More questions about whether researchers can trust OpenAI with unpublished math
TL;DR Highlight
A mathematician's Mastodon post sparks debate over trusting OpenAI with unpublished research
Who Should Read
Developers who paste confidential or unpublished ideas into LLM APIs and chatbots
Core Mechanics
- The source text is nearly empty
- The captured source is only a Mastodon post title and a JavaScript notice.
- The post body is not available in the source.
- Only the author is confirmed
- The author is Andreas Thom, replying to the tristanbuckmaster account as part 1/3 of a thread.
- No specific claims or numbers appear in the source.
- The debate lives in the comments
- These are the only source facts we can state.
- See the community reactions section for the issues raised.
Community Reactions
- Opinion that training use can't be verified
- Commenters said outsiders can't verify OpenAI's claim that it did not train on those chats.
- Others noted data provenance is hard to track.
- Counterargument that the evidence is weak
- One commenter said the claim rests only on someone saying they discussed the topic with AI.
- A comment quoted OpenAI saying the prompts could not have influenced the system, including training.
- Wish for a canary phrase experiment
- A commenter asked whether anyone had planted a unique phrase in a training-enabled chat to see if a model later knows it.
- They doubted a single transcript leaves much of an impression.
- Experience dating work publicly
- One commenter said they do their research openly on GitHub with dates.
- The goal is to make it obvious if an AI lab later reproduces their work.
How to Apply
- Check training settings on your LLM tools
- Open the data-use settings for every LLM product and API you use today.
- Turn off training use if it's on and record the terms date.
- Classify the data you put in prompts
- Label data as allowed, mask first, or forbidden: unpublished ideas, customer data, source code.
- Send forbidden data only through local models or paths with contractual training exclusion.
- Timestamp your unpublished work
- Commit or date important ideas before sharing them with any model.
- Use the record as evidence if priority is disputed later.
Terminology
Related Papers
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.
90% of US working hours are spent on tasks not covered well by AI benchmarks
Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage
Postgres 19 delayed after AI reviewers find more bugs than humans can fix
LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted
Show HN: Gemini 4 Argon plays StarCraft
LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.
A Plausible Explanation Is Enough to Mislead Multimodal Judges
Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.
Neural networks are clean
Looking at neurons directly shows most have one main job, challenging superposition.