The internet did what the internet does. Within hours of AI Anthropic's announcement, YouTube thumbnails and social-media posts were declaring that artificial intelligence had "cracked the secret of DNA," "unlocked the code of life," or produced something comparable to CRISPR. Those headlines are not merely exaggerated. They describe something quite different from what actually happened.
There are inevitably a few technicalities here, but at this stage of the AI game technicalities are the least of our worries. Put very simply, AI has not unlocked the meaning of human DNA. It has used enormous computational power to sift through a gigantic biological database and identify an intriguing arrangement involving a known type of enzyme. That could eventually prove important, but nobody yet knows what the newly identified system actually does.
On 23 September 2026, Anthropic published a preprint and company account describing the first public result from its new life-sciences group and Bay Area laboratory. The company set roughly 950 Claude AI agents loose on an enormous collection of genetic information. They worked for about 21 hours and consumed roughly 210 million tokens while searching through a database reportedly containing around 1.9 billion groups of proteins.
The target was a type of enzyme called a reverse transcriptase. There is no need to become a molecular biologist to understand the basic point. Enzymes are proteins that help chemical reactions happen, while reverse transcriptases are enzymes capable of copying information from RNA into DNA. They are already well known to science. The question was whether the AI could find interesting examples or arrangements hidden inside an amount of genetic data far too large for scientists simply to sit down and read.
The Claude agents assembled and examined enormous numbers of candidates. One of them eventually noticed something peculiar next to a reverse transcriptase in the genome of a bacteriophage, which is simply a virus that infects bacteria. Alongside the enzyme was an unusual pattern of repeated DNA. The agent described the pattern as "spectacular" and thought that its arrangement looked somewhat like CRISPR.
Anthropic has called the resulting three-part arrangement ART, for array-associated reverse transcriptases. The important point is that the reverse transcriptase itself was not newly discovered. What appears to be new is recognising that the enzyme, another nearby gene whose function remains mysterious, and this repeated stretch of DNA may belong together as some kind of biological system.
This is where the internet hype machine went into overdrive. CRISPR became famous because scientists discovered that certain repeated sections of bacterial DNA formed part of a defence system against viruses. That natural machinery was eventually developed into an extraordinarily powerful method for editing genes. Consequently, tell people that an AI has found something that "looks like CRISPR" and it is only a matter of minutes before someone announces that artificial intelligence has discovered CRISPR 2.0.
But looking like CRISPR and working like CRISPR are entirely different things. Scientists do not yet know what ART does. There is no demonstration that it edits genes, targets particular genetic sequences or functions as a bacterial defence mechanism. It could turn out to perform an important biological job, but it could also have a much more mundane function associated with the life cycle of the viruses in which it occurs.
Anthropic is quite open about this uncertainty. The company says the primary function remains unknown and laboratory investigation is continuing. CEO Dario Amodei likewise acknowledged that the system's precise function, possible usefulness in biotechnology and ultimate scientific significance have yet to be established. That is rather different from the online story that AI has suddenly deciphered one of the secrets of life.
There was nevertheless a genuine discovery here. Follow-up searches found numerous examples of these ART arrangements in giant bacteriophages, and laboratory experiments showed that the repeated DNA is copied into several small pieces of RNA. Something biologically interesting therefore appears to be happening. The problem is that scientists do not yet know what that something is.
The claim that AI is now "doing biology without humans" needs similar qualification. Human researchers decided what problem to investigate. Humans built the computer systems through which the Claude agents operated. Humans provided the scientific tools and databases, and humans performed the laboratory experiments needed to establish that the interesting DNA identified by the AI really was producing RNA.
What the AI did was still impressive. Rather than being told exactly which genetic sequence to examine, the agents searched enormous quantities of information, compared possibilities, consulted scientific literature and decided which candidates deserved closer attention. One agent chose to examine the raw DNA surrounding a particular enzyme and noticed a pattern that human researchers had apparently not previously recognised as a potentially unified system.
That represents an interesting change in the scientific workflow. Think of the AI less as Einstein suddenly discovering a new law of nature and more as several hundred extraordinarily fast research assistants sent into a library containing more material than any human research team could possibly read. They can search, compare and bring unusual things back to the scientists. The scientists must then determine whether those things actually mean anything.
Feng Zhang, one of the scientists who helped transform CRISPR into a genome-editing technology, reviewed Anthropic's preprint and described the association between the RNA-producing repeats and reverse transcriptases as genuinely intriguing. He also regarded it as an interesting demonstration of how AI agents might contribute to biological discovery. But the crucial scientific qualification remained: the finding requires further investigation.
There is another reason for caution that comes from Anthropic's own account. When the company repeated the same AI search, subsequent runs did not rediscover the interesting DNA array because the agents failed to examine the relevant neighbouring DNA. When models were directly shown the genetic sequence, they could often recognise the unusual repeats. When they had to navigate the databases and research tools themselves, their success rate fell.
That is important. We are not looking at a machine that can be instructed to "go and make a biological discovery" and reliably return with one. There was an element of contingency in what happened. One AI agent happened to investigate the right piece of surrounding DNA, noticed something interesting and reported it to the scientists. Other runs of essentially the same process missed it.
None of this makes the experiment unimportant. Quite the opposite. Biology has accumulated gigantic databases containing genetic information from organisms, viruses and microbes across the planet. There is now vastly more information than human scientists could individually inspect. AI may prove extremely useful precisely because it can rummage through these enormous digital warehouses looking for patterns that human researchers have overlooked.
Anthropic has also built a laboratory so that computational suggestions can be tested in the physical world. That may prove more important than the present ART discovery itself. An AI can search billions of pieces of biological information and say, in effect, "this looks odd; somebody should investigate it." Human scientists can then perform experiments to determine whether the apparent pattern represents genuine biology or merely an interesting coincidence.
That combination could greatly accelerate some forms of scientific research. AI performs the brute-force search, scientists exercise judgment, and laboratory experiments decide whether either of them was right. It is much less spectacular than an artificial intelligence sitting down and unlocking the secret of life, but it is also much more believable.
So what actually happened? AI did not decode human DNA. It did not discover the meaning of the human genome. It did not create a new gene-editing technology, and it certainly did not solve the mystery of life. A swarm of Claude agents searched an enormous genetic database and one agent noticed an unusual arrangement surrounding a known type of enzyme in viruses that infect bacteria. Scientists subsequently confirmed that part of this arrangement produces RNA, suggesting that the pattern may indeed have some biological function.
What that function is remains unknown. ART could eventually become a useful biotechnology, reveal an interesting new mechanism used by bacteriophages, or turn out to be considerably less important than the excitement surrounding its discovery suggests. Science has now reached the stage where experiments rather than headlines have to decide.
The larger story, however, is worth watching. AI's contribution to science may initially be less like replacing the great scientist and more like supplying science with an army of tireless research assistants. Give them access to mountains of information that no human being could read in a lifetime, let them search for anomalies and patterns, and occasionally one may return carrying something genuinely interesting.
That is impressive enough without pretending that AI has unlocked the secret of DNA. The internet turned an intriguing pattern into a revelation because "AI spotted something strange in a huge genetic database" does not generate quite the same clicks as "AI has decoded life." As usual, the real science is slower, messier and considerably more interesting than the headline.
https://theconversation.com/an-ai-model-has-found-a-new-crispr-like-biological-system-heres-what-it-means-for-science-292777
https://www.youtube.com/watch?v=JSA3mbGQ9qA
https://www.anthropic.com/news/claude-discovers-novel-enzyme-system
https://www-cdn.anthropic.com/22573675ada52a8ca8a97a1a4b4326b2f208a071.pdf