The AI rules, plainly

What the agencies actually say about AI in proposals.

The rules are new, uneven, and widely misunderstood. Here is what each funder has actually published, in plain language — and what it means for a research team. (Summaries, not legal advice; policies as of mid-2026 — always check the current issuance.)

NSF

AI-assisted fabrication is research misconduct

The PAPPG's research-misconduct definition now explicitly covers fabrication, falsification, or plagiarism committed with AI-based tools. A citation an AI invented is treated no differently from one you invented. Proposer disclosure of generative-AI use is required — non-disclosure is treated as misrepresentation — and reviewers are barred from feeding your proposal into unapproved AI tools. What it means: use AI if you like, but you own every word and every reference, and you'd better be able to show which is which.

NIH

Substantially AI-developed applications are not original

NIH does not accept applications substantially developed by AI as the applicant's original work, under Guide Notice NOT-OD-25-132, effective September 25, 2025. What it means: AI as an assistant is survivable; AI as the author is disqualifying.

NASA

Disclose and document — and mind export control

NASA guidance calls for documenting generative-AI use in proposals (which tool, how used) and bars putting export-controlled content into public AI tools. What it means: keep a record of AI's role, and never paste sensitive technical content into a consumer chatbot.

DoD

No published proposal-AI rules

The Department of Defense has published no agency-wide rules on AI in proposals. What it means: silence is not safety — accountability defaults entirely to the PI and the institution, under the general standards for accuracy and integrity that federal submissions already carry.

DOE

No proposal-specific AI policy

The Department of Energy likewise has no proposal-specific AI policy. What it means: the same — you own it all, with no safe-harbor rules to point to.

The pattern across all five:  no agency rewards AI use, every agency holds humans accountable for it, and fabrication is a major issue. The safe posture is the same everywhere: the PI reviews and owns every word, AI assistance stays grounded in your own documents with a citation on every claim, and AI's role is documented. Where an agency permits AI-assisted drafting, Edvisable drafts with you under exactly those controls; where it doesn't, the same engine verifies instead of drafting. That posture is what Edvisable enforces by default.

Want this as a one-pager for your team?

Book a conversation and we'll send the plain-language summary — useful even if we never talk again.

Book a pilot conversation