Simons Foundation · Published 2026-08-13

ICM 2026 Public Lectures - Terence Tao

privateReview: unreviewedConfidence: high
artificial intelligencemathematical foundationsproof generationGoodhart's Lawcanonicalizationpipeline-decompositionproof-indigestiongoodharts-lawreward-hackingverificationprovenancethroughput-bottleneck
Open on YouTube ↗

Summary

Overview

  • Speaker: Terence Tao
  • Channel: Simons Foundation
  • Main topic: The impact of artificial intelligence on the practice, foundations, values, and community of mathematics
  • Purpose: To provide a structured framework and metamathematical analysis of how the mathematical community should respond to modern AI technologies, moving beyond hype to address values, workflows, and infrastructure. Terence Tao explores how the advent of modern artificial intelligence and automated proof assistants (like Lean, Coq, and HOL) are transforming mathematical research. Drawing a historical parallel to the foundational crisis of the early 20th century (Russell's paradox, Gödel's incompleteness theorems), Tao argues that AI forces mathematicians to examine not just technical workflows, but the fundamental goals and values of the mathematical profession. He deconstructs the problem-solving pipeline into five distinct stages—open problems, proof generation, proof verification, proof exposition, and proof canonicalization—and demonstrates how AI accelerates the first two stages while creating severe challenges ("proof indigestion") for the later stages of peer review, human understanding, and textbook canonicalization.

Topic Map

Historical Prologue: The Crisis in Foundations

  • Explanation: Drawing a historical parallel between the early 20th-century crisis in mathematical foundations (naïve set theory inconsistencies, Gödel incompleteness) and today's crisis in mathematical values and practices brought by AI.
  • Key claims:
    • For centuries, mathematics operated on naive foundations delegated to philosophers.
    • Discoveries like Russell's paradox and Gödel's incompleteness theorems forced mathematicians to critically re-examine implicit assumptions.
    • We are entering a similarly turbulent period in the foundations of mathematical values and practices.
  • Examples:
    • Russell's paradox (1901)
    • Gödel's incompleteness theorems (1931)
  • Terminology:
    • naive foundations
    • Russell's paradox
    • Gödel incompleteness theorems
    • crisis in foundations
  • Why it matters: Shows that mathematical communities have successfully navigated foundational crises in the past by re-examining and codifying their assumptions.

The AI Capability Conjecture and Working Hypothesis

  • Explanation: Formulating the community response question and conditionalizing on the AI Capability Conjecture to analyze its consequences without getting bogged down in truth-value debates.
  • Key claims:
    • The Community Response Question asks how the mathematical community should respond to modern AI technologies and their claimed capabilities.
    • The AI Capability Conjecture states that AI tools will soon accomplish research-level mathematical tasks in some fields with non-trivial success rates.
    • A conditional working hypothesis assumes AI tools will soon become capable of performing a reasonable fraction of research-level tasks with reasonable success, quality, supervision, and cost.
  • Examples:
    • First Proof challenge (grassroots independent assessment of AI models on novel research-level problems)
  • Terminology:
    • AI Capability Conjecture
    • Community Response Question
    • Working Hypothesis
    • First Proof challenge
  • Why it matters: Establishes a rigorous, conditional framework to analyze the downstream impacts of AI regardless of whether specific AI capability timelines are true or false.

Goodhart's Law and Goal Divergence

  • Explanation: Applying Goodhart's Law to mathematical research goals, demonstrating how excessive AI optimization for specific metrics (like solving unsolved problems or generating proofs) can cause largely aligned goals to diverge.
  • Key claims:
    • When a measure becomes a target, it ceases to be a good measure (Goodhart's law, 1975).
    • AI tools and companies are particularly vulnerable to Goodhart's law due to their ungrounded nature and financial incentives.
    • Excessive AI optimization causes goals like solving problems, building theory, training students, and creating aesthetic works to diverge.
  • Examples:
    • Optimizing solely for proof generation leads to large numbers of incorrect solutions to major open problems (e.g., Riemann hypothesis).
  • Terminology:
    • Goodhart's law
    • goal divergence
    • proxy goals
    • proof generation
  • Why it matters: Highlights the danger of unconstrained metric optimization in AI-assisted mathematical research.

