A researcher chooses a problem, formulates an intuition, steers the research. A model explores a thousand paths, proposes dozens, formalises what holds up. A verifier validates. The result is published. Who found it? This is not a question of vanity: attribution decides careers, funding, and who is believed next time.
We covered this morning a proof obtained in exactly this configuration. Let's take the time to consider the question it raises.
What attribution actually does
People think it rewards. It mainly does three other things, and that is why it matters.
It designates someone accountable. If the result is wrong, someone must answer for it. That is what distinguishes a scientific publication from a rumour, and it is the point we made in our article on proof: a result is established when someone stakes their credibility on it.
It indicates who to talk to. A cited researcher is someone you can question, who can clarify, defend, acknowledge an error. A model can do none of that.
It steers resources. Positions, funding and collaborations follow the signatures. That is where the stakes become concrete.
Why the model is not an author
The answer is more robust than one might think, and it does not rest on a judgement about the nature of machines.
An author, in the scientific sense, is someone who answers for what they sign. They can be questioned, challenged, contradicted. They can retract a paper. They can be wrong and admit it.
A model cannot take responsibility for anything, because there is no one there to take it. This is not a question of merit or contribution: it is a question of structure. Attribution presupposes an entity that persists, that can be held to account, and that loses something if it is wrong.
This is exactly the reasoning we set out in our article on authenticity: what defines the author is not having produced something, it is answering for it.
The more awkward question
There remains a problem this answer does not solve.
If the model did most of the exploratory work, what is left for the researcher? The honest answer is: the choice of problem, the intuition of direction, and the judgement about what matters.
These are real contributions, and they are exactly the ones we identified as resistant. We reported on a test where models largely failed to recover the core idea of a paper from its premises: these systems exploit the known and stumble on what is not.
But we must acknowledge what changes: the part of the work that made the contribution visible is shrinking. A researcher who spent three years on a proof demonstrated their investment through that duration. One who steers a model for three weeks has produced the same result without the evidence of effort. We developed this in our article on merit: the result no longer testifies to the effort.
This situation is not without parallel. An astronomical discovery made with a multi-billion telescope is not attributed to the telescope, nor to its designers, but to the team that chose where to look and knew how to interpret. A particle physics paper can have thousands of signatories, with established conventions on who appears and why. These disciplines have built attribution rules for work where the instrument does most of the measurable labour. They work imperfectly, and they exist.
What seems right
Three principles, and they are more modest than a theory of authorship.
Declare the usage. Not to stigmatise it, but so the reader knows how the result was obtained. That is what the publication we discussed this morning does, and it is good practice.
Attribute to those who answer. The signatories are those who take responsibility, whatever the tool used. That remains the rule.
Describe the division of labour. Saying what the researcher contributed and what the model produced is more useful than a debate over authorship. That is what the big-collaboration disciplines already do.
What to take away
The question in the title has a clear formal answer: the author is the one who answers for the result, and that will never be a model.
It has a less clear answer when it comes to what we admire. We admired a blend of intelligence and endurance, without distinguishing them because they went together. Endurance can now be delegated, and only intelligence is left to credit — that is, something rarer, less visible, and much harder to measure in a job application.
That may be the real task ahead: not deciding whether a machine can be an author, but relearning to recognise a contribution that can no longer be proven by time spent.