. But its local affiliate also opposed the bill.

“These people voting no should be called out in public and told, ‘You don’t care about the elderly population,’” Mr. Nguyen said.

Advertisement

A few camera laws cover only nursing homes, but the majority also include assisted-living facilities. Most mandate that the resident (and roommates, if any) provide written consent. Some call for signs alerting staff and visitors that their interactions may be recorded.

The laws often prohibit tampering with cameras or retaliating against residents who use them, and include “some talk about who has access to the footage and whether it can be used in litigation,” added Lori Smetanka, executive director of the National Consumer Voice.

It’s unclear how seriously facilities take these laws. Several relatives interviewed for this article reported that administrators told them that cameras weren’t permitted, then never mentioned the issue again. Cameras placed in the room remained.

Why the legislative surge? During the Covid-19 pandemic, families were locked out of facilities for months, Ms. Smetanka pointed out. “People want eyes on their loved ones.”

Changes in technology probably also contributed, as Americans became more familiar and comfortable with video chatting and virtual assistants. Cameras have become nearly ubiquitous — in public spaces, in workplaces, in police cars and on officers’ uniforms, in people’s pockets.

Advertisement

Initially, the push for cameras reflected fears about loved ones’ safety. Kari Shaw’s family, for instance, had already been victimized by a trusted home care nurse who stole her mother’s prescribed pain medications.

So when Ms. Shaw, who lives in San Diego, and her sisters moved their mother into assisted living in Maple Grove, Minn., they immediately installed a motion-activated camera in her apartment.

Their mother, 91, has severe physical disabilities and uses a wheelchair. “Why wait for something to happen?” Ms. Shaw said.

In particular, “people with dementia are at high risk,” added Eilon Caspi, a gerontologist and researcher of elder mistreatment. “And they may not be capable of reporting incidents or recalling details.”

More recently, however, families are using cameras simply to stay in touch.

Advertisement

Anne Swardson, who lives in Virginia and in France, uses an Echo Show for video visits with her mother, 96, in memory care in Fort Collins, Colo. “She’s incapable of touching any buttons, but this screen just comes on,” Ms. Swardson said.

Art Siegel and his brothers were struggling to talk to their mother, who, at 101, is in assisted living in Florida; her portable phone frequently died because she forgot to charge it. “It was worrying,” said Mr. Siegel, who lives in San Francisco and had to call the facility and ask the staff to check on her.

Now, with an old-fashioned phone installed next to her favorite chair and a camera trained on the chair, they know when she’s available to talk.

As the debate over cameras continues, a central question remains unanswered: Do they bolster the quality of care? “There’s zero research cited to back up these bills,” said Clara Berridge, a gerontologist at the University of Washington who studies technology in elder care.

“Do cameras actually deter abuse and neglect? Does it cause a facility to change its policies or improve?”

Advertisement

Both camera opponents and supporters cite concerns about residents’ privacy and dignity in a setting where they are being helped to wash, dress and use the bathroom.

“Consider, too, the importance of ensuring privacy during visits related to spiritual, legal, financial or other personal issues,” Lisa Sanders, a spokeswoman for LeadingAge, said in a statement.

Though cameras can be turned off, it’s probably impractical to expect residents or a stretched-thin staff to do so.

Moreover, surveillance can treat those staff members as “suspects who have to be deterred from bad behavior,” Dr. Berridge said. She has seen facilities installing cameras in all residents’ rooms: “Everyone is living under surveillance. Is that what we want for our elders and our future selves?”

Ultimately, experts said, even when cameras detect problems, they can’t substitute for improved care that would prevent them — an effort that will require engagement from families, better staffing, training and monitoring by facilities, and more active federal and state oversight.

Advertisement

“I think of cameras as a symptom, not a solution,” Dr. Berridge said. “It’s a Band-Aid that can distract from the harder problem of how we provide quality long-term care.”

The New Old Age is produced through a partnership with KFF Health News.

Continue Reading

Science

Gray whales are dying off the Pacific Coast again, and scientists aren't sure why.

