The Lever on Paper
Two mathematicians would rather go slower and neither can see a way to do it. One of them has endorsed a declaration, backed by the International Mathematical Union, that asks governments for public oversight of the AI industry.
In mathematics, Terence Tao said in an interview with New Scientist that circulates on X, there is a lot of good work to be done with the tools that already exist. There is no need on our end to accelerate any further, and the safety risks of going faster are not worth it. Then he named the mechanism: the big tech companies are locked into a prisoner’s dilemma, and this is a very toxic dynamic.
«Nobody wants to go this fast, but everyone feels like they have no choice.»
He added that he does not know how to cool the temperature down.
Timothy Gowers, also a Fields medallist, wrote on his blog on 26 July that he would have preferred these developments to happen at a slower pace, and that he sees no practical way to slow them down.
Tao does not know how to cool the temperature. Gowers sees no practical way. Both describe a constraint that no participant can loosen, and the logic holds: a prisoner’s dilemma is not solved by the people inside it.
Gowers then takes the premise somewhere else. If AI removes the motivation to spend years reaching the level of a research mathematician, the literature could expand enormously while the community that understands it thins out. Resisting that will require giving less priority to some current values, he writes, and he names one: ownership of mathematical results. He narrows the scope. If a formalised solution to an important open problem is obtained with a single prompt, and someone takes on the work of digesting and explaining it, that person should receive the lion’s share of the credit. It is a rule for one case, not a programme for the discipline.
He adds that there is simply no point in hoping that the system of incentives will not be disturbed, and that he is not necessarily too worried if the mechanisms for hiring, funding and recognition are disrupted. A collapse of the publishing system, he writes, he would regard as a beneficial consequence of AI. For him, giving something up is not a loss.
Tao has endorsed the Leiden Declaration on Artificial Intelligence and Mathematics, which has more than 4,400 signatories and is endorsed by the International Mathematical Union. His statement on the declaration’s page says he wholeheartedly endorses its statements and recommendations. Gowers did not sign. His reason, he writes, is not so much that he disagrees with it in any concrete way, but that it makes confident assertions and recommendations he feels somewhat uncertain about.
The declaration is not a general appeal. It has separate sections of recommendations for individual mathematicians; for mathematical organisations and not-for-profit research funders; for policymakers in government and elsewhere; and for commercial AI. The policymaker section has four items: protect the rights of authors, don’t believe the hype, regulate the AI industry, and invest in public computational infrastructure. The third is the one that matters here. It says the signatories stand with others in calling for significantly increased public oversight.
That is the closest thing in the document to the lever the two men say they cannot see: not a request to the participants in the race, but a request to the people who could change the rules they race under. It is also less than a lever. It asks for oversight, not for a slower pace, and it does not say what oversight would change about what defecting pays.
So one of the two has put his name to a request addressed to governments, while the other, who did not, says he sees no practical way to slow things down. Endorsing a request is not slowing anything down, and what the declaration achieves is a separate question that this text does not try to answer. In the excerpt of the interview that circulates on X, Tao does not mention it; the full interview may.
What the declaration shows is narrower than a solution. The instrument exists, it is addressed to the people who could use it, and it carries the discipline’s own endorsement. Whether it points at the thing both men say is missing is the question it leaves open.
The difference between the two is not in the diagnosis. It is in the price. Gowers says what he is willing to lose. Tao, in the excerpt, does not say what he would give up.