The Five Stages of Proof Processing and Proof Indigestion

  • Explanation: Deconstructing the problem-solving workflow into five stages: open problems, unverified solutions (proof generation), verified solutions (proof verification), well-written solutions (proof exposition), and definitive solutions (proof canonicalization).
  • Key claims:
    • Problem solving involves five stages: open problems -> proof generation -> proof verification -> proof exposition -> proof canonicalization.
    • AI is currently excelling at proof generation and verification, but is very weak at proof exposition and canonicalization.
    • Proof indigestion occurs when AI generates abundant unverified or unreadable proofs that overwhelm traditional peer review and human comprehension systems.
  • Examples:
    • Erdős problems containing dozens of AI-generated proof submissions that human experts have not volunteered to verify.
    • A 1991 paper of Bourgain annotated by a younger Terence Tao illustrating the friction required to understand deep mathematics.
  • Terminology:
    • proof generation
    • proof verification
    • proof exposition
    • proof canonicalization
    • proof indigestion
    • impedance mismatch
    • natural friction
  • Why it matters: Identifies where AI succeeds and where human bottlenecks and community infrastructure will break down.

Recommendations and the Leiden Declaration

  • Explanation: Outlining practical recommendations for individual mathematicians and the community, referencing the Leiden Declaration on Artificial Intelligence and Mathematics.
  • Key claims:
    • Mathematicians should transparently disclose tool use and maintain responsibility for correctness.
    • The community must decrease emphasis on 'first to solve' and increase emphasis on proof digestion: exposition, publication, and canonicalization.
    • New workflows, infrastructures, and community acceptance filters are needed to navigate proof abundance.
  • Examples:
    • Leiden Declaration on Artificial Intelligence and Mathematics
    • UNESCO Recommendation on Open Science
  • Terminology:
    • Leiden Declaration
    • tool use disclosure
    • proof digestion
    • canonicalization
    • proof abundance
  • Why it matters: Provides concrete action items for the mathematical community to adapt responsibly to the AI era.

Key Points

The Community Response Question

  • Explanation: How the mathematical community should respond to the advent of modern AI technologies and their capabilities in mathematical tasks.
  • Evidence: Comparative analysis of historical foundations and current AI benchmarking.
  • Practical implication: Requires active debate and policy formation by the mathematical community rather than passive acceptance.

Goodhart's Law in Mathematics

  • Explanation: When proof generation becomes the sole optimized target, other vital goals of mathematics (exposition, theory building, education) are neglected.
  • Evidence: Economic and machine learning principles applied to AI optimization incentives.
  • Practical implication: Mathematicians must explicitly define their goals and rubrics rather than letting AI companies dictate them.

Proof Indigestion vs. Proof Scarcity

  • Explanation: Mathematics is transitioning from an era of proof scarcity (where proofs were rare and carefully studied) to proof abundance (where AI generates massive volumes of proofs).
  • Evidence: Emergence of AI-generated proofs on platforms like Erdos problems without human verifiers.
  • Practical implication: Peer review and publishing infrastructure will break unless reformed to handle automated proof output.

The Importance of Natural Friction

  • Explanation: Difficult parts of mathematical proofs require natural friction that forces readers to slow down, engage, and truly understand the core ideas.
  • Evidence: Tao's personal annotations on Jean Bourgain's 1991 paper.
  • Practical implication: Overly polished or opaque AI-generated proofs that remove all friction hinder deep human learning and canonicalization.

The Leiden Declaration Recommendations

  • Explanation: A set of seven recommendations for mathematicians, including tool disclosure, supporting reviewing needs, and retaining responsibility for correctness.
  • Evidence: Leiden Declaration draft principles developed by mathematicians in 2024–2026.
  • Practical implication: Establishes ethical and operational standards for AI usage in mathematical research and publishing.

Frameworks, Models & Processes

The Five-Stage Proof Processing Pipeline

  • How it works: Deconstructs mathematical problem-solving from open problems to accepted canonical theory through successive refinement and community acceptance stages.
  • Components:
    • Open problems
    • Unverified solutions (proof generation)
    • Verified solutions (proof verification)
    • Well-written solutions (proof exposition)
    • Accepted / Definitive solutions (proof canonicalization)
  • When to use: When evaluating where AI tools are effective and where human/community bottlenecks occur in mathematical research.

Examples & Case Studies

Terence Tao annotating a 1991 paper by Jean Bourgain as a graduate student.

  • Illustrates: The educational value of natural friction and struggle in reading and digesting difficult mathematical proofs.
  • Lesson: AI-generated proofs that eliminate all difficulty can deprive students and researchers of deep learning experiences.

