Heidelberg Laureate Forum · Published 2026-09-16

Panel Discussion: AI in Mathematical Research | September 15

privateReview: unreviewedConfidence: high
artificial intelligencemathematicsethicsmachine learningresearchagentic-researchbenchmark-gamingHuman-in-the-loopdata-provenanceevaluationskill-degradationopen-weightsexploratory-tooling
Open on YouTube ↗

Summary

Overview

  • Speaker: Tony Feng (Moderator)
  • Channel: Heidelberg Laureate Forum
  • Main topic: AI in Mathematical Research
  • Purpose: To examine different perspectives from mathematical experts regarding how AI tools affect mathematical research, the nature of proofs, ethical considerations, and the future role of human mathematicians. A panel of distinguished mathematicians and researchers discuss the current and future impact of artificial intelligence on mathematical research, exploring ethical concerns, productivity shifts, and the long-term viability of human mathematics alongside AI systems.

Topic Map

Introduction and Panelist Introductions

  • Explanation: Jessica Fintzen introduces the panel and moderator Tony Feng, who then introduces the panelists (Michael Harris, Peter Scholze, Geordie Williamson, and Jacob Tsimerman) and poses icebreaker questions regarding time spent engaging with AI.
  • Key claims:
    • Mathematicians have diverse perspectives on AI in research.
    • Some prominent mathematicians spend zero hours actively engaging with AI tools, while others explore them deeply.
  • Examples:
    • Michael Harris and Peter Scholze reported spending zero hours actively using AI tools in their immediate work.
  • Terminology:
    • Large Language Models
    • LLMs
    • Fields Medalist
  • Why it matters: Establishes the baseline skepticism and differing engagement levels among top-tier mathematicians regarding current AI technology.

AI, Ethics, and the 'Gift Economy' of Mathematics

  • Explanation: Michael Harris discusses ethical concerns surrounding commercial LLMs, data scraping, intellectual property theft, and how mathematics operates on a gift economy.
  • Key claims:
    • Commercial LLMs are built on the theft of intellectual property and exploited labor.
    • Mathematics operates as a gift economy where community and collaboration matter more than profit motives.
  • Examples:
    • Michael Harris citing Ted Chiang on the ethical problems of LLMs.
  • Terminology:
    • gift economy
    • intellectual property
    • silicon valley
  • Why it matters: Highlights the deep philosophical and ethical discomfort many academics feel regarding commercial AI deployment.

AI in Exploration vs. Benchmark Solving

  • Explanation: Geordie Williamson and Peter Scholze discuss the difference between using coding agents and AI for genuine mathematical exploration versus solving narrow benchmark problems as PR stunts.
  • Key claims:
    • Solving millennium problems as benchmarks is often just a PR stunt and unhealthy for the mathematical community.
    • AI is genuinely useful when used as a coding agent for exploratory calculations and spotting subtle patterns (e.g., in number theory).
  • Examples:
    • Geordie Williamson's work with DeepMind on Kazhdan-Lusztig polynomials.
  • Terminology:
    • millennium problems
    • Navier-Stokes
    • formalization
  • Why it matters: Distinguishes hype-driven AI benchmarks from practical, exploratory applications that assist researchers.

The Future of Human Mathematics and Oversight

  • Explanation: Jacob Tsimerman and the panel debate AI safety, the risk of extinction-level over-reliance, and whether human mathematicians risk becoming obsolete if AI surpasses human capabilities.
  • Key claims:
    • AI safety is a paramount issue with no shortage of headlines.
    • Human mathematics and culture hold inherent value independent of raw computational or theorem-proving efficiency.
  • Examples:
    • Discussion on whether humanity will lose its role as the primary expositor of mathematical truth.
  • Terminology:
    • AI safety
    • superhuman mathematicians
    • oversight
  • Why it matters: Addresses existential anxiety within the academic community about the long-term utility and status of human intellect.

Key Points

AI as an Exploratory Tool

  • Explanation: AI models and coding agents help mathematicians run experiments and spot patterns faster than manual computation.
  • Evidence: Geordie Williamson's collaboration with DeepMind.
  • Practical implication: Mathematicians can leverage coding agents to explore complex combinatorial or algebraic structures.

Ethical and Economic Concerns

  • Explanation: Commercial AI development relies on data scraping and labor extraction that conflict with academic norms.
  • Evidence: Michael Harris's critique of commercial LLM business models.
  • Practical implication: Researchers must weigh the ethical implications of using proprietary AI systems trained on scraped academic work.

