Top of page

There is no single rulebook for AI use in research. Requirements differ by funder and by journal, and they continue to change. But the major guidance has converged on a consistent set of expectations, and several firm lines now carry real consequences. This page is for Reynolda Campus faculty preparing a proposal, reviewing one, or writing up results.

Four principles that hold everywhere

Whatever the funder or publisher, these four expectations are constant.

Using AI to write or edit a proposal

NIH 

NIH will not treat an application, or any section of one, that is substantially developed by AI as the applicant’s original idea, and will not consider it. This policy took effect for due dates on or after September 25, 2025, and is now reflected here and in the NIH Grants Policy Statement. NIH has deliberately not set a numeric threshold for what counts as “substantially developed,” so your own judgment is what governs. Think of it as a spectrum.

Lower-risk uses:

Higher-risk uses:

The closer AI moves toward authoring the science, the greater the originality risk.

NSF 

NSF asks proposers to indicate in the project description whether and how generative AI was used to develop the proposal, and holds you responsible for the accuracy and authenticity of everything you submit, including AI-assisted content.

If you submit to both agencies, the safe default is the same: keep the scientific substance your own, and disclose any AI assistance.

If you serve as a peer reviewer

Never upload a proposal or manuscript you are reviewing into a public AI tool.

NSF treats anything entered into AI tools outside its firewall as having entered the public domain, which breaks the confidentiality of merit review, can create legal liability, and violates the reviewer confidentiality pledge. NIH likewise prohibits using generative AI to analyze or formulate critiques of applications under review. Journal peer review carries the same near-universal prohibition.

The harm is concrete: uploading surrenders another researcher’s unpublished ideas, and your control over them, with no way to get them back.

Writing up and publishing results

Most journals follow the International Committee of Medical Journal Editors (ICMJE), and the pattern is consistent across major publishers:

A useful rule of thumb: assistive use (polishing your own words) usually needs little or no disclosure; generative use (creating new content) must be disclosed; and AI image generation is frequently prohibited outright. Always confirm your target journal’s specific policy before submitting.

Protecting sensitive and confidential information

This is where funder AI guidance, IRB obligations, and University data-security policy all meet. Whatever a funder permits in general, do not enter sensitive material into any AI tool whose data-handling terms have not been cleared by the University. That includes:

When in doubt about a tool or a use case, contact before you proceed.

Primary sources

These are the authoritative notices. Funder policy changes frequently — confirm against the live notice before relying on any detail here.


Prepared June 2026. Funder and publisher AI policies are evolving rapidly; verify each section against the live primary sources before each submission cycle.