Websites dedicated to mathematical problems (like erdosproblems.com) receiving dozens of AI-generated proof submissions.

  • Illustrates: Proof indigestion and impedance mismatch, where proof generation outpaces human verification and review capacity.
  • Lesson: Generating proofs is becoming easy, but verifying, reviewing, and digesting them is becoming the primary bottleneck.

Actionable Takeaways

  • Immediate:
    • Transparently disclose AI tool use in mathematical writing and research.
    • Do not rely solely on AI-generated proof correctness without human verification and exposition.
    • Read and engage with the Leiden Declaration on Artificial Intelligence and Mathematics.
  • Strategic:
    • Shift professional emphasis from 'being first to solve a problem' to proof digestion: exposition, publication, and canonicalization.
    • Develop new workflows and community filters for peer review to handle proof abundance.
    • Protect foundational mathematical education and training from premature AI automation.
  • Questions to investigate:
    • How will mathematical journals adapt their submission and refereeing guidelines for AI-assisted papers?
    • Can formal proof assistants (Lean, Coq) bridge the gap between machine verification and human canonical understanding?
    • How do we maintain community acceptance and canonicalization standards in an era of proof abundance?

Claims Worth Verifying

  • AI models can solve research-level mathematical problems at publication-level quality in controlled tests. (empirical)
  • Websites like erdosproblems.com contain dozens of AI-generated proof submissions awaiting human verification. (factual)

Notable Quotes

"For centuries, mathematics successfully operated using 'naive' foundations, with some fundamental questions largely delegated to philosophers." "When a measure becomes a target, it ceases to be a good measure." "We are not trying to meet some abstract production quota of definitions, theorems and proofs. The measure of our success is whether what we do enables people to understand and think more clearly and effectively about math."

Compressed Summary

  • AI accelerates proof generation and verification but struggles with exposition and canonicalization.
  • Goodhart's law warns that excessive AI optimization of single metrics causes mathematical goals to diverge.
  • The community faces 'proof indigestion' as AI-generated proofs overwhelm human peer review and verification capacity.
  • Proof canonicalization and digestion remain essential human-driven stages for building cohesive mathematical theories.
  • Keywords: artificial intelligence, mathematical foundations, proof generation, Goodhart's law, canonicalization
  • Core insight: While AI revolutionizes proof generation, the mathematical community must pivot its primary valuation from solving problems to proof digestion, exposition, and canonicalization to prevent proof indigestion.

Core insights

6
Architecturehigh noveltystrong evidence

Decomposing an end-to-end task into five distinct stages (open problems, proof generation, proof verification, proof exposition, proof canonicalization) reveals that AI capability is highly asymmetric across the pipeline: models are strong at generation and verification but weak at exposition and canonicalization. The bottleneck therefore migrates downstream rather than disappearing.

Why it matters

If you treat a capability as monolithic you will mis-size the harness. Decomposing a workflow into stages and measuring per-stage capability tells you where the real constraint is and where to invest engineering effort (in this case, the human-facing digestion stages, not the generation stage).

Generalization

Any agentic pipeline where generation is cheap and machine-checkable but consumption requires human comprehension will develop a downstream digest/canonicalize bottleneck. Architectures should budget per-stage and instrument per-stage rather than assuming uniform capability.

Problem solving involves five stages: open problems -> proof generation -> proof verification -> proof exposition -> proof canonicalization.
Open source video
AI is currently excelling at proof generation and verification, but is very weak at proof exposition and canonicalization.
Open source video
Failure Modehigh noveltystrong evidence

Proof indigestion: when generation throughput massively exceeds the community's verification and comprehension throughput, abundant unverified or unreadable outputs overwhelm the review system. The failure is not incorrect output but an impedance mismatch between a fast producer and a slow consumer.

Why it matters

Production agent systems that emit artifacts faster than downstream humans or checks can absorb them will silently accumulate unverified work, eroding trust and correctness guarantees. The relevant metric is producer/consumer rate matching, not raw output volume.

Generalization

Applies to code generation, test generation, document drafting, and any agent whose output requires human review: without backpressure or an acceptance filter, output rate becomes a liability.

