8,281 papers reveal statistical signatures that expose AI-assisted mathematical proofs
RexDouglass · x · 2026-11-09
A team led by Harvard statistician Zheng Tracy Ke presents an arXiv study on detecting AI involvement in proofs, distinct from AI-polished writing.
- Built 5 cohorts of 8,281 math papers: historical arXiv, disclosed AI-use papers, OpenAI-released manuscripts, plus self-generated AI papers, with 700 matched pairs on the same problems.
- Introduces ProofTrace, representing each paper as a graph of immediate proof support (GIPS) to extract interpretable features.
- PCA shows a qualitative gradient: AI-reworded papers stay near human ones; AI-generated cohorts shift in a shared direction; disclosed-AI papers sit in between, drifting toward AI cohorts over time.
- Matched-pair features: AI-rebuilt papers cite external results only 0.2% (vs 6.1% human), use intermediate steps directly 87% (vs 58%), and have fewer separate proof strands (0.14 vs 0.24).
- Conclusion: AI-assisted proofs carry statistically significant organizational "signatures" that can be detected.
More from AGI Musings
- Defending Dario's 2009 $10k Donation: GiveDirectly Didn't Even Exist Yet — austinc3301 · 2026-11-09
- Anthropic S-1 claims Claude jumped 40% to 90% on open-ended engineering tasks in 9 months — trevposts · 2026-11-09
- Berkeley-led working group: a hardwired decade-long pause of frontier AI is feasible — DavidSKrueger · 2026-11-09
- Dario's grad-school donation math: valuing an adult death at 2-3x an infant's — austinc3301 · 2026-11-09
- Yoshua Bengio: we should just stop training the most advanced AI models — DavidSKrueger · 2026-11-09
- March for an AI Treaty: protests in six cities on Nov 14 call for banning ASI — DavidSKrueger · 2026-11-09