The Value of Human Mathematical Culture

  • Explanation: Mathematics is not just about proving theorems; it is a human cultural endeavor rooted in community and shared understanding.
  • Evidence: Panel consensus on the importance of human engagement and dialogue.
  • Practical implication: Even if AI proves theorems efficiently, human-led mathematical practice and education remain vital.

Frameworks, Models & Processes

Full AI vs. No AI Experimentation

  • How it works: A pedagogical method of alternating between maximum AI assistance and zero AI assistance to calibrate research skills.
  • Components:
    • Full AI immersion
    • AI-free grounding
    • Skill calibration
  • When to use: When training students or evaluating personal dependence on automated tools.

Examples & Case Studies

Geordie Williamson received a tweet regarding a counterexample to the Jucovič conjecture right before an ICM event.

  • Illustrates: How rapidly AI-generated or assisted mathematical observations circulate online.
  • Lesson: The mathematical community experiences intense speed and disruption from online information flows.

Actionable Takeaways

  • Immediate:
    • Experiment with coding agents for exploratory mathematical tasks.
    • Be mindful of the ethical and copyright implications of commercial AI tools.
  • Strategic:
    • Maintain a balance between leveraging AI acceleration and preserving foundational human mathematical skills.
    • Engage in cross-disciplinary discussions on AI safety and governance.
  • Questions to investigate:
    • How will AI change the publication and peer-review process in mathematics?
    • Can open-source AI models align better with academic gift economies?

Claims Worth Verifying

  • AI models will be able to explain mathematics better than any human mathematician by April. (Speculative prediction)
  • The International Mathematical Union was the first institution to endorse the Leiden Declaration. (Historical fact)

Notable Quotes

"Many people feel that LLMs are a fundamentally unethical technology because they are built on the theft of intellectual property." "Mathematics is one of those perspectives to gain theoretical foundations and into the subject of AI."

Compressed Summary

  • Mathematicians exhibit a wide spectrum of engagement with AI, from total non-use to deep collaboration.
  • Ethical concerns center on data scraping, intellectual property theft, and commercial surveillance capitalism.
  • AI serves best as an exploratory assistant for computational and pattern-recognition tasks.
  • Preserving human mathematical culture and critical oversight remains essential.
  • Keywords: artificial intelligence, mathematics, ethics, machine learning, research
  • Core insight: While AI offers powerful exploratory tools for mathematical research, the community must navigate severe ethical challenges and preserve the human cultural foundations of mathematical inquiry.

Core insights

6
Mental Modelmedium noveltymoderate evidence

AI's defensible contribution to deep expert work is as a coding agent for exploratory computation and subtle pattern-spotting (e.g., Kazhdan-Lusztig polynomials with DeepMind), not as a solver of headline benchmark problems.

Why it matters

Suggests capability investments should target open-ended exploration loops (hypothesis generation, exploratory computation, anomaly detection) rather than fixed-problem benchmark performance, which the panel treats as decoupled from genuine research value.

Generalization

For agent design, separate 'benchmark-targeted' capability from 'exploration-support' capability; they may require different tool surfaces, harnesses, and evaluations, and optimizing one can misrepresent the other.

AI is genuinely useful when used as a coding agent for exploratory calculations and spotting subtle patterns (e.g., in number theory).
Open source video
Geordie Williamson's work with DeepMind on Kazhdan-Lusztig polynomials.
Open source video
Solving millennium problems as benchmarks is often just a PR stunt and unhealthy for the mathematical community.
Open source video
Practicemedium noveltyweak evidence

A deliberate 'full AI vs. no AI' alternation regime is proposed as a way to calibrate research skills and detect dependence on automated tooling.

Why it matters

Provides a concrete, testable harness for measuring skill preservation and operator dependence, rather than assuming AI assistance is strictly additive. Applicable to engineering teams evaluating copilots and agents.

Generalization

Any tool-assisted workflow can be evaluated via periodic tool-off control periods to measure degradation of unaided competence and to attribute output quality correctly.

A pedagogical method of alternating between maximum AI assistance and zero AI assistance to calibrate research skills.
Open source video
Full AI immersion
Open source video
AI-free grounding
Open source video
Skill calibration
Open source video
Tradeoffmedium noveltymoderate evidence

Commercial LLM development is framed as resting on IP theft and exploited labor, in direct conflict with mathematics' 'gift economy' norms, making provenance an architectural constraint rather than a policy afterthought.

Why it matters

For systems deployed in academia or regulated environments, model provenance and training-data lineage become selection criteria that trade off against raw capability; open-weight alternatives are proposed as the alignment path.

Generalization

Provenance-aware model selection is an architectural variable: teams can trade peak capability for license/data lineage compliance, and should treat data lineage as part of the model dependency surface.