Published

on

Gray whales are dying off the Pacific Coast again, and scientists aren't sure why.

Gray whales are dying in large numbers, again.

At least 70 whales have perished since the start of the year in the shallow, protected lagoons of Mexico’s Baja California peninsula where the animals have congregated for eons to calf, nurse and breed, said Steven Swartz, a marine scientist who has studied gray whales since 1977. And only five mother-calf pairs were identified in Laguna San Ignacio, where most of the wintering whales tend to congregate, Swartz said.

That’s the lowest number of mother-calf pairs ever observed in the lagoon, according to annual reports from Gray Whale Research in Mexico, an international team of researchers — co-founded by Swartz — that has been observing gray whales in Laguna San Ignacio since the late 1970s.

The whales are now headed north. In just the last two weeks, three gray whales have died in San Francisco Bay, one of which was described by veterinarians and pathologists at the Marine Mammal Center in Sausalito as skinny and malnutritioned. Evaluations on the two other deaths are still being conducted.

Alisa Schulman-Janiger, who has led the Los Angeles chapter of the American Cetacean Society’s gray whale census at Rancho Palos Verdes since 1979, said the number of whales she and her volunteers have observed migrating north this spring and swimming south this past winter is the lowest on record.

Advertisement

“We didn’t see a single southbound calf, which has never happened in 40 years,” she said.

Schulman-Janiger and other researchers aren’t sure why the whales are dying, although she and others believe it could be from lack of food based on the depleted conditions in which some of the whales have been found.

Eastern North Pacific gray whales cruise the Pacific coastline every year as they migrate 6,000 miles north from the Baja peninsula to their summer feeding grounds in Arctic and sub-Arctic regions. There, the leviathans gorge themselves on small crustaceans and amphipods that live in the muddy sediment of the Bering, Chukchi and Beaufort seas, before they head back south to loll, cavort and mingle in balmy Mexican waters.

The animals migrate through a gantlet of perils as they navigate some of the world’s most heavily shipped regions, maneuver through discarded fishing lines and gear, dodge pods of killer whales waiting to tear apart defenseless calves, and swim through waters polluted with microplastics, toxic chemicals and poisonous algae.

Most of the time, the bulk of them make the journey just fine.

Advertisement

But in 2019, large numbers of the whales began to die.

Starting that spring, biologists at the Laguna San Ignacio research station recorded roughly 80 dead whales in Mexican waters, and just 41 mother-calf pairs in the lagoon. They also noticed — using photographs and drone imagery — that roughly a quarter of the animals were “skinny.”

“You can see it in photographs,” said Schulman-Janiger, who described skinny whales as looking like they had necks because a thick fat pad that typically covers the area behind the skull is gone. “And you can see their scapulae,” she said, referring to the animals’ shoulder blades.

“You shouldn’t see a whale’s shoulder blades,” she said.

Then, as the hungry whales migrated north in 2019, large numbers began stranding on the beaches of California, Oregon, Washington and Alaska. By the end of that year, researchers had documented 216 dead whales on the beaches and near shore waters of the North American Pacific coastline.

Advertisement

A federal investigation by the National Oceanic and Atmospheric Administration into what is known as an unexplained mortality event was launched in 2019. The investigation allowed for scientists across multiple disciplines and institutions to gather and share knowledge to determine the cause of the die-off.

The cause of the deaths was never definitively established, and the investigation was closed in 2023 as the number of strandings fell into a range considered normal. Many researchers concluded a change in Arctic and sub-Arctic food availability (via massive changes in climate) was the driving factor. Their assessment was supported by the observations of malnutrition and skinniness in the whales and similar events and observations in other Arctic animals, including birds, seals, crabs and fish.

They also noticed that many of the whales had started feeding in areas — such as San Francisco Bay and the Los Angeles and Long Beach harbors — where such behaviors had never before been seen.

In the last two weeks, several gray whales have been observed in San Francisco Bay, including a near record high of nine on a single day. Reports of feeding behaviors had also been made, including off the city of Pacifica.

