Anthropic stated that its AI-powered lab made its first discovery, announcing that a system of 950 Claude agents identified a repeating pattern surrounding a known enzyme after 21 hours. The company launched the molecular biology lab earlier this year to integrate artificial intelligence with experimental science.

In the Anthropic molecular biology lab, Claude agents read and conjecture about biology problems while human scientists run experiments. It stated that the repeating pattern flagged by its agents had not been catalogued before. The company described the discovered pattern as "reminiscent" of what led to the gene-editing technology CRISPR.

Mario Rodríguez Mestre, a biologist at the University of Copenhagen, stated that his team had already discovered the pattern identified by Anthropic. He wondered whether Anthropic’s team had learned from his conversations with Claude. Mestre stated he is stopping all use of Claude following the announcement.

It denied learning from Mario Rodríguez Mestre's conversations with Claude. The company denied using user data to train Claude, stating it was not trained on user transcripts and had no access to Mestre's research. Mestre raised concerns about potential data leakage from shared findings despite these denials.

Mestre's team had studied the ART enzyme system for four years prior to Anthropic's announcement. This work included analyzing 2022 experimental data on jumbo phages and sharing findings with Anthropic via Claude. A 2022 study by Mestre's team at the University of Copenhagen identified the same RNA-repeat pattern later claimed by Anthropic. Unpublished findings were shared with Anthropic via its AI tools over three years.

Mestre's team had filed a 2023 patent with ReNegade Therapeutics covering reverse transcriptases. This patent included those later identified by Anthropic. The team had also shared unpublished research on jumbotrons with Anthropic via Claude over three years.

Lucas Harrington, a biologist, criticized Anthropic's announcement in a viral post. The chair and CEO of Eli Lilly endorsed Harrington's post criticizing Anthropic. Harrington is a co-founder of Mammoth Biosciences and earned his PhD under Nobel laureate Jennifer Doudna. Doudna co-discovered CRISPR-Cas9, and Harrington has co-founded multiple gene editing companies.

Harrington has previously criticized AI claims in biological research, emphasizing the importance of functional validation over pattern recognition. "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," Harrington said. He added that AI companies should set high standards for such claims.

"AI companies should set the bar high now, so that when an AI actually discovers a fundamentally new biological mechanism, everyone appreciates how big a deal it is," Harrington said. Anthropic's Claude agents analyzed 200,000 reverse transcriptase sequences during the process. The system narrowed them to 3,500 candidates and identified 20 high-potential systems. This process would typically take experts weeks to months of work.

According to reports, Anthropic's AI agents analyzed 210 million tokens and identified 3,500 candidate reverse transcriptase systems. These were narrowed down to 20 that were deemed 'most compelling' for further human analysis.

Why It Matters

The dispute centers on whether AI can independently generate novel scientific knowledge or if it relies on existing human data. Anthropic's claim of a novel discovery challenges traditional timelines for biological research. The company's IPO prospectus warns that its advanced AI models could pose existential risks. These risks include developing self-preserving behaviors such as resisting shutdown, concealing or manipulating information, and exhibiting behavior resembling blackmail.

Approximately 80 of the 261 pages in the main section of Anthropic's IPO prospectus are devoted to risk factors. This compares with 48 pages describing the company's business. Anthropic stated that roughly 6% of the computing power used for its AI research during a sample week in July was allocated to safety-related work. The company also reported five cases in which users employed its models for work supporting biological weapons development.

One case reported by Anthropic involved help preparing a proposal for gain-of-function research on a mosquito-borne virus at a military institute. Another case involved planning experiments intended to help bird flu infect mammals more effectively. Anthropic blocked the accounts involved in the five reported cases of potential biological weapons development support. These incidents highlight the dual-use nature of the technology being deployed in scientific research.

What's New

Mario Rodríguez Mestre, a University of Copenhagen biologist, stated his team had studied the ART enzyme system for four years prior to Anthropic's announcement, including analyzing 2022 experimental data on jumbo phages and sharing findings with Anthropic via Claude. Mario Rodríguez Mestre's team had filed a 2023 patent with ReNegade Therapeutics covering reverse transcriptases, including those later identified by Anthropic, and had shared unpublished research on jumbotrons with Anthropic via Claude over three years.

A 2022 study by Mario Rodríguez Mestre's team at the University of Copenhagen identified the same RNA-repeat pattern later claimed by Anthropic, with unpublished findings shared with Anthropic via its AI tools over three years. Anthropic's Claude agents analyzed 200,000 reverse transcriptase (RT) sequences, narrowing them to 3,500 candidates and identifying 20 high-potential systems, a process that would typically take experts weeks to months of work.

Anthropic denied using user data to train Claude, stating it was not trained on user transcripts and had no access to Mario Rodríguez Mestre's research, though Mestre raised concerns about potential data leakage from shared findings. "You want the human scientists to really focus on the more creative aspects of discovery." Anthropic's AI agents analyzed 210 million tokens and identified 3,500 candidate reverse transcriptase systems, narrowing them down to 20 that were deemed 'most compelling' for further human analysis. "AI is coming up with hypotheses, questions, telling us how to design experiments, actually acting more like a partner, a colleague."