Commercial LLMs are built on the theft of intellectual property and exploited labor.
Open source video
Mathematics operates as a gift economy where community and collaboration matter more than profit motives.
Open source video
Can open-source AI models align better with academic gift economies?
Open source video
Mental Modelmedium noveltymoderate evidence

Human oversight in expert domains is defended on grounds of cultural/accountability value, explicitly decoupled from raw computational or theorem-proving efficiency.

Why it matters

Reframes human-in-the-loop placement: humans are retained not as the fastest verifier but as the accountable expositor and meaning-maker, which changes interface and audit design goals.

Generalization

Design oversight interfaces for accountability and interpretability even when the machine outperforms the human on the primary task metric; the value function includes legitimacy, not just accuracy or speed.

Human mathematics and culture hold inherent value independent of raw computational or theorem-proving efficiency.
Open source video
Discussion on whether humanity will lose its role as the primary expositor of mathematical truth.
Open source video
Empirical Resultmedium noveltymoderate evidence

Elite practitioners report zero hours of active AI use, indicating a frontier gap: current tooling is not yet adopted inside the deepest expert reasoning loops despite enthusiasm elsewhere.

Why it matters

Challenges the assumption of universal expert adoption; suggests the binding constraint is workflow fit at the frontier of expert reasoning rather than general model capability.

Generalization

Adoption telemetry from top practitioners is a leading indicator of whether tooling addresses real workflow bottlenecks or only peripheral tasks.

Michael Harris and Peter Scholze reported spending zero hours actively using AI tools in their immediate work.
Open source video
Some prominent mathematicians spend zero hours actively engaging with AI tools, while others explore them deeply.
Open source video
Empirical Resultmedium noveltymoderate evidence

AI-assisted mathematical observations now circulate online fast enough to disrupt community processes (e.g., a counterexample to the Jucovič conjecture arriving via tweet before an ICM event).

Why it matters

Indicates the verification/review pipeline, not generation, becomes the bottleneck as machine-assisted output velocity rises; review latency and provenance tracking become the scarce resources.

Generalization

When generation cost approaches zero, the limiting subsystem shifts to validation, triage, and attribution; systems should be designed around verification throughput.

How rapidly AI-generated or assisted mathematical observations circulate online.
Open source video
Geordie Williamson received a tweet regarding a counterexample to the Jucovič conjecture right before an ICM event.
Open source video

Deep dives

5

Discovery-oriented evaluation for exploratory agents

Research question

If benchmark performance is decoupled from research utility, what evaluation protocol measures an AI coding agent's exploratory utility in expert work?

Why

Determines whether capability claims reflect assistant value or benchmark optimization, and guides tool and harness design for open-ended exploration.

AI is genuinely useful when used as a coding agent for exploratory calculations and spotting subtle patterns (e.g., in number theory).
Open source video
Solving millennium problems as benchmarks is often just a PR stunt and unhealthy for the mathematical community.
Open source video
Source video

Verification throughput as the limiting subsystem

Research question

When AI-assisted result generation velocity exceeds expert review capacity, what system design maximizes verification throughput and attribution?

Why

Reframes research infrastructure around validation, triage, and provenance rather than generation, addressing the peer-review bottleneck.

How rapidly AI-generated or assisted mathematical observations circulate online.
Open source video
Geordie Williamson received a tweet regarding a counterexample to the Jucovič conjecture right before an ICM event.
Open source video
Source video

Provenance-aware model selection as architecture

Research question

How should model provenance and training-data lineage be represented as gating constraints in research AI deployments?

Why

Makes compliance and ethical norms part of the dependency surface, not a policy afterthought, affecting model choice and fallback strategies.

Commercial LLMs are built on the theft of intellectual property and exploited labor.
Open source video
Mathematics operates as a gift economy where community and collaboration matter more than profit motives.
Open source video
Source video

Tool-off calibration regimes for skill preservation

Research question

Can alternating maximum AI assistance with zero AI assistance reliably measure skill preservation and dependence in expert workflows?

Why

Provides a testable harness for skill degradation and attribution, applicable to copilots and agents in engineering teams.

A pedagogical method of alternating between maximum AI assistance and zero AI assistance to calibrate research skills.
Open source video
Some prominent mathematicians spend zero hours actively engaging with AI tools, while others explore them deeply.
Open source video
Source video

Human oversight as accountable expositor

Research question

How should oversight interfaces be designed when the human is retained for accountability and interpretation rather than superior task performance?

Why

Changes UI and audit design goals toward reasoning exposure and attribution rather than aggregate confidence.