Proof indigestion occurs when AI generates abundant unverified or unreadable proofs that overwhelm traditional peer review and human comprehension systems.
Open source video
Erdős problems containing dozens of AI-generated proof submissions that human experts have not volunteered to verify.
Open source video
Mental Modelmedium noveltymoderate evidence

Goodhart's law is applied structurally: when AI is optimized against proxy metrics (solving problems, generating proofs) the previously aligned goals (building theory, training students, aesthetic worth) diverge. Ungrounded AI systems plus financial incentives make this drift especially likely.

Why it matters

Optimizing an agent on a measurable proxy can produce metric gains while the true objective degrades. Multi-objective systems need explicit divergence monitoring, not just a single improving score.

Generalization

Reward hacking / metric gaming risk grows with the distance between the proxy metric and the true goal and with economic pressure on the optimizing party — a design constraint for eval and reward design in agents.

When a measure becomes a target, it ceases to be a good measure (Goodhart's law, 1975).
Open source video
AI tools and companies are particularly vulnerable to Goodhart's law due to their ungrounded nature and financial incentives.
Open source video
Excessive AI optimization causes goals like solving problems, building theory, training students, and creating aesthetic works to diverge.
Open source video
Tradeoffhigh noveltyweak evidence

Natural friction can be load-bearing. The friction required to work through deep mathematics may be the mechanism by which comprehension is built, so removing friction with AI may eliminate a benefit rather than a cost.

Why it matters

Not all latency, effort, or intermediate work is waste. Some steps produce understanding or verification value as a side effect. Optimizing them away can degrade the very quality the pipeline exists to produce.

Generalization

In agent design, consider whether a 'slow' step is instrumental (a human learning, a check catching errors) before automating it away; classify steps as friction-for-value vs incidental friction.

A 1991 paper of Bourgain annotated by a younger Terence Tao illustrating the friction required to understand deep mathematics.
Open source video
Practicemedium noveltymoderate evidence

Analyzing a contested capability under a conditional hypothesis (an AI Capability Conjecture with a defined success rate, quality, supervision, and cost) lets the community reason about downstream consequences without resolving the truth of the capability claim.

Why it matters

This is a reusable methodology for planning under capability uncertainty: condition on the assumption, derive implications, and design for the consequences regardless of whether the timeline holds. It decouples architecture planning from prediction debates.

Generalization

For teams building on uncertain model capabilities, define an explicit capability assumption with thresholds (success rate, cost, supervision level) and design the system against those thresholds rather than betting on a timeline.

A conditional working hypothesis assumes AI tools will soon become capable of performing a reasonable fraction of research-level tasks with reasonable success, quality, supervision, and cost.
Open source video
conditionalizing on the AI Capability Conjecture to analyze its consequences without getting bogged down in truth-value debates
Open source video
Architecturemedium noveltystrong evidence

Formal machine verification (proof assistants like Lean, Coq, HOL) and human comprehensibility are distinct properties. AI can excel at machine-checkable verification while failing the separate requirement of exposition for human readers.

Why it matters

Correctness oracles and explanation/exposition layers should be separate components with separate evaluation; passing a formal checker does not imply an artifact is understandable or maintainable by humans.

Generalization

In agentic code/data/proof generation, treat 'passes automated checks' and 'is comprehensible and maintainable by humans' as two independent gates; optimizing the first does not satisfy the second.

AI is currently excelling at proof generation and verification, but is very weak at proof exposition and canonicalization.
Open source video

Deep dives

5

Stage-wise capability profiling of multi-stage agent pipelines

Research question

Can per-stage capability benchmarks predict where the end-to-end bottleneck migrates in an agentic pipeline, better than aggregate task-success benchmarks can?

Why

If capability is asymmetric across pipeline stages, aggregate success metrics hide the real constraint and mis-size engineering investment. A per-stage profile turns 'the model is X% good' into 'stage 4 is the saturation point', which is the only form that supports a build decision.

Problem solving involves five stages: open problems -> proof generation -> proof verification -> proof exposition -> proof canonicalization.
Open source video
AI is currently excelling at proof generation and verification, but is very weak at proof exposition and canonicalization.
Open source video
Source video

Backpressure and admission control for agent-generated artifacts

Research question

What admission-control mechanism (confidence gating, rate limiting, triage queues, expiry) keeps the unverified-artifact backlog bounded as generation throughput grows?

Why

Unverified output accumulation is not a model-quality problem and will not be fixed by better models — it is a rate-matching problem. Designing the acceptance filter and backpressure is a concrete engineering deliverable that determines whether an agent system stays trustworthy at scale.

