Wiki › Thinking next to AI

AI, Mathematics and Research Incentives

Research rewards after cheaper AI proof production: worthwhile questions, understanding, exposition, technical apprenticeship and new directions.

Core idea

When generating mathematical proofs becomes cheaper, producing a new theorem becomes a weaker proxy for a researcher's understanding and contribution. Frank Calegari argues for rewarding the choice of worthwhile questions, structural insight, exposition, and new research directions rather than simply accelerating the existing race to publish results.

The source is a November 1, 2026 guest essay on Terence Tao's blog, not an essay authored by Tao. It advocates institutional changes; it does not establish that all mathematical proof production has become automated.

Proof, verification and understanding

These are different achievements:

• Proof production: obtaining an argument, potentially entirely through AI.
• Verification: establishing that an argument is correct; formal verification in Lean is one possible safeguard, not something every cited AI result has already undergone.
• Understanding: explaining why a result matters, how it connects to other mathematics, and what it makes possible next.
• Research direction: identifying worthwhile unexplored terrain rather than only finishing inherited conjectures.

For Calegari, choosing an important problem, obtaining an AI-generated proof, understanding it thoroughly, and using it to develop new questions can be a substantial human contribution even without adding an ingredient to the original proof. Conversely, priority and novelty alone are not sufficient evidence of understanding.

This distinction extends Cognitive Surrender from individual tool use to academic incentives: an institution can reward plausible output while failing to cultivate the expertise that could judge it.

Proposed institutional changes

1. Evaluate all work as potentially AI-assisted. Ask what ideas, explanation, or research direction a paper contributes, not whether the author performed a machine-reproducible feat unaided. This is an evaluation standard, not a requirement to use AI.

2. Change hiring. Use direct mathematical conversations for graduate students, postdocs and faculty rather than relying mainly on publication prestige. Distinguish deep knowledge from fluent presentation; Calegari values the former more highly.

3. Raise journal standards. Reward clear, useful exposition and substantive ideas; produce fewer, better papers. He floats temporarily abandoning journals that cannot adapt, not abandoning the recording of mathematical results.

4. Broaden graduate education while retaining technical depth. Students need a view beyond their specialty and demonstrable ability on some class of problems, including deep mathematics without AI.

5. Improve seminars. Explain the larger questions, connections, obstacles and promising avenues before the latest technical novelty.

6. Take collective ownership and responsibility. Situate machine-produced results within the accumulated human mathematical tradition, rather than assigning all intellectual credit to a company or its promoters. This is a cultural-credit argument, not a legal ownership claim.

Why solved problems need not end a field

Calegari uses the classification of finite simple groups as an analogy: completing a famous goal can cool professional interest without exhausting a subject's deeper questions. AI could similarly lower barriers to ambitious cross-field programs. His arithmetic/hyperbolic-geometry example illustrates research he would like to explore, not a reported solved problem or demonstrated AI discovery.

The emotional loss of familiar professional practices can be real even when the discipline still has abundant questions. This links to AI Job Grief without implying that optimism settles employment or funding outcomes.

Education and connections

The proposal complements AI Education With and Without Machines: both preserve unaided foundations while opening more ambitious assisted work. It also adds an institutional layer to AI and Mental Effort: personal commitment to understanding is harder to sustain if hiring and publication reward only rapid output.

The parallels to AI-Native Product Work are wiki synthesis: when implementation becomes cheaper, selection, taste and explanation may become more important. Mathematical truth, research credit and product value remain different kinds of evaluation.

Evidence limits and open questions

• The essay reports an October 6 OpenAI release of over 700 research papers and criticizes writing, attribution and some results. This ingest preserves those as the author's claims; it does not independently validate the release, the papers or their correctness.
• Proposals about journals, interviews and education are normative, not measured evidence that these reforms improve outcomes.
• How can committees assess understanding without rewarding fluency, prestige, access to compute, or personal networks instead?
• Who will fund checking, attribution repair, synthesis and exposition when proof output grows faster than expert attention?
• How should technical apprenticeship and broad exploration be balanced as tool capabilities change?