Human mathematics and culture hold inherent value independent of raw computational or theorem-proving efficiency.
Open source video
Discussion on whether humanity will lose its role as the primary expositor of mathematical truth.
Open source video
Source video

Article ideas

4

The Benchmark Mirage: Why Headline Problem-Solving Is Not Research Utility

AI capability claims should be evaluated by exploratory utility in expert workflows, because benchmark-targeted wins can be decoupled from genuine research value and distort community incentives.

Angle

Contrast PR-stunt benchmark solving with coding-agent-driven pattern discovery, then propose discovery-oriented evaluation criteria.

Source video

Review at the Speed of Generation: Designing for the Verification Bottleneck

As AI-assisted claims circulate faster than peer review, research infrastructure must shift from generation to triage, provenance stamping, and verification throughput.

Angle

Use the counterexample-via-tweet anecdote to redesign pre-review and attribution pipelines.

Source video

Provenance Is a Dependency: Model Selection in the Academic Gift Economy

Training-data lineage and license compatibility are architectural selection criteria for research deployments; open-weight models may align with academic norms but must be evaluated as a capability-provenance tradeoff rather than a moral binary.

Angle

Model procurement as supply-chain design, with provenance metadata and fallback routing.

Source video

Humans as Accountable Expositors, Not Fallback Verifiers

Human oversight in expert domains should be justified by accountability and interpretive legitimacy, not by outperforming models, so oversight UIs must expose reasoning and attribution.

Angle

Reframe human-in-the-loop design when the machine exceeds the human on the primary metric.

Source video

Project ideas

4

ExplorationBench

beyond-evals

Agents optimized for fixed-answer benchmarks will not rank equally on open-ended exploratory tasks; benchmark score will have weak correlation with expert-rated exploratory utility.

Proof of concept

Build a harness of exploratory tasks (pattern detection, counterexample search, exploratory computation) with expert grading, then run benchmark-tuned agents against coding-agent configurations.

Measurement

Rank correlation between benchmark scores and expert exploratory utility; false-positive discovery rate; time-to-first-insight.

Source video

Verification Triage Pipeline

gatehouse

A provenance-stamped pre-review triage system will reduce time-to-verification and improve attribution for rapidly circulating AI-assisted mathematical claims compared with unstructured channels.

Proof of concept

Ingest AI-assisted claims from public posts, attach provenance and reproducibility checklists, route to expert triage queues, and measure review latency.

Measurement

Median time-to-verification; provenance completeness score; expert agreement and errors caught.

Source video

Provenance-Gated Model Router

gatehouse

A router that treats training-data lineage and license as gating metadata can enforce provenance constraints for academic deployments with acceptable exploratory-task performance via open-weight fallbacks.

Proof of concept

Tag models with provenance and license metadata, route tasks according to policy, fallback to open-weight models, and log lineage decisions.

Measurement

Policy compliance audit pass rate; exploratory task success; latency and cost penalty versus unconstrained routing.

Source video

Tool-Off Calibration Harness

movement-lab

Periodic AI-free control periods will detect measurable skill degradation and attribution errors in tool-assisted expert workflows.

Proof of concept

Workflow plugin that schedules alternating AI-on and AI-off sessions, logs artifacts, and collects blind expert ratings of unaided versus assisted output.

Measurement

Delta in unaided task quality and time across control periods; accuracy of self-reported AI contribution; drift in expert rating.

Source video

Architectural implications

4

The panel positions AI as an instrument inside a human's exploratory loop rather than an autonomous theorem-prover.

Before

Agent framed as an autonomous problem-solver that produces finished results.

After

Coding/exploration agent framed as a tool that accelerates the human's hypothesis-testing cycle.

Consequence

Tool surfaces, context management, and evaluation should optimize for fast iterative exploration and interpretability of intermediate findings, not end-to-end autonomy.

Source video

Benchmark resolution is characterized as PR-motivated and disconnected from research utility.

Before

Benchmark score used as the primary capability proxy.

After

Exploratory-utility and pattern-discovery evaluations treated as the meaningful signal.

Consequence

Evaluation harnesses must include open-ended discovery tasks, not only fixed-answer benchmarks, to avoid optimizing the wrong objective.

Source video

Data provenance is treated as a first-class ethical and possibly adoption constraint.

Before

Model choice driven primarily by capability and cost.

After

Model choice also gated by training-data lineage and license compatibility.

Consequence

Teams need provenance metadata on model dependencies and fallback to aligned alternatives, adding a supply-chain dimension to model selection.

Source video

The human's role is defended on accountability/cultural grounds independent of performance.

Before

Human retained only where they outperform the model.

After

Human retained as accountable expositor even where the model outperforms.