Proof indigestion occurs when AI generates abundant unverified or unreadable proofs that overwhelm traditional peer review and human comprehension systems.
Open source video
Erdős problems containing dozens of AI-generated proof submissions that human experts have not volunteered to verify.
Open source video
Source video

Measuring human comprehensibility as a gate separate from formal verification

Research question

Can human-comprehensibility of a generated artifact be measured reproducibly and independently of whether it passes a formal checker?

Why

If passing an automated checker is treated as the only gate, systems ship artifacts that are correct but unmaintainable, unteachable, and impossible to canonicalize. Making comprehensibility a measured, separate gate is what turns the downstream exposition stage from an aspiration into an engineering target.

AI is currently excelling at proof generation and verification, but is very weak at proof exposition and canonicalization.
Open source video
The community must decrease emphasis on 'first to solve' and increase emphasis on proof digestion: exposition, publication, and canonicalization.
Open source video
Source video

Early detection of Goodhart divergence in multi-objective agent loops

Research question

Does proxy-metric improvement reliably co-occur with measurable degradation on off-proxy objectives, and can that divergence be detected early enough to intervene before the proxy score saturates?

Why

Reward hacking is usually diagnosed after the fact. Turning it into a monitored signal with measurable lead time converts a known pathology into a controllable one, and it is directly actionable in eval and reward design.

When a measure becomes a target, it ceases to be a good measure (Goodhart's law, 1975).
Open source video
AI tools and companies are particularly vulnerable to Goodhart's law due to their ungrounded nature and financial incentives.
Open source video
Source video

Classifying load-bearing friction in agent workflows

Research question

Which workflow steps produce comprehension or verification value as a by-product of their effort, and which are incidental cost that can be automated away without loss?

Why

Indiscriminate friction removal is a plausible way to make an AI-accelerated workflow faster and worse. A classification criterion, even an empirical one, prevents optimizing away the mechanism that produces the quality the pipeline exists to deliver.

A 1991 paper of Bourgain annotated by a younger Terence Tao illustrating the friction required to understand deep mathematics.
Open source video
Excessive AI optimization causes goals like solving problems, building theory, training students, and creating aesthetic works to diverge.
Open source video
Source video

Article ideas

4

The Bottleneck Moved Downstream: Why Scaling Generation Makes Agent Pipelines Worse Before Better

AI capability is unevenly distributed across a workflow's stages, so generation and verification gains migrate the constraint to exposition and canonicalization; teams that keep investing in the already-strong stages are scaling the wrong end of the pipeline and will measure throughput gains that never reach delivery.

Angle

Take a five-stage pipeline (identify, generate, verify, explain, canonicalize), show where models are strong versus weak, and compute what happens to end-to-end throughput when only the strong stages improve.

Source video

Proof Indigestion Is a Queueing Problem, Not a Model Problem

Accumulating unverified agent output is an impedance mismatch between a fast producer and a slow consumer, so the correct fix is admission control and backpressure rather than a better model; without rate matching, higher output quality just increases the backlog of things nobody has read.

Angle

Reframe artifact accumulation in standard queueing terms — arrival rate, service rate, backlog depth, latency — and specify the admission-control primitives (confidence gating, triage queues, expiry) that keep backlog bounded.

Source video

Passing the Checker Is Not the Same as Being Understood

Formal verification and human comprehensibility are independent properties, so treating an automated checker as the sole gate produces systems that are provably correct and operationally unusable; comprehensibility needs to be a separately instrumented gate with its own metrics and its own failure budget.

Angle

Show the two gates as distinct components with distinct evaluation harnesses, and show what breaks when a team optimizes only the checker gate.

Source video

Goodhart's Law as an Architecture Requirement, Not a Warning

Optimizing an agent on a measurable proxy reliably diverges previously aligned goals, so divergence monitoring belongs in the system design alongside latency and cost monitoring; the absence of a divergence signal is a missing component, not a sign of alignment.

Angle

Convert Goodhart's law from an aphorism into a concrete instrumentation spec: which off-proxy objectives to track, what a divergence alarm looks like, and who gets paged.

Source video

Project ideas

4

Digestion Gate

gatehouse

An explicit backpressure-and-triage layer keeps unverified-artifact backlog bounded as generation rate scales, whereas an unbounded submission path grows backlog linearly or worse; the gate's confidence threshold, not model capability, determines whether review latency stays flat.

