Short answer: it depends on the account you type into. Personal ChatGPT, Codex, and Gemini accounts may train on your prompts until you opt out, and the opt-out covers future chats only. Personal Claude accounts train when model improvement is on. Business, Enterprise, Edu, and API accounts exclude your data by default. For unpublished ideas, the one control you can check is keeping text and audio on your machine.
What happened in September 2026
On September 8, 2026, OpenAI announced that its AI system had produced a proof for the Navier–Stokes problem, one of the Clay Mathematics Institute’s seven Millennium Prize Problems, and had checked it in the Lean proof assistant. Within a day, NYU mathematician Tristan Buckmaster alleged that his unpublished work, typed into OpenAI’s Codex, may have fed that result, and a second mathematician raised the same question about an earlier OpenAI result. OpenAI denies using their work. As of this writing, nobody outside OpenAI can confirm or rule it out.
| Date (2026) | What happened | Source |
|---|---|---|
| Sep 1 | OpenAI starts its Navier–Stokes run “after hearing a rumor which we later realized was related to Levent Alpöge, an Anthropic employee, and Tristan Buckmaster, a math professor at NYU.” | OpenAI |
| Sep 7 | Buckmaster and Alpöge post their own results on related fluid equations, the night before OpenAI’s announcement. | Scientific American |
| Sep 8 | OpenAI announces the proof. Buckmaster, who had used Codex while working on the problem, alleges OpenAI’s system may have drawn on his sessions. OpenAI: “While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.” | TechCrunch, Axios |
| Sep 9 | Andreas Thom (TU Dresden) says he and a colleague had discussed related work with ChatGPT for months before OpenAI’s August non-sofic group result, and that OpenAI researcher Mark Sellke answered his question in one line: “Regarding your conversations with ChatGPT: that did not happen.” | Thom on Mathstodon, Cybernews |
| Sep 10 | OpenAI updates its post: Buckmaster’s Codex prompts from the prior two months “could not have influenced the system in any way, including through training.” | OpenAI |
| Sep 11 | Fields Medalists publish “A Severe Misalignment of AI in Mathematics.” The Clay Institute says “we will provide updates.” | Terence Tao, Clay Institute |
| Sep 17 | An OpenAI spokesperson tells Nature: “no user inputs past July 3rd could have influenced this system in any way.” Nature also reports the Clay Institute will weigh the proof only after peer-reviewed publication. | Nature |
| Sep 21 | OpenAI forms an Advisory Group on Mathematics and AI. Its announcement does not address user data. | OpenAI, TechCrunch |
OpenAI mathematician Sébastien Bubeck said at a press briefing: “We did not use their prompt or proofs to prompt our models” (Scientific American). The allegation remains unproven. Between September 8 and September 17, OpenAI’s statements moved from “cannot rule out” to a July 3 cutoff.
The part no user can check
Two details from the dispute matter more than the verdict.
First, Buckmaster told Nature he had three ChatGPT accounts and had opted out of training on only two of them (Nature). One account in three still allowed training. A lab with ten students, each holding a personal account or two, has up to twenty toggles to keep track of.
Second, Thom turned training off on June 29. In his own words, the setting “is still only a promise whose implementation users cannot audit, and it is prospective: it does not answer what happened to earlier conversations or to derivatives already selected” (Mathstodon). OpenAI’s help page agrees on the timing: after you opt out, “your new conversations won’t be used to train OpenAI models” (OpenAI Help).
He also named the gap in the word OpenAI used:
De-identification may remove a name; it does not remove the intellectual content of a mathematical idea.
Stripping your name from a transcript protects your identity. It does nothing for the idea inside it, and in research the idea is the asset.
Your ideas leave through three doors
Chat assistants. You paste a half-finished lemma into ChatGPT, Claude, or Gemini to sanity-check it. What happens next depends on the account type (table below).
Coding agents. Codex, Claude Code, and similar tools read your repository, your notes, and your scratch files. On a personal ChatGPT plan, the ChatGPT training setting “also applies to your Codex tasks,” and Codex has a separate “Include environments” setting on top of it (OpenAI Help).
Meeting recorders. You say the idea out loud first: in a lab meeting, an advisor one-on-one, or a call with a collaborator. Cloud note takers upload that audio. Otter’s privacy policy lists “training our proprietary AI technology on de-identified audio recordings and on transcriptions” among its uses (Otter.ai). That is the same “de-identified” defense, applied to your spoken ideas.
Universities already restrict the third door. Northwestern IT says “Transcription services not listed below, such as Rev or Otter.ai, should not be used with Northwestern research data” (Northwestern IT). We compiled what 12 universities say in Is Otter.ai allowed for university research?
Which accounts train on your content
Checked against each vendor’s own policy pages on September 25, 2026. Policies change, so follow the link before you rely on a row.
| Account | Trains on your content by default? | Source |
|---|---|---|
| ChatGPT Free, Plus, Pro | Yes, until you turn off “Improve the model for everyone” | OpenAI Help |
| Codex on a personal ChatGPT plan | Follows the same ChatGPT setting; “Include environments” is separate | OpenAI Help |
| ChatGPT Business, Enterprise, Edu | No, unless your organization opts in | OpenAI |
| OpenAI API | No, unless you opt in. Abuse-monitoring logs kept up to 30 days | OpenAI API docs |
| Claude Free, Pro, Max | Yes while the model-improvement setting is on. You choose in Privacy settings; data used for training is kept up to five years | Anthropic |
| Claude for Work, Claude API | No. API inputs and outputs deleted within 30 days by default | Anthropic Privacy Center, retention |
| Gemini app, personal account | Yes while Keep Activity is on, which is the default for adults. Reviewers read a subset of chats | Google, Keep Activity |
| Google Workspace, Workspace for Education | No, without your organization’s permission | Google Workspace |
| Otter.ai | Trains on de-identified audio and transcripts | Otter.ai |
| Local models (Whisper, or an LLM in Ollama or LM Studio) | Nothing leaves the machine, so nothing reaches a vendor. Check with the network off | Northwestern IT |
A checklist for your lab
- Use institution accounts for research chats. If your university licenses ChatGPT Edu, Claude for Education, or Gemini in Workspace, those accounts exclude your data from training by default. Personal accounts do not.
