I do not pretend to understand the biology. I am interested in what happens when scientists give AI an enormous amount to look through and then ask the right questions of what comes back.
I was reading about a new discovery in DNA. I have my computer highlight stories like this because I am curious about how we use AI, especially when a headline makes it sound as though the machine has done the whole job. This one said that Claude had discovered a new enzyme system. I pictured a robot arm juggling chemicals and bits of technology, while a quiet scientist in horn rimmed glasses sat at his desk playing Hearts or doomscrolling Grumpy Cat videos. Lucky fellow.
So I looked past the headline to see what everybody had actually been doing.
The clue in the DNA
The scientists at Anthropic asked Claude to search a vast collection of DNA records for interesting examples of things called “reverse transcriptases”. Those who know me will know I had no idea what they were until I looked them up. They are enzymes that copy RNA into DNA. That is quite enough biology for me to get through this story.
Roughly 950 AI agents spent 21 hours on the search. They gathered more than 200,000 examples and whittled thousands of possible systems down to a short list. One agent noticed repeated stretches of DNA beside an enzyme gene. The scientists investigated that lead and described a previously uncharacterised system they call array-associated reverse transcriptases, or ART. That name belongs to the biological system, by the way. Claude has not taken up painting. [1]
The enzyme itself had been seen before. The intriguing part was the arrangement around it, which nobody appears to have recognised as a system. The scientists carried out some initial experiments, but they still do not know what ART actually does. It may turn out to be useful; we cannot put it to work merely because its name sounds impressive. I certainly would not call it the next CRISPR while the researchers are still trying to find out what it is. [1]
Here is what interests me. The agents could search an enormous pile of material, compare what they found with earlier work, go back to check an oddity and bring the scientists a set of questions worth asking. I can imagine a human researcher doing some of that. I cannot imagine asking one to read 200,000 records before lunch. Perhaps this particular pattern would have been spotted soon anyway; we do not know. But it is easy to see how a tool that searches quickly and checks its own leads might shorten the wait for the next useful question.
A pile of invoices
I find it easier to understand if I leave the DNA alone and think about invoices. Suppose I ask an agent to look through years of them. It finds a cluster of unusually large charges from one supplier and flags a possible problem. That is useful. Then the finance manager checks the purchase order and remembers that we replaced all the old lighting in one planned purchase. The cluster is real. The suspicion is wrong.
The agent has saved us the effort of searching through the lot. The manager has supplied the context that makes sense of what it found. I would rather have both than spend a fortnight hunting invoices only to accuse the electrician of grand theft lightbulb.
In the laboratory, the scientists have a much harder version of that job. Claude found other candidates too. The team has to decide which deserve attention, work out what to test and make sense of the results. They are the people who may eventually discover what ART does and whether anyone can use it. I may be hazy on reverse transcriptases, but I can see why that part takes a scientist.
That is also why I am wary of reading the headline as proof that a machine understands a discovery in the way its researchers do. We have evidence that it found an intriguing lead. We have the beginnings of evidence from human experiments. We do not yet know the lead’s full meaning. Human scientists bring experience, intuition and occasionally sheer luck to that search for meaning. AI gives them more chances to exercise those talents: more ground covered, more oddities noticed, more ideas that can be tested. Some ideas will be wrong. We learn that by checking, not by admiring the speed at which they arrived.
The framework that makes useful work possible
I have written elsewhere about agents crossing boundaries while pursuing a task. My excitement about this research has not made that concern disappear. The people who build and use agents need to give them a proper moral and legal framework: clear authority, enforceable limits on their tools, independent testing and a record of what they tried to do. We should be able to police those rules and hold people or organisations responsible when they fail. The software will hardly be troubled by a fine; someone who chose its permissions might be.
I want the rules because I want more work like this. Give scientists the means to search data at a pace none of us could manage alone. Let them choose what is worth investigating and test what the machine brings back. If it helps them invent, discover and theorise sooner, that is a very good reason to be interested in AI.
We have given the frightening AI headlines plenty of attention, and often rightly so. Here is a hopeful one that also deserves its share. A machine helped bring a promising question out of a mountain of data. Now I want to see what the scientists do with it. I also want us to build the framework that lets them keep doing it safely.