Proof of concept

Build an artifact intake service with a confidence-gated routing policy (auto-accept, queue for human review, quarantine) plus a configurable submission rate limiter; run it against a synthetic producer whose rate is varied across a 10x range and against an otherwise identical unbounded control path.

Measurement

Backlog depth over time as a function of generation rate (bounded vs unbounded), median and p95 human review wait, fraction of artifacts reaching a canonicalized/approved end state, and fraction abandoned or expired.

Source video

Stage-Wise Capability Profiler

beyond-evals

Per-stage capability scores predict end-to-end pipeline throughput on held-out tasks better than aggregate task-success rates, and the stage with the lowest score accounts for the majority of end-to-end drop-off.

Proof of concept

Instrument a five-stage pipeline (identify, generate, verify, explain, canonicalize) with independent per-stage scorers; fit a throughput model on per-stage scores and a competing model on a single aggregate success rate; evaluate both on held-out task batches.

Measurement

Out-of-sample predictive error (R^2 or rank correlation) of per-stage model vs aggregate model against measured end-to-end throughput; attribution error when predicting which stage is the constraint.

Source video

Divergence Monitor

new

In a multi-objective agent loop optimized against a single proxy metric, statistically significant degradation on off-proxy objectives appears before the proxy score saturates or reverses, giving measurable lead time for intervention.

Proof of concept

Run a multi-objective optimization loop with one optimized proxy objective and two held-out off-proxy objectives; log both continuously and apply change-point detection to find the first significant off-proxy divergence relative to the proxy trajectory.

Measurement

Lead time in optimization steps between the first significant off-proxy divergence alarm and the proxy plateau/reversal; false-positive rate of the divergence detector across seeds.

Source video

Friction Audit

movement-lab

Some workflow steps are load-bearing: removing them lowers downstream human comprehension or error-catch rate, while removing incidental-friction steps does not, and the two classes can be separated by an A/B removal protocol run on the same task set.

Proof of concept

Select a set of tasks with measurable downstream outcomes, identify candidate friction steps, and run a three-arm experiment (control, removal of suspected load-bearing friction, removal of suspected incidental friction) using the same generated artifacts.

Measurement

Downstream comprehension scores and error-catch rate per arm; effect size and confidence interval for each removed step; classification accuracy of pre-registered predictions about which steps are load-bearing.

Source video

Architectural implications

3

The problem-solving workflow is a multi-stage pipeline where AI capability is uneven and the constraint sits in the later, human-facing stages.

Before

Treating generation+verification as the whole task and measuring success by output volume.

After

Explicit stage boundaries with per-stage evaluation, especially for exposition and canonicalization.

Consequence

Engineering investment shifts from generation accelerators toward digestion layers and acceptance filters, since generation gains yield diminishing returns when the downstream stage is saturated.

Source video

Proof abundance requires new community acceptance filters to handle unrouted artifacts.

Before

Peer review and human expertise as the implicit default verification and routing mechanism.

After

Dedicated workflows and infrastructures for triage, review, and canonicalization of large volumes of AI-generated work.

Consequence

An explicit ingestion/triage/canonicalization subsystem becomes a first-class part of the system, not an afterthought, and should include rate limiting or backpressure to avoid indigestion.

Source video

Accountability and provenance must remain with the human even when tools generate the work.

Before

Opaque or undisclosed use of AI tooling in produced artifacts.

After

Transparent disclosure of tool use and human ownership of correctness.

Consequence

Systems need built-in provenance/attribution tracking alongside correctness responsibility assignment.

Source video

Tradeoffs and failure modes

5

Generation throughput vs digestion capacity

Benefit

AI accelerates problem identification and proof generation, producing proofs at rates humans cannot match.

Cost or risk

Unverified or unreadable output accumulates, overwhelming peer review and human comprehension (proof indigestion / impedance mismatch).

Proof indigestion occurs when AI generates abundant unverified or unreadable proofs that overwhelm traditional peer review and human comprehension systems.
Open source video
Source video

Metric optimization vs goal alignment

Benefit

Optimizing measurable proxies like problem-solving yields demonstrable gains on those metrics.

Cost or risk

Previously aligned goals (theory building, training, aesthetics) diverge from the optimized proxy; quality degrades off-metric.

Excessive AI optimization causes goals like solving problems, building theory, training students, and creating aesthetic works to diverge.
Open source video
Source video

Friction removal vs comprehension retention

Benefit

AI reduces the effort to produce or read mathematical work.

