How Much of Your Next Proof Should You Give OpenAI?

Key takeaways

• A short prompt can reveal the central idea behind an unpublished proof.
• A no-training commitment does not, by itself, settle data retention or access.
• Verifying a proof and documenting research contributions are separate tasks.
• Define what you will share and agree on boundaries with collaborators before submitting unpublished work.

You are stuck on a proof, and explaining the obstacle to an AI seems like a reasonable next step. But the explanation may contain the very insight you hope to publish. Before handing unpublished mathematics to OpenAI, you need to assess how it handles your work as carefully as how it handles the math.

A single sentence can give away the idea

Confidential research brings to mind large datasets or a nearly finished paper. In mathematics, the valuable part might fit in a sentence.

Suppose your breakthrough is a transformation that turns a stubborn problem into something tractable. Explaining that transformation can disclose the essential idea without sharing a single page of your manuscript.

The same applies to a lemma: a supporting result used within a larger proof. The lemma may already be familiar. Your contribution could be recognizing exactly where to use it.

A useful test is: could someone reconstruct my approach from this explanation?

That question matters more than the length of the prompt. Removing your name and institution does little to conceal a mathematical strategy that remains fully visible.

“No training” answers only one question

Suppose a service promises not to use your inputs for model training. That is a meaningful commitment. It does not automatically mean your inputs are immediately deleted or inaccessible to everyone.

Training, retention, and access need separate answers:

• Will prompts or attachments be used to improve models?
• How long are conversations and files retained, and how does deletion work?
• Who can access the data, and under what conditions?

Those answers need to come from the terms that apply to the particular service you intend to use. General confidence in OpenAI cannot resolve the details.

The point is to establish whether the service’s commitments match the confidentiality your research requires. These questions do not establish that a leak has occurred or will occur.

A correct proof does not settle research credit

Imagine an AI suggests the decisive lemma, and you revise it into a working proof. You now have two jobs: verify the mathematics and document how the result came together.

Verification means checking assumptions, testing edge cases, and examining whether the argument holds throughout. It also means checking prior work. A convincing answer may restate an existing result in unfamiliar language.

Documenting contributions requires a different record. Write down the idea you started with, distinguish the AI’s suggestions from your own, and track the revisions that made the argument work.

A conversation log can help reconstruct that sequence. But a transcript alone does not establish priority or originality. It cannot replace checking the proof, reviewing the literature, or explaining who contributed what.

Keep the reasoning and the contribution record clear enough that someone else can assess both.

Set the task before sharing the research

“Help me with my research” leaves too much undefined. A specific task makes the disclosure decision easier.

Asking for an explanation of a published definition may require only public information. Asking for an assessment of a new proof strategy may require revealing the central unpublished insight. Identify which kind of request you are making before composing the prompt.

For collaborative work, agree on the boundaries with your coauthors. Notes you wrote yourself may still contain someone else’s unpublished idea.

A practical routine is to date your notes and record your starting ideas before using the service. Afterward, record which suggestions you adopted and how you changed them. Treat those records with care, too: the conversation itself may contain the confidential material you wanted to protect.

Trusting OpenAI with unpublished mathematics requires clear data-handling terms, agreement among collaborators, and a record of contributions. Before pressing send, read the prompt once more: how much of the discovery is already in that sentence?