Terence Tao's "Mathematics 2.0" vision and AI proof warning

Tao's 'Math 2.0' post argues that mathematics' core institutions - tenure, prizes, journal publication, and the prestige of being first to crack a famous open problem - were all built around an assumption that proofs are scarce and that solving a hard problem signals rare human expertise ^[1]. His framing is blunt about what has changed: 'We will transition from an era of proof scarcity to an era of proof abundance' ^[2]. That shift breaks the field's implicit incentive structure, because problem-solving was never supposed to be mathematics' end goal - it was a proxy for understanding. Tao's sharpest line states this directly: 'Further blind optimization of problem-solving alone is now actively harmful to the long-term health of mathematics' ^[3]. Once AI can generate proofs cheaply and at scale, a field that still rewards raw solve-counts is optimizing for the wrong metric - producing more answers while the thing those answers were supposed to indicate, human insight, goes untested and undeveloped.

Tao coined 'proof indigestion' to describe a scenario where AI-generated theorems pass formal Lean verification faster than humans can explain, teach, or absorb them - he notes that 'immense volumes of machine-generated theorems pass Lean verification, yet even the engineers who initiated the runs cannot explain the underlying mechanisms' ^[4]. He pairs this with a second, related worry specific to open problems: he describes them as 'lighthouses that are less destinations than guides for exploring the surrounding landscape,' arguing that what mathematicians learn while attempting a hard problem - the failed approaches, the new techniques, the surrounding territory mapped along the way - is often worth more than the final answer ^[5]. An AI system that reaches the destination without generating any of that surrounding insight does not just fail to help; in Tao's framing it risks 'sterilizing' the problem's neighborhood, destroying the motivation for years of human exploratory work that would otherwise have continued around it.

Tao's post did not arrive in a vacuum. Days earlier, OpenAI released more than 700 manuscripts (reported elsewhere as 370-plus solved problems) claiming solutions to numerous open mathematical problems, including partial progress on Millennium Prize problems, at an average of roughly three hours of compute per solution ^[7]. The release triggered mathematicians reacting 'with shock and disgust,' including Fields Medalist Hugo Duminil-Copin's 'The shock is immense. I am paralysed,' alongside questions like 'How much beauty have we lost?' ^[8]. The backlash was already building before the release: 25 Fields Medalists, including Tao and Peter Scholze, had co-signed a September 11 declaration titled 'A Severe Misalignment of AI and Mathematics,' warning that AI companies' commercial incentives - speed, benchmark claims, publicity - clash with mathematics' slower, verification-based, attribution-conscious culture ^[9]. The Association for Human Mathematics went further, calling for a boycott and describing the release as 'not a demonstration of scholarship, but a demonstration of power.' Tao's 'Math 2.0' framework reads as the constructive half of that same reaction - less an attack on OpenAI specifically than an attempt to design incentive structures that could survive this kind of event recurring.

Tao's own answer is not to halt AI-assisted mathematics but to redesign what the field rewards. Through SAIR, the nonprofit he co-founded, he has pushed an 'Open Math Model' initiative offering transparent, inspectable model weights as a counterweight to closed frontier systems ^[10]. Alongside that, his Math 2.0 proposals include 'open exposition problems' that reward clear explanation over first-solve priority, living networks of research results, large formal libraries like Mathlib, and ablation studies that test how resilient a theorem's proof actually is - all designed to push the field's incentives back toward understanding rather than solve-counts. Tao frames the obstacle plainly: 'the primary bottleneck to achieving Math 2.0 is not technology... It is imagination' ^[6]. None of this addresses what happens if an AI lab simply has no reason to adopt exposition-rewarding norms the mathematics community builds for itself.

Not every mathematician reads Tao's warning the same way. University of Toronto mathematician Dan Litt took the opposite view of the OpenAI release, calling the wave of AI-generated solutions 'great for mathematics' ^[7]- a direct counterpoint to Scholze's insistence that 'the goal is and always will be the human understanding of mathematics' ^[8]. That split roughly tracks a broader divide in how the field is processing Math 2.0: one camp treats concerns like Tao's as understandable but ultimately a status defense by an established field resistant to automation, while the other treats them as a narrow, specific pedagogical argument - that solving problems instantly denies the next generation of mathematicians the training ground of attempting them. Tao himself has tried to locate the disagreement one level up, framing the field's crisis as being about values rather than raw capability, and calling on mathematicians to respond collectively: 'we have to organise, become activists, get a little political' ^[6]. The unresolved tension is whether incentive redesign, Tao's bet, can actually compete with the speed and publicity incentives driving AI labs toward exactly the kind of release that triggered this debate.