Did Anthropic's AI really make a scientific discovery on its own?

An expert at the University of Copenhagen said his team had been sharing its research with the company's AI model, Claude, and that its new finding matched their work

By Carl Zimmer 

 

When Anthropic, the artificial intelligence giant, unveiled findings from its new biology lab last week, its scientists claimed to have used AI agents to discover new enzymes with promise for biotechnology.

 

Some scientists were quick to cast doubt on the achievement. Then Mario Rodríguez Mestre, a computational biologist, said this weekend that he and his colleagues had been studying the enzymes and their associated molecules — which Anthropic calls ARTs — for four years.

 

Dr. Mestre and his colleagues have yet to publish their findings. But for the past three years, they have regularly used Anthropic’s AI models as they have written code, drafted manuscripts and performed other tasks.

 

In doing so, Dr. Mestre said he and colleagues shared key findings about the enzymes with Anthropic.

 

“So, for me, the most important question is not ‘Were ARTs already known?’ They were,” Dr. Mestre said in an interview. The issue instead, he added, is whether Anthropic’s AI actually reasoned its way to the results, or whether it was guided to them based partly on his own work.

 

“My concern is that this information was used to train future versions of the models,” Dr. Mestre said of his own research. “I think it’s important to raise this possibility.”

 

In a statement on Sunday, Anthropic said: “We are not aware of any previously published work describing the ART system we recently found. Claude was also not trained on any user transcripts, and our molecular biology team has no such access, either.”

 

Anthropic said that Claude’s key contribution was identifying an array of RNA molecules and a protein associated with reverse transcriptase in a system. But Dr. Maestre said that he and his colleagues had made that discovery over a year ago.

 

Similar worries surfaced recently when researchers at OpenAI claimed to have solved a longstanding mathematics puzzle. A leading mathematician noted that he had been using the company’s AI to work toward a solution for some time.

 

Anthropic, which is expected to go public at a valuation potentially as high as $2 trillion, has touted biological research as a central part of its mission.

 

“We believe this type of application will accelerate rapidly in the years ahead, and that we can make profound progress in the near future on diseases that have plagued humanity for millennia,” Dario Amodei, the chief executive of Anthropic, said in a speech to the United Nations on Wednesday.

 

In their report last week, Anthropic researchers said they instructed AI agents to search databases containing billions of gene sequences. The scientists did not describe the precise prompts they had used.

 

The databases contain DNA collected from the ocean, soil and other habitats. The genes belong mostly to microbes and viruses; scientists have only determined what a fraction of them do.

 

Anthropic’s scientists tasked the AI with finding genes that direct production of enzymes called reverse transcriptases. These enzymes read the sequence of an RNA molecule and make a matching DNA molecule with the same sequence.

 

Bacteria use some reverse transcriptases to sense invading viruses. Some viruses rely on reverse transcriptases as weapons against other viruses.

 

The Anthropic scientists provided the AI with five examples of reverse transcriptases. The agents wrote their own software to explore the databases for similar genes and to look closely at promising candidates.

 

The agents identified genes for possible reverse transcriptases in bacteria-infecting viruses known as jumbo phages. Each of those sequences was located close to an array of genes for RNA molecules, which may help the enzymes carry out their jobs.

 

The Anthropic scientists named the enzymes array-associated reverse transcriptases, or ARTs. In a report posted to their website, they said that their AI had taken “the first steps of a biological discovery on its own.”

 

On Sunday, Dr. Mestre disputed that claim. “They are the same systems we have been studying for years,” he said. And they are the same enzymes, he said, that he had described to Anthropic’s AI

 

Dr. Mestre first encountered the enzymes in jumbo phages in 2022 while searching for new reverse transcriptases. He worked briefly as a consultant for ReNegade Therapeutics, which filed a patent in 2023 on a number of reverse transcriptases, including the ones Dr. Mestre found. He is listed as a co-inventor on the patent filing.