Asked whether the researchers at NOAA are noting these concerning observations and anticipating the possibility of another die-off, Michael Milstein, an agency spokesman, said the number of strandings along the Pacific coast is still low — just seven in California and one in Washington. The annual average is about 35.

Advertisement

He said it was too early in the whales’ northward journey to know for sure.

John Calambokidis, senior research biologist and co-founder of the Cascadia Research Collective, a marine mammal research center based in Olympia, Wash., agreed with Milstein: “We are just entering our main period of strandings (April to June) so a little early to draw any conclusions.”

And despite Schulman-Janiger’s concerns, she too said it is early — and that La Niña ocean conditions may be partly to blame for the low number of animals observed thus far.

She said reports from Mexico indicate many gray whales migrated farther south than they typically do, and have been seen swimming around the Gulf of California — off the coasts of Loreto, Cabo San Lucas and Puerto Vallarta.

Gray whales swim from Alaska to Baja California, where they mate and give birth.

Advertisement

(Carolyn Cole / Los Angeles Times)

She said that is good news if the low counts are due to the whales just being late. But worrisome if already food-stressed whales are having to tack on an additional 800 miles to their journey.

“It’s a very weird year for gray whales, and a concerning year given their body condition, the strandings and the very low calf estimates,” she said.

Advertisement
Continue Reading

Science

Jeremiah Ostriker, Who Plumbed Dark Forces That Shape Universe, Dies at 86

Published

on

Jeremiah Ostriker, Who Plumbed Dark Forces That Shape Universe, Dies at 86

Jeremiah Ostriker, an astrophysicist who helped set off a revolution in humankind’s view of the universe, revealing it to be a vaster, darker realm than the one we can see, ruled by invisible forms of matter and energy we still don’t understand, died on Sunday at his home on the Upper West Side of Manhattan. He was 87.

His daughter Rebecca Ostriker said the cause was end-stage renal disease.

Over more than four decades, mostly at Princeton University, Dr. Ostriker’s work altered our understanding of how galaxies form and evolve as he explored the nature of pulsars, the role of black holes in the evolution of the cosmos and what the universe is made of.

Before the 1970s, most astronomers believed that galaxies were made up mostly of stars.

“Ostriker was arguably the most important single figure in convincing the astronomical community that this natural and seductive assumption is wrong,” David Spergel, the president of the Simons Foundation, which supports scientific research, wrote in 2022, nominating Dr. Ostriker, his mentor, for the Crafoord Prize, the astronomical equivalent of a Nobel. He cited Dr. Ostriker’s “eloquent advocacy for the then-radical new model in which the visible stars in galaxies were only a minor pollutant at the center of a much larger halo of dark matter of unknown composition.”

Advertisement

Dr. Ostriker’s work, he said, was “the grandest revision in our understanding of galaxies” in half a century.

Jerry Ostriker, as he was known to friends and colleagues, a man with a prickly sense of humor and a soft but commanding voice, was willing to go wherever the data and scientific calculations led him, and was not shy about questioning assumptions — or having fun. Prominently displayed in his home was a youthful photo of himself, taken in Cambridge, Mass., driving a motor scooter as his wife, Alicia Ostriker, seated behind him, lifts a bottle of wine to her lips. (A close look shows the cork still in the bottle.)

“He had the quickest wit of any scientist I have encountered,” said James Peebles, a Nobel physics laureate and a colleague of Dr. Ostriker’s at Princeton. “And I don’t remember ever matching him in a spontaneous debate” on any issue.

Asked in a 1988 oral history interview for the American Institute of Physics if he had favored any of the models of the universe being batted about in the 1970s, when he entered the field — whether the universe was finite or infinite, whether it had a beginning or was somehow always here, whether it would expand forever or crash back down in a big crunch — he said he had not.

“Scientists have followed their own biases, and my principle bias at the time was being contemptuous and intolerant of all of these people who had specific models,” he answered. “How could they be so certain when the evidence was as confusing and inconclusive?”