- Audit every personal account, old ones included. Check the training toggle on each. Buckmaster’s third account is the cautionary case.
- Use temporary chats for early ideas. A ChatGPT temporary chat “is not used to improve OpenAI models while it remains temporary,” though OpenAI “may keep a copy of a temporary chat for up to 30 days for safety purposes” (OpenAI Help). Claude’s incognito chats “are not used to improve Claude, even if you have enabled Model Improvement” (Anthropic Privacy Center).
- Keep meeting audio and summaries on the device. Record lab meetings and one-on-ones with a local transcriber, summarize them with a local model in Ollama or LM Studio instead of a cloud chatbot, and turn away note-taking bots that join calls. Northwestern tells researchers to “Turn off all internet connections from your device to check that the software is working locally on your computer and not sending any data to the cloud.”
- Put AI transcription in your IRB protocol. Georgia State tells researchers who use AI-powered transcription services that they “must describe these plans in the IRB application as a data sharing practice” (Georgia State).
- Date your ideas. A preprint, a dated lab-notebook entry, or a git commit gives you evidence of when you had an idea, whatever any model later produces.
Where Quietly fits
We make Quietly, a Mac app that records meetings with no bot in the call and transcribes them on your Mac with Whisper, the open-weights speech model OpenAI released. Transcription and the cleanup pass run on-device, and Quietly saves each meeting as a Markdown file in a folder you choose. The built-in Chat can run on a local model through Ollama or LM Studio, so you can ask a lab meeting what it decided without the transcript leaving your Mac. You can also connect OpenAI or Anthropic with your own API key; then Chat sends the meeting you ask about (transcript, notes, and details such as attendees) to that provider, under API terms that exclude training by default. Northwestern’s guidance says: “Do not use AI transcription software that sends data to third-party services or requires you to provide an API key in order to use the software.” Quietly’s transcription does neither, and Chat does neither when it runs on a local model in Ollama or LM Studio. Quietly runs on macOS only. The researcher setup, including a sentence for your IRB protocol, is on our page for researchers.
FAQ
Did OpenAI use mathematicians’ private chats to solve Navier–Stokes?
Nobody outside OpenAI can say. Buckmaster raised the allegation about Navier–Stokes, and Thom raised a similar one about OpenAI’s non-sofic group result. OpenAI says its researchers and agents did not see Buckmaster and Alpöge’s work before it went public, that Buckmaster’s recent Codex prompts could not have influenced the system, and that no user inputs after July 3 could have. Thom says OpenAI’s reply to him “did not mention the setting or any account-specific check.” As of September 25, 2026, no independent audit has appeared and the Clay Institute has not ruled on the proof.
Does ChatGPT train on my research?
On Free, Plus, and Pro accounts, OpenAI may use your conversations to train models until you turn off “Improve the model for everyone” in Settings, then Data controls. The same toggle covers Codex tasks on those plans. ChatGPT Business, Enterprise, Edu, and the API exclude your data by default.
Does turning off training remove what a model already learned?
No vendor promises that. OpenAI’s help page says the setting stops “your new conversations” from training its models. Thom’s complaint rests on this gap: an opt-out covers what you type next, not what you typed before.
Is Otter.ai safe for research interviews?
Otter’s privacy policy says it trains its AI on de-identified audio recordings and transcriptions, and several universities, including Northwestern, Michigan, and Temple, restrict or prohibit it by name. See what 12 universities say about Otter.ai.
Can I transcribe research interviews without uploading audio?
Yes. Whisper runs on a laptop, and Northwestern IT lists local Whisper-based software for research interviews, with two conditions: test it with the network off, and skip software that sends data to third parties or requires an API key. You can run it yourself (our walkthrough) or use an app that wraps it. Summaries can stay local too if you run an LLM in Ollama or LM Studio.
Sources
- OpenAI, Navier–Stokes announcement and investigation update, Sep 8 and Sep 10, 2026
- OpenAI, Advisory Group on Mathematics and AI, Sep 21, 2026
- Nature, Who gets credit in the AI era? OpenAI maths bombshell sparks debate, Sep 17, 2026
- TechCrunch, OpenAI’s feud with mathematicians is only escalating, Sep 11, 2026
- Andreas Thom, posts on Mathstodon (1/3, 2/3), Sep 9, 2026
- Clay Mathematics Institute, statement on the Navier–Stokes announcement, Sep 11, 2026
- Terence Tao, A Severe Misalignment of AI in Mathematics, Sep 11, 2026
- Otter.ai, Privacy Policy, effective June 16, 2026
- Northwestern IT, AI transcription of research interviews
Update log
- September 25, 2026: Published. All vendor policies and quotes checked on this date.