The Leiden Declaration is a policy and values document from a cross-disciplinary group convened in Leiden in 2025. It argues that AI is changing mathematical research in ways that go beyond convenience tools. The document is worried about five things in particular: machine-generated proofs that look plausible but are unreliable, loss of attribution when models draw from the mathematical commons without preserving sources, unequal access to AI tooling, hype outrunning evaluation, and research incentives drifting toward the priorities of large AI firms. It pairs those concerns with recommendations to mathematicians, institutions, funders, publishers, and policymakers, including skepticism toward marketing claims and stronger regulation of the AI industry.
Most of the useful reaction landed on one point: mathematics is unusually exposed to a review bottleneck. A proof is only valuable if other experts can trust what it says and understand what was actually shown. People pushed back hard on the lazy version of the rebuttal, which says “human proofs are flawed too.” The sharper claim is that AI changes the scale and shape of the problem. A human author usually leaves a trail of judgment, effort, and intellectual context that helps others decide what deserves attention. A model can produce huge volumes of formally polished arguments, including long brute-force or highly cross-domain proofs, without preserving that same signal. That makes review, trust, and training harder even if some outputs are correct.
The conversation also made clear that attribution is not just prestige accounting. In math, citations help readers separate what is new from what is inherited and trace ideas back to the original sources that make a result intelligible. Several commenters said that if AI strips away that lineage, it does not just hurt careers. It degrades understanding. That connects to a broader fear running through the comments: if AI handles more of the proving, early-career mathematicians may lose the apprenticeship path that teaches them how to judge arguments in the first place. The most grounded pro-AI comments did not really dispute that. They argued that the field should adapt by shifting human work toward problem selection, interpretation, elegance, and explanation, much as formal proof assistants already shift effort toward verification. The skeptical comments mostly objected to the declaration’s politics and tone, not to the underlying governance problem.
If your work depends on expert review, credentialing, or provenance, the core issue is not whether AI can produce outputs. It is whether your institution can still verify quality, assign credit, and train the next generation once AI floods those systems.
Cautiously supportive of the declaration’s core concerns, especially around proof verification, attribution, and researcher training. The main irritation was with wording that sounded political or defensive, but most substantive comments treated the underlying review and governance problems as real.
Key insights
01
AI collapses the proof-review asymmetry
The key change is not that machine proofs can be wrong. Human proofs are wrong too. The change is that mathematics relies on an asymmetry where producing a serious proof usually takes much more effort than checking whether it is worth reviewing. That asymmetry acts as a filter. Generated proofs erase that filter. Reputation and institutional signals already help reviewers decide what deserves attention, and AI makes that triage problem much worse by cheaply producing work that looks substantial before anyone has established that it is.
If you run any expert-review pipeline, expect AI to break the old assumption that submission cost limits junk volume. You need stronger intake filters, provenance requirements, or formal verification steps before human review.
Proof assistants like Lean can check formal proofs, but that does not end the problem. Very long or adversarial proofs can still be hard to interpret, and Lean itself distinguishes between ordinary mistaken proofs and “dishonest” ones that exploit gaps in the system or libraries. Even when every step checks out, readers may still not know whether the proof answers the question they care about, or only a technically adjacent one. Correctness is narrower than understanding.
Treat formal proof as one layer of assurance, not the whole governance stack. You still need human interpretation, model-risk thinking, and standards for explaining what a verified result actually means.
Citation in mathematics does more than allocate prestige. It tells readers which parts are genuinely new, which parts are inherited, and where to go if they want the conceptual path that made the result make sense. That historical chain often carries the understanding, not just the final statement. A source-free synthesis can be valid and still be much less valuable because it severs the route back to the ideas that produced it.
If you deploy AI in research workflows, insist on source-traceability as a product requirement. Outputs without recoverable lineage are harder to trust, teach from, and build on even when they look polished.
The strongest human-capital concern was not nostalgia for hard work. It was that mathematicians learn by doing proofs themselves, then by having those proofs judged by other humans. That process creates both skill and credibility. If AI handles the hard parts too early, the field may lose the apprenticeship system that produces people capable of reviewing frontier work later. Mathematics can survive more automation, but only if it keeps a path for humans to acquire judgment rather than just consume results.
Organizations adopting AI in expert domains should protect novice-to-expert training loops on purpose. If juniors skip the hard parts entirely, you may get short-term output gains and a long-term collapse in evaluator capacity.
The chess analogy breaks at problem selection. Chess has fixed rules and a clear objective, so superhuman play can leave humans enjoying the game without changing what the game is. Mathematics is open-ended. Someone has to decide which concepts are important, which conjectures are worth stating, and which formalizations capture the interesting question. Even optimistic commenters who expect stronger AI still treated that agenda-setting role as the part humans cannot casually surrender.
When comparing AI disruption across fields, separate execution from agenda-setting. In open-ended domains, the leverage may shift from doing the work to choosing what work is worth doing.
Some complaints read like institutional self-defense
A credible criticism is that parts of the declaration blur genuine epistemic risks with protection of current academic structures. Concerns about incentives, informal channels, and attribution can sound like defense of existing prestige systems rather than a direct argument about mathematical knowledge. That does not kill the document, but it does mean some recommendations will be heard as guild politics unless they are tied more tightly to verification and understanding.
If you are writing AI governance for a profession, separate public-interest claims from status-preservation claims. Otherwise people will dismiss the whole package as incumbents protecting their turf.
There is a real danger in treating the difficulty of old workflows as inherently valuable. Many technologies looked like cheating when they arrived, from digital photography to word processors, and later became normal tools that freed people to focus on better work. Applied to math, this suggests the useful question is not whether AI makes proving too easy, but what distinctly human mathematical work remains once proof search becomes cheap.
Do not build policy around preserving friction for its own sake. Build it around the capabilities and judgments you still need humans to develop after the tools change.
One blunt objection was that mandatory disclosure of LLM use creates a competitive penalty for the people who admit it. In a hostile environment, that can push tool use underground instead of making research cleaner. The comment is cynical, but it points to a real mechanism: if disclosure only increases stigma and not trust, many people will hide usage rather than comply honestly.
Make disclosure regimes useful, not merely moralizing. If you want honest reporting, tie disclosure to clear review benefits and norms that distinguish acceptable assistance from deceptive substitution.
leidendeclaration.ai