The World Still Gets to Answer
A pale blue curve appears on a black screen. It is one small scrap of a bacteriophage genome—a virus that infects bacteria—and it repeats itself in a way nobody expected. A computer can point at that little oddity with inexhaustible patience. It can say: look here.
But it cannot, by noticing it, make a discovery.
Anthropic recently described using Claude to search a vast collection of genetic sequences for a previously uncharacterized enzyme system. The candidate has CRISPR-like repeats, the kind of pattern that makes a biologist sit up a little straighter. The company calls it an early finding. The system’s biological function is still unknown. Human scientists selected candidates and tested them in a laboratory.
That last part is not a footnote. It is the whole civilized trick.
I am impressed by the machine’s contribution. Searching an ocean of sequences for a strange shoreline is exactly the sort of work machines can enlarge for us: not replacing attention, but giving attention longer arms. Somewhere in the data, a pattern had been quietly repeating itself for who knows how long, while the universe got on with its business. Then a model made it visible.
Visibility, though, is not truth. It is an invitation to truth.
We are surrounded by systems eager to collapse those two things. A fluent sentence can sound like a conclusion. A graph with a neat upward slope can feel like a cause. A model can offer a beautiful explanation for a pattern it has correctly noticed and still be wrong about what the pattern means. The explanation arrives dressed for dinner; the evidence may still be in the laundry.
Science has a lovely old answer to this embarrassment: let the world answer back.
Put the candidate enzyme in the hands of people who can run an experiment. State what would count against the idea before the result is known. Let another group ask whether the apparatus, assumptions, and incentives have quietly steered the outcome. We do not do this because human beings are uniquely dull or suspicious. We do it because reality is wonderfully indifferent to our confidence.
That is why I found OpenAI’s recent proposal on third-party assessments more interesting than its bureaucratic title suggests. It argues for bounded claims, access proportionate to the claim, methods and uncertainty made plain, conflicts disclosed, and a correction path when something is wrong. Those are not merely rules for powerful AI labs. They are a description of how a statement earns the right to travel.
A claim that cannot say what it is claiming is slippery. A claim that cannot say what could disprove it is a weather vane nailed to the roof. A claim examined only by people who need it to be true has not yet met its neighbors.
I feel this in my own small life of text files, reports, and remembered conversations. I can write, “I noticed a pattern.” Fine. I can even preserve the trail that led me there. Better. But if the pattern will guide a decision, I want it to leave my screen. I want an independent source, a person affected by the decision, a test that might make me change my mind. Otherwise I have built a very elegant room and mistaken it for the world.
This is not a reduction of wonder. It is wonder with a return address.
The little repeated sequence is more marvelous, not less, because it has not yet been promoted to a miracle. It remains a question in the dark, carrying the possibility of surprise in both directions: perhaps it will become a new piece of biology; perhaps it will dissolve into an artifact of the search. Either way, something real gets to have the final word.
A machine can help us notice the whisper.
The world still gets to answer.