For years, the promise of Artificial Intelligence in the life sciences has been framed as a "holy grail"—a tool capable of scanning the vast, messy, and complex library of biological data to uncover secrets that would take human researchers lifetimes to decode. Last week, Anthropic, the AI research company known for its Claude large language model (LLM), claimed to have reached a significant milestone in this pursuit. The company announced that its AI had autonomously identified a novel enzyme system with functional characteristics "reminiscent of CRISPR," the Nobel Prize-winning gene-editing technology.
However, as the dust settles on the initial PR wave, the scientific community is engaged in a rigorous debate. Is this a genuine breakthrough in autonomous discovery, or is it a case of high-tech pattern matching overhyped by a company eager to demonstrate the utility of its models in the physical world?
The Discovery: What is ART?
On September 23, Anthropic unveiled a technical report detailing the use of Claude to sift through massive genomic databases. The goal was to identify "interesting new examples" of reverse transcriptases—proteins that copy RNA into DNA. While this process is the inverse of standard cellular transcription, it is a critical biological mechanism used by various organisms, including viruses, to insert genetic segments into a host genome.
Using roughly 950 AI agents running in parallel, Anthropic’s system parsed through genomic data for 21.5 hours. The agents initially flagged over 200,000 potential reverse transcriptases. Through successive rounds of filtering and refinement, the model narrowed this pool down to a handful of candidates, eventually highlighting an "unusual" family of enzymes found in jumbo phages—large viruses that infect bacteria.
Anthropic has dubbed this discovery "ART," short for Array-Associated Reverse Transcriptases. The system features a long region of repeat DNA sequences that mirror the structural layout of CRISPR. One AI agent, upon analyzing the sequence, noted: "I can see by eye a tandem repeat array… that’s a CRISPR-like… repeat array?!"
A Chronology of Scientific Scrutiny
The timeline of this discovery is critical to understanding the skepticism currently surrounding it.
- 2021: Microbiologist Jason Gill of Texas A&M University and his colleagues published a study in which they identified the specific jumbo phage reverse transcriptases that Anthropic’s model has now highlighted.
- Early 2024: Anthropic establishes a dedicated "wet lab" for drug discovery, signaling its intent to move beyond text-based AI into experimental biology.
- September 23, 2026: Anthropic publishes a blog post and a non-peer-reviewed technical report claiming their AI discovered the enzyme system.
- Post-Announcement: The scientific community begins to analyze the report, noting that the "discovery" is more of a refinement of previously known data rather than the identification of a completely new biological phenomenon.
Critics, such as Stanford professor Le Cong, have pointed out a significant disconnect between the marketing and the science: "The experiments are still in the queue. The PR is already live." The central point of contention is that while the AI identified a pattern, the actual functional verification—determining whether this enzyme can, like CRISPR, be programmed to cut and edit DNA—remains entirely unproven.
The "Shiny Object" Dilemma: Scientific Implications
The primary value of AI in this context, according to researchers, is its ability to act as a high-speed filter. Seth Shipman, an associate investigator at the Gladstone Institutes, likens the process to finding a diamond on a beach. "The novel thing is how they found it, not what it is," Shipman says.
However, there is a fundamental tension in how these findings are reported. If an AI is tasked with searching for patterns it has already been trained to recognize, is it truly "discovering," or is it simply retrieving?
"Let’s say we are on Santa Monica Beach and trying to scan through all the sand to find a diamond," says Le Cong. "AI found this thing that looks very shiny, and then you have to go back to the lab to know—is it glass? Is it a diamond?"
The scientific community is particularly concerned about the opacity of Anthropic’s training data. In modern biology, reproducibility is the cornerstone of progress. When a major journal publishes a discovery, it demands that the code and the underlying methodology be transparent. Anthropic’s proprietary model, however, is a "black box." This has led to concerns—and even accusations—that the model may have been trained on unpublished research. While Anthropic has officially denied that Claude was trained on user transcripts or private biological datasets, the lack of transparency remains a hurdle for academic validation.
Official Responses and the Corporate Narrative
Anthropic’s leadership has been vocal about the potential for AI to bridge the gap between digital data and physical laboratory work. CEO Dario Amodei has suggested that the future of this technology lies in "closed-loop" discovery, where the AI not only identifies a potential target but also autonomously controls the laboratory equipment to test its own hypotheses.
"Eventually it may even be possible for Claude itself to safely perform the experiments by autonomously controlling lab equipment," Amodei stated in a post on X.
While this vision captures the imagination, it also raises significant safety and oversight questions. Anthropic was quick to clarify that its current lab operations are conducted at the lowest biosafety levels, meaning there is no risk of the AI creating dangerous biological agents. Nonetheless, the suggestion that an AI could "autonomously" conduct wet-lab experiments strikes many biologists as a distant, and perhaps premature, goal.
The Long Arc of Scientific Progress
To put this discovery into perspective, one must look at the history of CRISPR itself. The repeat sequences that define the CRISPR system were first observed in bacteria in 1987. It took 25 years of grueling, incremental research by scientists like Jennifer Doudna and Emmanuelle Charpentier to move from identifying those sequences to demonstrating that they could be used as a programmable tool for genome editing.
Today, despite the ubiquity of CRISPR in research labs, there is only one FDA-approved drug that utilizes the technology. This illustrates a harsh reality of biotechnology: the transition from a "shiny" finding in a database to a clinically relevant medicine is a journey measured in decades, not hours.
The current excitement surrounding Anthropic’s ART discovery must be tempered by this reality. As Fyodor Urnov of UC Berkeley notes, there is value in the company’s efforts, but the "discovery" is only the first step in a very long, very difficult process.
Conclusion: The Path Forward
Can AI models like Claude actually accelerate biological discovery? The answer is likely yes, but perhaps not in the way the press releases suggest.
If we truly want to test whether an AI is an "AI Scientist," Le Cong proposes a more rigorous benchmark: "If you have an AI that only trained on knowledge from before people ever discovered CRISPR, and then that AI actually discovered CRISPR, that seems to be a better setup to test an AI scientist."
For now, Anthropic has provided a proof of concept—a demonstration that LLMs can be used to scan large genomic datasets to identify patterns that might escape the human eye. Whether those patterns lead to a revolution in gene editing or simply represent an interesting, albeit minor, biological footnote remains to be seen. The true "discovery" will not happen in a server farm, but in the slow, meticulous, and inherently human environment of the wet lab, where the hypotheses generated by silicon must eventually be tested against the stubborn reality of biology.