Advertisement

Jeremiah Paul Ostriker was born on April 13, 1937, on the Upper West Side, the second of four siblings. His father, Martin Ostriker, ran a clothing company, and his mother, Jeanne (Sumpf) Ostriker, was a public-school teacher. Babe Ruth lived around the corner, and the children used to chase his car for autographs.

“I must have been the classic nerd child,” Dr. Ostriker wrote in a memoir published in the Annual Review of Astronomy and Astrophysics in 2016. He first became interested in science when he was 4: His mother started reading science books aloud to get him to sit still for an oil portrait, and the readings stuck.

After graduating from the Ethical Culture Fieldston School in the Bronx, Jerry Ostriker went to Harvard University, where he planned to study chemistry. Instead, he switched to physics, which appealed to what he called his “cosmic perspective.”

“I probably spent more time on literature than I spent on science,” he said in the oral history interview.

He soon began commuting to Brandeis University to visit Alicia Suskin, a former Fieldston classmate who was an aspiring artist and poet. They were married in 1958, while they were still undergraduates.

Advertisement

Ms. Ostriker, a professor emerita of English at Rutgers University, became an award-winning poet and has often written her husband into her work. In turn, he found poetry in astrophysics. “As an astrophysicist, you get a perspective on humankind,” he said, describing it as “sweating on this little grain of spinning sand.”

In addition to his wife and his daughter Rebecca, an editor for the opinion section of The Boston Globe, Dr. Ostriker is survived by two other children, Eve Ostriker, an astrophysicist at Princeton, and Gabriel Ostriker, a data engineer; a sister, Naomi Seligman; two brothers, Jon and David; and three grandchildren.

After graduating from Harvard in 1959, Dr. Ostriker worked at the United States Naval Research Laboratory for a year before enrolling in graduate school at the University of Chicago, splitting his time between the university’s Yerkes Observatory and the physics department, where he worked under the future Nobel laureate Subrahmanyan Chandrasekhar.

He earned his Ph.D. in 1964. After a postdoctoral year at the University of Cambridge in England, where he rubbed elbows with future black hole eminences like Stephen Hawking and Martin Rees, Dr. Ostriker joined Princeton as a research scientist. He remained there for 47 years, rising through the ranks to become chairman of the astronomy department and provost of the university.

At Princeton, Dr. Ostriker wrote a series of papers that would lead astronomy to the dark side.

Advertisement

He wondered whether galaxies, like stars, could break apart if they rotated too fast. The question was particularly relevant to so-called disc galaxies like the Milky Way, which are shaped sort of like a fried egg, with a fat, yolky center surrounded by a thin, white flat of stars.

Working with Dr. Peebles, he constructed a computer simulation and found that disc galaxies were indeed unstable. They would fall apart unless there was something we couldn’t see, a halo of some additional invisible mass, lending gravitational support.

Whatever this stuff called dark matter was — dim stars, black holes, rocks, exotic subatomic particles left over from the Big Bang — there could be a lot of it, as much as 10 times the mass of ordinary atomic matter.

It was one of the first theoretical arguments that there must be more to galaxies than could be seen in starlight. In the 1930s, the astronomer Fritz Zwicky had suggested that most of the mass in galaxies was “dark.” His idea was largely ignored until Dr. Ostriker and Dr. Peebles published their paper in 1973.

The reaction from the scientific community was predominantly hostile, Dr. Ostriker said. “I couldn’t see particularly why,” he said in the oral history. “It was just a fact.”

Advertisement

A year later, incorporating more data from galaxy clusters and other star systems, he and his colleagues argued that, in fact, most of the mass in the universe was invisible.

By the early 1980s, the idea of dark matter had become an accepted part of cosmology, but there remained conundrums, including calculations that suggested that stars were older than the universe in which they lived.

The missing ingredient, Dr. Ostriker and the theoretical physicist Paul Steinhardt, then at the University of Pennsylvania, suggested in 1995, was a fudge factor known as the cosmological constant. Einstein had come up with this concept in 1917, but had later abandoned it, considering it a blunder.