 

After working for ReNegade, Dr. Mestre joined the University of Copenhagen to study reverse transcriptases in more detail, including the ones he found in jumbo phages. He and his colleagues nicknamed them “jumbotrons.”

 

Anthropic released Claude in 2023, a year after Dr. Mestre began researching jumbotrons. Like many scientists, he found it useful for an assortment of tasks.

 

“I completely switched my way of doing science,” he said.

 

During his conversations with Claude, Dr. Mestre shared a draft of both his dissertation and of a manuscript describing jumbotrons. He and his colleagues uploaded additional details while using Claude to perform various administrative tasks.

 

The New York Times reviewed Slack messages, figures and other materials produced by Dr. Mestre documenting his research and his conversations with Claude over the past several years. They seem to show that the AI was extensively involved in his research.

 

Seth Shipman, a bioengineer at the Gladstone Institutes and the University of California, San Francisco, said that he was familiar with Dr. Mestre’s work.

 

“We’ve been talking with him about these for years,” he said. “Many of us know about these systems.”

 

Dr. Mestre said he could not be sure that Anthropic had relied on his work to seek out jumbotrons. But he was skeptical, he said, that the AI agents could have managed to find them so swiftly.

 

Jumbotrons are rare, and their RNA genes are difficult to distinguish from random genetic sequences. Dr. Mestre also found some striking overlap between his work and the work presented by the Anthropic scientists.

 

The scientists claimed that their AI had independently retrieved the raw data from a 2022 experiment in which researchers recorded all the viral RNA molecules produced by a microbe infected by jumbo phages.

 

Anthropic’s AI reported that the microbes created high levels of the RNA molecules in ARTs. The scientists argue that this evidence supports their hypothesis that the reverse transcriptases they claim to have discovered rely on these RNA molecules to carry out a job for the jumbo phages.

 

Dr. Mestre said that he and his colleagues analyzed the same data a year ago, which helped them identify the RNA genes associated with the jumbotrons. They shared this unpublished information with Claude.

 

“The resemblance was striking to us,” Dr. Mestre said.

 

Without a closer look at what the Anthropic AI actually did, Dr. Mestre said that it would be impossible to say for sure that it was guided by his research.

 

“I mean, coincidences happen,” he said. “Everyone in the field is doing more or less the same things. But we were aware of these repeats more than one year ago.”

 

Dr. Mestre noted that AI companies also have been accused of training models on the work of artists and writers without copyright permission. “It’s already happened,” he said.

 

Dr. Shipman, of U.C.S.F., said that he and his colleagues, too, have leaned heavily on Anthropic’s tools for their research. “We’re also telling them what we’re working on,” he said.

 

The company’s announcement last week raised serious concerns among scientists who study these enzymes, he added.

 

“I’ve heard from a bunch of people in our field who say, ‘Uh oh, are they training on my prompts?’” Dr. Shipman said. “We can’t not think about it.”

 

The jumbotron controversy echoes another scientific conflict that played out earlier this month. Another artificial intelligence start-up, OpenAI, claimed to have cracked a major unproved problem in math.

 

But the New York University mathematician Tristan Buckmaster, who had been working with a colleague on his own solution, said that they had used OpenAI’s models in their work.

 

In a statement, Dr. Buckmaster raised the possibility that OpenAI had borrowed from their efforts. “Is it ethical to use customer’s data to try to scoop their customer?” he later asked on social media.

 

OpenAI later denied that its models had been trained on Dr. Buckmaster’s conversations. (The New York Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to AI systems. The two companies have denied those claims.)

 

Dr. Mestre had become increasingly enthusiastic about using AI for his science. Just last month, he began using Claude Science, a new Anthropic AI tool tailored to researchers.

 

But after Anthropic’s announcement last week, he is winding down his projects with Claude and shifting to other models.

 

“I’m shutting down everything,” he said.