Anthropic's biological discovery sparks debate about the role of artificial intelligence in scientific research

Officials in artificial intelligence companies are betting on a radical transformation in biological research, similar to what large language models have achieved in programming, software development, and mathematics. However, the announcement by Anthropic, the developer of the Claude model, that it had made a new biological discovery, sparked controversy regarding the authenticity of the discovery and the extent of the independence of artificial intelligence in making scientific discoveries.

The company said that its artificial intelligence agents have discovered a “new enzymatic system” in the genetic code of viruses, which may represent a new gene editing mechanism, similar to the CRISPR-Cas9 technology widely used in laboratories, whose innovators won the Nobel Prize in Chemistry in 2020. The company’s CEO, Dario Amodei, described the discovery as having great scientific potential, and wrote on the “X” platform: “Today we announced a discovery led by Claude of a molecular machine that we believe "It may represent a new mechanism for gene editing."

Anthropic said that the discovery was made “only with the general guidance of our scientists,” but experts who were not involved in the research stressed that the results are still preliminary and do not prove that gene editing has occurred or explain the mechanism behind it. The research has not been published in a peer-reviewed scientific journal.

Claims about the originality of the discovery Aaron Inglehart, an assistant professor at the University of Minnesota, said that the results are “preliminary and interesting,” but they do not yet prove that gene editing occurred. The researchers did not determine the function of the discovered viral sequences, or whether they could be used in practice.

Days after Anthropic's announcement, Mario Rodriguez Mestre, who used Claude extensively in his recent doctoral research at the University of Copenhagen, said his unpublished research addresses the same viral patterns. “It's almost the same discovery, not just studying two related protein families or similar biological systems,” Mistry told Ofoq News.

He explained that he had previously shared drafts of his thesis, analyzes and materials describing these systems through his Claude account, but stressed that he did not accuse the company of intentionally taking over his work, saying: “I do not claim that Anthropic intentionally took over our work, nor can I prove that.” He added that information related to his research may have reached the model during his training, but he believes that the most likely possibility is that the research process was based on prior scientific knowledge and much greater human guidance than the description of “independent discovery” suggests.

Ilya Finkelstein, a professor of molecular biological sciences at the University of Texas, Austin, said it is possible that different research groups will reach the same results, but Mistry's claim is dangerous, and may make scientists hesitant to share their unpublished research results with artificial intelligence tools. The Anthropic researchers said they asked Claude to search a large database of DNA sequences for “interesting new examples” of the enzyme reverse transcriptase (RT), which copies RNA into DNA.

According to the company, 950 artificial intelligence agents, capable of planning and executing tasks independently, spent 21 hours analyzing the data, and identified 3,500 enzymes, before narrowing the list to 20 promising enzymes. She added that an analysis of this type could take weeks or months by an experienced scientist.

The company said an agent detected a repetitive pattern of DNA sequences next to a gene responsible for an unusual reverse transcriptase enzyme. After additional analysis, he concluded that he had discovered a new biological system, and submitted a report for human review.

Laboratory scientists then analyzed and tested the result, describing the system as “never previously described.” It is found in bacteriophages, which are viruses that attack bacteria, and is similar in structure to the CRISPR system found naturally in bacteria. Finkelstein noted that four of the study's six authors are molecular biologists and experts in the field, and that the lead researcher, Peter Yun, previously worked in the laboratory of Jennifer Doudna, one of the pioneers of CRISPR-Cas9 technology and winner of the Nobel Prize in Chemistry in 2020, jointly with Emmanuelle Charpentier.

He argues that this means that AI agents benefited from the guidance of leading experts. Finkelstein wrote in a blog post on his lab's website that providing experts with massive computing power and advanced models could lead to further discoveries in bioinformatics.

But other researchers believe that talk of independent discoveries led by artificial intelligence is exaggerated. Gustavo Sodry, professor of genomic neuroimaging and artificial intelligence at King's College London, said that analyzing data on a scale beyond human capabilities contributes to the progress of research, but does not necessarily mean achieving a scientific breakthrough.

He added: “In my experience, most of the work is in bridging the gap between spotting an interesting pattern and arriving at useful knowledge.” Frontiers of AI in Science Earlier this year, Anthropic launched its Claude Science research tool, while competing companies have unveiled similar tools, including OpenAI's GPT-Rosalind, Microsoft's Quine, and Google DeepMind's Co-Scientist.

These tools combine large language models with specialized scientific data, such as DNA sequences. But its reliability depends on the data it was trained on.

For example, Google DeepMind's AlphaFold tool, whose developers won the 2024 Nobel Prize in Chemistry for its contribution to predicting protein structures, draws its power from specialized data collected over decades. With fears mounting that AI agents are spiraling out of control, biological discoveries may allow companies to present a more positive image of their products.

However, achieving tangible scientific results may require integrating these models into laboratory experiments using robots. Yuval Ilani, assistant professor of biochemical technologies at Imperial College London, said that biology is different from mathematics, and that artificial intelligence needs to be embodied in the physical world through robots capable of carrying out precise experimental tasks.

“Based on what I've seen in robotics and automation, we're still a long way from achieving that,” he added. “So it will be a while before we can justify the hype around AI in biology.”

Anthropic recognizes the limits of what AI can accomplish on its own, and is investing in developing hardware that allows its agents to operate physical equipment safely. Soudry stresses that the role of humans in scientific research is not diminishing, warning that the danger is not that scientists are excluded, but rather that they stop using their judgment to evaluate the importance of the results.

The next step for Anthropic researchers is to conduct laboratory experiments to understand the discovered pattern and determine whether it can be practically employed in biotechnology. “Certainly, some of these tasks could be assigned to robots in the future,” Sodry said.

But questions worthy of scientific research do not remain hidden in databases waiting to be discovered, but rather stem from us, from what we care about and what worries us.”