Mathematician accuses OpenAI of training Astra on his unpublished work; OpenAI denies

Mathematician Andreas Thom posted a long Mastodon thread presenting multiple pieces of evidence suggesting OpenAI may have used his private research conversations with Gábor Kun about the Gromov soficity conjecture to train Astra—the very conjecture at the heart of the non-sofic group problem OpenAI later claimed Astra solved. After Valerio Capraro relayed the accusation on X, it spread widely with boosts from Gary Marcus and Schmidhuber; OpenAI has denied the accusation, but the controversy continues.

Confirmed

• Thom publicly listed evidence on Mastodon suggesting his conversations with Kun may have been used in Astra's training (per m5, m8, first-hand relay by Capraro)
• Valerio Capraro amplified the accusation on X (m5, m8 are his own posts); Gary Marcus (m6) and Schmidhuber (m4) reshared it
• Per m9, Thom is a group theorist at TU Dresden; he argues the core derivation Astra used to crack the "Gromov soficity conjecture" closely matches the non-sofic group technique he and Kun developed over roughly two decades—a path never published publicly
• Quant researcher henryquantum reshared, urging OpenAI and other frontier labs to clarify data provenance seriously and transparently (per m3)
• OpenAI has denied the accusation (per m9)

Unconfirmed

• Whether OpenAI actually used Thom's conversation data remains supported only by Thom's own evidence, framed with words like "possibly" and "suspects"; the substance and basis of OpenAI's denial have not been disclosed

Why it matters

• The case strikes at the core dispute over AI training-data compliance: whether private user conversations or unpublished research can be used for training, especially when the lab then claims a breakthrough built on it
• If true, it would implicate both intellectual property of unpublished academic work and user privacy, which is why numerous researchers and prominent scholars (including Schmidhuber) spread the story; OpenAI's denial has not quelled the doubts and public pressure keeps building

2026-09-10 ~ 2026-09-11 · 9 related posts

Full story(20 episodes)→

• Episode 1: Rumors claim GPT-6 cracks Navier-Stokes, credibility doubtful(2026-09-05, 2 posts)
• Episode 2: OpenAI Claims 10,000 Agents Solved Navier-Stokes in 88 Hours(2026-09-09, 13 posts)
• Episode 3: Existence Proof Effect: Rumor of Solvability Drove OpenAI's Navier-Stokes Attempt(2026-09-09, 3 posts)
• Episode 4: Mathematicians Accuse OpenAI of Racing to Claim Their Unfinished Proof(2026-09-09, 6 posts)
• Episode 5: Anthropic Faces Ethics Backlash Over Math Proof and Scooped Collaboration(2026-09-09, 2 posts)
• Episode 6: Rumors swirl that OpenAI nears Hodge and BSD conjectures as Millennium Prize problems fall(2026-09-10, 27 posts)
• Episode 7: OpenAI Claims Navier-Stokes Solved, Progress on a Second Millennium Problem(2026-09-10, 21 posts)
• Episode 8: Mathematician accuses OpenAI of training Astra on his unpublished work; OpenAI denies(2026-09-10, 9 posts)
• Episode 9: Navier–Stokes Authorship Dispute Fuels Debate Over AI Labs' Scooping Culture(2026-09-10, 6 posts)
• Episode 10: OpenAI denies Codex prompts influenced training as "stolen chat data" claims unravel(2026-09-11, 8 posts)
• Episode 11: OpenAI's Millennium Prize Claim Mired in Authorship and Scooping Dispute(2026-09-11, 10 posts)
• Episode 12: AI Circles Speculate Millennium Prize Problems May Fall to AI(2026-09-11, 2 posts)
• Episode 13: Math Community Splits Over AI's Role After OpenAI's Millennium Prize Claim(2026-09-11, 8 posts)
• Episode 14: Rumor: OpenAI Aims Internal Model at Riemann Hypothesis and P vs NP(2026-09-11, 7 posts)
• Episode 15: OpenAI's Navier-Stokes Proof Passes Rechecks but Faces Scrutiny(2026-09-11, 2 posts)
• Episode 16: OpenAI's 10,000-Agent Navier-Stokes Claim Sparks Math Community Credit War(2026-09-12, 18 posts)
• Episode 17: Misread as Anti-AI, Terence Tao Sparks Debate Over the Value of Mathematics(2026-09-12, 18 posts)
• Episode 18: Fields Medalists' Open Letter on AI Sparks Cross-Disciplinary Debate(2026-09-12, 6 posts)
• Episode 19: OpenAI Claims AI Agent Solved the Navier–Stokes Blow-up Problem(2026-09-13, 8 posts)
• Episode 20: OpenAI Claims 10,000 Agents Solved Navier-Stokes Problem in 88 Hours(2026-09-14, 11 posts)

• [source] Mathematician alleges OpenAI trained Astra on his unpublished work on Gromov's soficity conjecture — ValerioCapraro · 2026-09-10
• OpenAI accused of training Astra on mathematicians' soficity conjecture work — GaryMarcus · 2026-09-10
• Mathematician alleges OpenAI's Astra may have trained on his unpublished soficity work — basedjensen · 2026-09-10
• Mathematicians ask whether OpenAI used their data to crack group theory problems — rbhar90 · 2026-09-10

5 near-duplicate retellings: ValerioCapraro · ctjlewis · iamtrask · gklambauer · 新智元