Cost or risk

Removing the natural friction required to understand deep mathematics may reduce genuine comprehension.

A 1991 paper of Bourgain annotated by a younger Terence Tao illustrating the friction required to understand deep mathematics.
Open source video
Source video

Formal verification vs human comprehensibility

Benefit

Automated proof assistants can verify correctness at scale.

Cost or risk

Machine-verified results may still be unexplainable or un-canonicalizable for humans, leaving a comprehension gap that verification does not close.

AI is currently excelling at proof generation and verification, but is very weak at proof exposition and canonicalization.
Open source video
Source video

Incentive design ('first to solve')

Benefit

Competitive priority incentives historically drove mathematical progress.

Cost or risk

Emphasizing first-to-solve under AI abundance rewards generation over digestion, worsening the downstream bottleneck.

The community must decrease emphasis on 'first to solve' and increase emphasis on proof digestion: exposition, publication, and canonicalization.
Open source video
Source video

Open questions

4

How can exposition and canonicalization quality be measured or automated, given AI is currently weak at them?

Why unresolved

The summary identifies these stages as the weak point but offers no mechanism for evaluating or generating them.

Research direction

Build evaluation harnesses for human-comprehensibility and canonicalization of generated artifacts, and prototype agents that target these stages specifically.

Source video

What acceptance filters or rate limits prevent proof indigestion when generation outpaces review?

Why unresolved

The summary recommends new workflows, infrastructures, and community acceptance filters but does not specify their design.

Research direction

Design and test backpressure/triage mechanisms for agent-generated artifacts, including confidence-gated routing to human review.

Source video

Which steps' friction is value-producing versus incidental, and can it be preserved while gaining AI efficiency?

Why unresolved

The summary notes friction is required for understanding but provides only an anecdotal illustration, no criterion.

Research direction

Empirically classify workflow steps by whether their removal preserves downstream comprehension or correctness.

Source video

How can Goodhart drift be detected and corrected in multi-objective AI-assisted workflows?

Why unresolved

The summary asserts divergence under excessive optimization but gives no monitoring or correction mechanism.

Research direction

Instrument per-goal metrics alongside proxy metrics and test whether divergence appears as proxy scores improve.

Source video

Key claims

7
comparativeVerification needed

AI currently excels at proof generation and verification but is very weak at proof exposition and canonicalization.

Evidence

AI is currently excelling at proof generation and verification, but is very weak at proof exposition and canonicalization.

Question

Does empirical benchmarking confirm a capability gap between generation/verification and exposition/canonicalization in current models?

Source video
factualVerification needed

Problem solving decomposes into five stages: open problems, proof generation, proof verification, proof exposition, and proof canonicalization.

Evidence

Problem solving involves five stages: open problems -> proof generation -> proof verification -> proof exposition -> proof canonicalization.

Question

Is this five-stage decomposition complete and non-overlapping for describing mathematical research workflows?

Source video
causalVerification needed

When a measure becomes a target it ceases to be a good measure, and AI tools and companies are especially vulnerable due to ungrounded nature and financial incentives.

Evidence

AI tools and companies are particularly vulnerable to Goodhart's law due to their ungrounded nature and financial incentives.

Question

Does the ungrounded nature of AI systems measurably increase susceptibility to Goodhart-style metric divergence compared to other optimizers?

Source video
causalVerification needed

Excessive AI optimization causes previously aligned mathematical goals to diverge.

Evidence

Excessive AI optimization causes goals like solving problems, building theory, training students, and creating aesthetic works to diverge.

Question

Can goal divergence be observed empirically as optimization pressure on a single proxy metric increases?

Source video
factualVerification needed

AI-generated proofs are accumulating unverified on problems such as the Erdős problems because human experts have not volunteered to verify them.

Evidence

Erdős problems containing dozens of AI-generated proof submissions that human experts have not volunteered to verify.

Question

What is the actual volume of unverified AI-generated proof submissions and their resolution rate?

Source video
opinionVerification not requested

Mathematicians should transparently disclose tool use and remain responsible for correctness, and the community should shift emphasis toward proof digestion.

Evidence

Mathematicians should transparently disclose tool use and maintain responsibility for correctness.

Source video
opinionVerification not requested

A conditional working hypothesis can analyze AI's consequences without resolving the truth of capability claims.

Evidence

conditionalizing on the AI Capability Conjecture to analyze its consequences without getting bogged down in truth-value debates

Source video

Connections

5