AI’s Real Gift to Science

Dario Amodei, the chief executive of Anthropic, has often said that AI will cure most diseases in the next 10 years. Late last month, he claimed that his company had taken a first step toward that future: Claude had, in less than 24 hours, discovered a new “enzyme system” that the company says resembles CRISPR, a gene-editing tool currently curing people of diseases. While announcing the discovery, Anthropic hinted that this was the type of finding that revolutionizes medicine and quoted a gene-editing pioneer who called it “genuinely intriguing.”

The company was open about not knowing what exactly it had spotted or whether the finding—the first released from a biology lab that the company established this spring—was really that significant. Still, many academics, along with the CEO of Eli Lilly, pointed out that the company had dressed up a very incremental finding as a major discovery. Fyodor Urnov, a scientist who coined the term genome editing and is working on a CRISPR project with Anthropic, told me that “the jury is out” on whether what Claude found is even an enzyme at all. Multiple researchers told The New York Times that, in fact, they had been aware of the supposedly new enzyme system for years, including a team at the University of Copenhagen that alleged Anthropic’s new results might have been stolen from their lab’s interactions with Claude. The company, in response, said that Claude isn’t trained using user transcripts. (Anthropic did not respond to a comment request for this article.)

In other words, Anthropic made a big deal about a finding that—although achieved in a novel, Claude-led way—may not lead to any biomedical advances, and that the company knew may not lead to any biomedical advances. Ironically, in making such grandiose statements, the company and other AI boosters obscured what Claude does appear to have accomplished: It sorted through reams of biological data that would have taken scientists much longer to analyze, if they could have at all. This capacity might not be a miracle, but for scientists facing mountains of data that have surpassed their capacity for analysis, it’s still meaningful.

Around the turn of the 21st century, researchers sequenced for the first time all 3 billion letters that make up human DNA. This effort, called the Human Genome Project, took 13 years and cost $2.7 billion. Today, sequencing a person’s full genome takes just over a day and costs as little as $200. Accordingly, the total amount of publicly available genomic data has become so large as to be unmanageable. In other fields, such as protein science, technological advances have led to a similar boom in the sheer amount of data being collected.

Trying to pull anything useful from all that data has become its own long-term project. In 2012, a committee of bioinformatics experts found that the “bottleneck in scientific productivity” had moved from data production to data analysis, but early efforts at the National Institutes of Health to deal with that bottleneck quickly sputtered out. If researchers could efficiently store, send, and examine all of the information they generated, they would likely be able to, for example, develop new drugs by identifying better genes or cellular parts to target. But it’s simply too massive for them to comprehensively attack.

The biological corpus is not, however, too large for AI agents. Bo Wang, an assistant professor at the University of Toronto and the chief AI scientist at a drug-discovery company, recently wrote that AI’s biggest impact in biology will come from its help “scaling scientific attention”: Setting agents loose on the enormous data sets can surface the outliers that warrant further human inspection. Agents can propose and test hypotheses for what might cause the unusual patterns they find, and predictive modeling can help scientists choose what experiments to carry out on their actual lab benches. Eventually, automated labs might even carry out those chosen experiments—a few already exist in the United States, and the Trump administration is spending hundreds of millions to establish more. Meanwhile, scientists remain divided over whether AI models actually steer them toward potentially paradigm-shifting experiments, or safer, less exploratory ones.

AI is already changing medicine: Models that use people’s medical history to predict their health risks, including their likelihood of developing certain kinds of cancer, are now used in most U.S. hospitals. At the same time, not every field of biology generates enough data for this approach to work. AI might be less useful in some domains that rely on less abundant microscopy data, such as the study of how cells and tissues communicate; such data tend to be harder to come by than genetic sequences, Stirling Churchman, a professor of genetics at Harvard Medical School, told me. Organelle biology and the science of cellular transport similarly require painstaking, time-consuming experiments.

Still, Churchman said, “there will be magic where there’s data.” She spent the summer at OpenAI working on a specially trained model for life-science researchers. She anticipates that by allowing scientists who can’t code to examine their data themselves, AI will enable researchers to “analyze data at least 10 times faster, freeing us to spend more time on the biological questions and experiments.”

The recent Anthropic finding is an example of how AI can identify outliers that might otherwise have gone unnoticed. Existing tools can comb through heaps of genomic data (and have already been used to identify bona fide new gene-editing systems). What AI can do is run those kinds of analyses at a much larger scale, while also evaluating intermediary results, Kendell Clement, who runs a computational-biology lab at the University of Utah, told me. Of course, that scale requires dumping in correspondingly huge resources: Anthropic directed about 950 Claude bots to simultaneously pore over a database that contained nearly 2 billion DNA entries.

AI companies have every incentive to keep evangelizing about the medical miracles they say their models will unlock, especially as backlash against them mounts. Amodei has an idea for how to quiet AI skeptics: “The thing that will work is actually curing cancer,” he recently wrote on X. That would certainly be a coup. But helping scientists work more quickly is, too, even if it doesn’t sound quite so revolutionary.