As Dr. Steinhardt recalled, he and Dr. Ostriker were “convinced that a universe with only dark and ordinary matter could not explain the existing observations.” But once they added the cosmological constant, everything came out right.

They were not the only ones with this idea. The cosmologists Michael Turner, now retired from the University of Chicago, and Lawrence Krauss, now retired from Arizona State University, also argued in favor of bringing back the constant. “To say Jerry was a giant in the field is an understatement,” Dr. Turner wrote in an email, adding, “Sparring with Jerry over science was a privilege and often a learning experience.”

Advertisement

Three years later, two competing teams of astronomers discovered that the expansion of the universe was being accelerated by a “dark energy” acting as the cosmological constant, pushing galaxies apart. The cosmological constant then became part of a standard model of the universe, as Dr. Ostriker and others had predicted.

In another series of papers, he and various collaborators transformed astronomers’ view of what was going on in the space between stars.

Dr. Ostriker and Renyue Cen, also of Princeton, concluded in 1999 that most ordinary atomic matter in the nearby universe was invisible, taking the form of intergalactic gas heated to millions of degrees by shock waves and explosions.

At Princeton, Dr. Ostriker helped set up the Sloan Digital Sky Survey, a collaboration — initially of Princeton, the University of Chicago and the Institute for Advanced Study in Princeton, N.J. — aimed at remapping the entire sky in digital form with a dedicated telescope at Apache Point Observatory in Sunspot, N.M.

“The survey is going to increase our knowledge and our understanding of the universe a hundredfold,” he told The New York Times in 1991. “The map is not going to show us how the universe began, but it will show us the nature and origin of large-scale structure, the most interesting problem in astrophysics today. With an answer to this problem, we will be able to better approach the question of how it all began.”

Advertisement

The survey, started in 1998, is now in its fifth iteration and has generated some 10,000 research papers and archived measurements of a half-billion stars and galaxies, all free to any astronomer in the world.

As provost, Dr. Ostriker led the effort to vastly expand the university’s financial aid program, changing many loans to grants that would not need to be repaid, making a Princeton education more egalitarian. In 2000, he was awarded the National Medal of Science by President Bill Clinton.

Dr. Ostriker retired from Princeton in 2012, just as his daughter Eve was joining the astronomy faculty there. He took a part-time position at Columbia University, returning to his childhood neighborhood.

“Growing up in New York City, I couldn’t see the stars,” he once told The Times. He found them anyway, and a whole lot more that we can’t see with or without the glare of streetlights.

It was a passion that never waned. Encountered recently by a reporter on the sidewalk in front of Columbia, Dr. Ostriker launched into an enthusiastic description of a promising new theory of dark matter.

Advertisement

Early in 2023, by then ailing, he took to his bed at home. But he kept up with his research by email and had regular pizza lunches with colleagues.

Apprised recently of results from the James Webb Space Telescope that seemed to reinforce his ideas about dark matter, he wrote in an email to his colleagues, “Keep up the good work.”

The dark universe he helped conjure half a century ago has developed a few small cracks, leading to new ideas about the nature of that dark matter.

“It’s a very, very, very specific and clear theory. So therefore, God bless it, it can be wrong,” Dr. Ostriker said in a recent interview. “That’s the way science proceeds. And what we know about it is that it is a little bit wrong, not a lot wrong.”

Dr. Rees, a cosmologist at the University of Cambridge and the Astronomer Royal, summed up Dr. Ostriker’s life this way: “Some scientists come up with pioneering ideas on novel themes; others write definitive ‘last words’ on already-established topics. Jerry was in the first category.”

Advertisement

“He wrote among the earliest papers — now classics — on the nature of pulsars, the evidence for dark matter and on galaxy formation and cosmology. His flow of papers continued into his 80s,” Dr. Rees added. “He enthusiastically engaged in new data and in computational techniques. He inspired younger colleagues and collaborators, not just at Princeton but around the world.”

Continue Reading

Trending