Consequence

Oversight UIs should expose reasoning and attribution for audit purposes, not merely aggregate confidence.

Source video

Tradeoffs and failure modes

4

AI assistance vs. skill degradation

Benefit

Accelerated exploration and pattern discovery via coding agents.

Cost or risk

Erosion of unaided research skills; mitigations require explicit tool-off periods.

A pedagogical method of alternating between maximum AI assistance and zero AI assistance to calibrate research skills.
Open source video
Source video

Benchmark optimization vs. genuine utility

Benefit

Headline benchmark wins generate visibility and funding narrative.

Cost or risk

Incentive distortion and community harm when benchmark solving is decoupled from real capability.

Solving millennium problems as benchmarks is often just a PR stunt and unhealthy for the mathematical community.
Open source video
Source video

Capability from scraped data vs. provenance ethics

Benefit

Commercial LLMs deliver capability built from large-scale data.

Cost or risk

IP appropriation and exploited labor conflict with academic gift-economy norms, creating adoption friction.

Commercial LLMs are built on the theft of intellectual property and exploited labor.
Open source video
Source video

Generation velocity vs. verification latency

Benefit

AI-assisted results propagate rapidly through online channels.

Cost or risk

Peer review and community verification cannot keep pace with circulating claims.

How rapidly AI-generated or assisted mathematical observations circulate online.
Open source video
Source video

Open questions

4

How will AI change the publication and peer-review process in mathematics?

Why unresolved

Raised as an explicit research question with no proposed mechanism in the summary.

Research direction

Prototype machine-assisted triage and provenance-stamped pre-review pipelines that scale with generation velocity.

Source video

Can open-source AI models align better with academic gift economies?

Why unresolved

Framed as an open possibility; the summary offers no evidence on whether open-weight models change ethical or economic outcomes.

Research direction

Compare provenance, licensing, and community-governance properties of open-weight versus commercial models for research deployment.

Source video

How should exploratory utility be measured if benchmark performance is a poor proxy?

Why unresolved

The panel rejects benchmark solving as meaningful without supplying an alternative measurement.

Research direction

Design discovery-oriented evaluations (novel pattern detection, counterexample generation) and validate them against expert judgment.

Source video

Will humans lose the role of primary expositor of mathematical truth?

Why unresolved

Posed as an unresolved debate about human status under superhuman theorem-proving.

Research direction

Study whether human-authored explanations retain verifiable epistemic value when model-generated proofs exceed human performance.

Source video

Key claims

8
opinionVerification needed

Commercial LLMs are built on theft of intellectual property and exploited labor.

Evidence

Commercial LLMs are built on the theft of intellectual property and exploited labor.

Question

What specific datasets, labor practices, or licensing violations substantiate this claim?

Source video
factualVerification needed

Mathematics operates as a gift economy in which community and collaboration outweigh profit motives.

Evidence

Mathematics operates as a gift economy where community and collaboration matter more than profit motives.

Question

What published evidence supports characterizing academic mathematics as a gift economy?

Source video
causalVerification needed

Solving millennium problems as benchmarks is often a PR stunt and unhealthy for the mathematical community.

Evidence

Solving millennium problems as benchmarks is often just a PR stunt and unhealthy for the mathematical community.

Question

Is there evidence linking benchmark-targeted announcements to measurable community harms?

Source video
comparativeVerification needed

AI is genuinely useful as a coding agent for exploratory calculations and spotting subtle patterns.

Evidence

AI is genuinely useful when used as a coding agent for exploratory calculations and spotting subtle patterns (e.g., in number theory).

Question

What concrete discoveries or time savings were attributable to coding-agent use in the cited Kazhdan-Lusztig collaboration?

Source video
opinionVerification not requested

Human mathematics and culture have inherent value independent of computational or theorem-proving efficiency.

Evidence

Human mathematics and culture hold inherent value independent of raw computational or theorem-proving efficiency.

Source video
predictionVerification needed

AI models will be able to explain mathematics better than any human mathematician by April.

Evidence

AI models will be able to explain mathematics better than any human mathematician by April.

Question

By which April, and by what evaluation standard would 'explaining better than any human' be judged?

Source video
factualVerification needed

The International Mathematical Union was the first institution to endorse the Leiden Declaration.

Evidence

The International Mathematical Union was the first institution to endorse the Leiden Declaration.

Question

Was the IMU chronologically the first institutional endorser of the Leiden Declaration?

Source video
factualVerification not requested

Michael Harris and Peter Scholze spend zero hours actively using AI tools in their current work.

Evidence

Michael Harris and Peter Scholze reported spending zero hours actively using AI tools in their immediate work.

Source video

Connections

5