Anthropic's AI Scientific Discovery Claim Triggers Scientific Controversy Over What Counts

Anthropic Claims AI-Driven Breakthrough in Biological Research

Anthropic announced on September 23 that its AI model Claude had identified an enzyme system with properties reminiscent of CRISPR, the gene-editing tool that won a Nobel Prize1

1

. The company deployed approximately 950 Claude agents running simultaneously, which identified the system in 21.5 hours after analyzing over 200,000 reverse transcriptase candidates from genomic databases2

2

. This finding represents the first result from Anthropic's molecular biology lab, established earlier this year to accelerate drug discovery3

3

. CEO Dario Amodei has repeatedly stated that AI will cure most diseases within the next 10 years, positioning this discovery as a first step toward that ambitious goal4

4

.

What Claude Actually Found in Jumbo Phages

The AI agents identified array-associated reverse transcriptases, or ART, found in jumbo phages—large viruses that infect bacteria3

3

. Reverse transcriptase enzymes copy RNA into DNA, the opposite of normal cellular processes, and are used by various organisms including HIV3

3

. What made this particular reverse transcriptase notable was its association with a long array of repeating DNA sequences resembling CRISPR's characteristic structure2

2

. One AI agent wrote: "I can see by eye a tandem repeat array ... that's a Crispr-like ... repeat array?!"2

2

. However, Anthropic acknowledged in its announcement that "we don't yet understand what this system does," though only a handful of known systems share these features, and all can cut, copy, and paste DNA2

2

.

Scientists Challenge the Discovery Claims

The announcement triggered immediate pushback from biologists who questioned whether this constitutes genuine AI scientific discovery. A viral post from one biologist, endorsed by the chair and CEO of drugmaker Eli Lilly, stated that "finding a weird cluster of genes and repeats is often the easy part. The hard part, and where the real discoveries come from, is figuring out what the system actually does"1

1

. Jason Gill, a microbiologist at Texas A&M University, revealed that he and his colleagues had already identified this same reverse transcriptase in a 2021 paper on jumbo phages2

2

. What might be new is Claude's identification of repeats around the enzyme hinting at a CRISPR-like system, but Gill noted that ART appears to have no "obvious relationship" to any known CRISPR system2

2

.

Computational Biologist Claims Prior Discovery and Raises Training Concerns

Mario Rodríguez Mestre, a computational biologist at the University of Copenhagen, disclosed over the weekend that his team had been studying these enzymes—which they nicknamed "jumbotrons"—for four years5

5

. Mestre and his colleagues filed a patent in 2023 through ReNegade Therapeutics covering these reverse transcriptases, where Mestre is listed as co-inventor5

5

. More concerning, Mestre revealed that for the past three years, he regularly used Anthropic's AI models while writing code, drafting manuscripts, and performing other research tasks, potentially sharing key findings about the enzymes with Anthropic5

5

. "My concern is that this information was used to train future versions of the models," Mestre stated5

5

. Anthropic denied these allegations, stating: "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"5

5

. Mestre has stopped all use of Claude anyway1

1

.

The Real Value: Pattern Recognition in Massive Genomic Databases

Despite the scientific controversy, experts acknowledge that AI in biology demonstrates genuine utility in managing overwhelming data volumes. The Human Genome Project took 13 years and cost $2.7 billion to sequence 3 billion letters of human DNA4

4

. Today, sequencing a full genome takes just over a day and costs as little as $200, creating publicly available genomic databases so massive they've become unmanageable4

4

. A 2012 committee of bioinformatics experts found that the "bottleneck in scientific productivity" had shifted from data production to data analysis4

4

. Identifying new reverse transcriptases can take months of mining genome databases manually, so the fact that Anthropic completed this in under 24 hours is impressive2

2

. "These models are good at finding patterns, better than a person staring at it with their eyeballs can," Gill explained2

2

.

Why the Framing Matters for Scientific Progress

The core issue extends beyond this specific finding to how AI companies frame their technology's role in science. AI companies aren't presenting their systems simply as tools scientists can use, like microscopes or supercomputers—they're insisting that AI systems are making discoveries themselves1

1

. This approach conflicts with how science actually works, where new knowledge typically emerges from collaboration and an ever-growing arsenal of tools1

1

. Once the standard becomes whether Claude itself made a discovery, all evidence becomes fodder for a binary debate: breakthrough or bust1

1

. Whittling 200,000 candidates down to a few worth exploring represents legitimate scientific work, and the fact that a general-purpose chatbot could accomplish this is notable, even if humans helped steer it and ultimately ran the experiments1

1

. Fyodor Urnov, a gene-editing expert at UC Berkeley who coined the term genome editing, told reporters that "the jury is out" on whether what Claude found is even an enzyme at all4

4

. Similar concerns emerged when OpenAI claimed its agents solved a million-dollar mathematics problem, only for skeptics to question whether it was the problem mathematicians actually cared about most1

1

. As Anthropic and OpenAI race to one-up each other, raising the bar for what constitutes AI-driven breakthroughs in science might be the last thing on their minds1

1

.