How to Edit a Grant Proposal (NIH / NSF / Horizon Europe) with AI

A Nature analysis in early 2026 reported that NIH proposals edited with AI assistance had measurably higher fundability scores than matched unedited proposals from the same labs. The effect was small but consistent. The same analysis flagged a more troubling pattern: a subset of AI-edited proposals showed reviewer comments suggesting the prose felt "uniform" or "templated," and those proposals scored worse than baseline. AI editing helps grant prose, but only when done right.

Grant proposals are different from journal manuscripts. The constraints are harder (strict page limits, prescribed sections, font requirements). The audience is different (program officers and review panels reading hundreds of proposals in days). The stakes are concentrated (one proposal, one cycle, six months of waiting). And the disclosure rules vary significantly across NIH, NSF, and Horizon Europe — getting them wrong creates funding integrity issues, not just embarrassment.

This guide walks through where AI legitimately helps in a grant proposal, where it doesn't, what each major funder currently allows and requires for disclosure, and a workflow that respects the constraints.

The constraints that shape grant writing

Before any editing, understand what you're working within. The constraints are not advisory.

Page or character limits are absolute. NIH R01 research strategy is 12 pages. NSF main project description is 15 pages. Horizon Europe RIA/IA proposals have a 45-page Part B limit. Going over by even a paragraph triggers automatic rejection at most agencies — your proposal isn't reviewed, period.

Font and formatting are checked. NIH requires Arial, Helvetica, Palatino Linotype, or Georgia at 11 point minimum. NSF requires Times New Roman, Helvetica, or similar at 11 point with 1-inch margins. Both check. Programs have rejected proposals for 10.5 point font.

Required sections are not optional. NIH proposals must include Specific Aims, Significance, Innovation, Approach. NSF requires Intellectual Merit and Broader Impacts addressed substantively. Horizon Europe requires Excellence, Impact, and Implementation. Missing any required section is fatal.

The reviewer reads fast. A program officer or panel reviewer may have 8-15 proposals to evaluate in a cycle. They skim. They scan headings. They form judgments in the first page or two. The Specific Aims (NIH) or first two pages (NSF) carry disproportionate weight.

These constraints define what editing means in grant context. Every cut, every restructuring, every word change has to respect them.

Where AI legitimately helps

These uses produce better proposals without creating integrity issues.

Specific Aims polish (NIH). The Specific Aims page is the single most-read part of an NIH proposal. AI is excellent for tightening the four key sentences (what we know, what we don't know, what we propose, what the impact will be). Edit, don't generate from scratch.

Significance section tightening. Significance sections are often padded — authors over-cite, over-frame, and over-explain. AI editing for clarity and brevity is straightforward here without changing the substance.

Jargon reduction for cross-disciplinary reviewers. NSF and Horizon Europe reviewers often come from adjacent fields. AI can flag discipline-specific jargon that needs definition, and suggest plainer alternatives that preserve precision. This is especially valuable for Broader Impacts (NSF) and Impact (Horizon Europe) sections, which non-specialists evaluate.

Consistency across the proposal. A long proposal drafted over weeks by multiple co-PIs accumulates inconsistencies — different acronym definitions, different verb tenses, different framing of the same concept. AI editing across sections catches these. Use our AI proofreader with a Comprehensive editing pass for this.

Bringing prose into the agency's house style. NIH proposals read differently from NSF proposals. Horizon Europe has its own evaluation criteria that drive specific framings ("excellence" claims, "impact pathway" language). AI editing can shift prose toward the agency's expected register without changing substance.

Cover letter polish. The cover letter (NIH) or proposal summary (NSF) is short enough to benefit from intensive editing. The cover letter techniques that work for journals apply, with agency-specific framing.

Where AI should NOT do the work

These uses cross the line into the substantive intellectual contribution that must be yours.

Preliminary data interpretation. What your pilot data show, what it means for your hypothesis, and what its limitations are — these are scientific judgments. AI can polish how you express them. It cannot make them for you.

Novel argumentation. The case for why your approach succeeds where others have failed, the gap your project fills, the mechanistic theory underlying your aims — these need to come from you. Reviewers can usually tell when this language is generic, and generic loses to specific in tight competitions.

Budget justification. Don't draft budget justification with AI. The numbers are too easy to get wrong, and reviewers and grants management officers check carefully. Write this yourself; have AI proofread it.

Letters of support and biosketches. Letters of support are signed by other people; you shouldn't be drafting them at all, let alone with AI. Biosketches are formulaic but require accurate personal history.

Responses to reviewer summary statements (resubmissions). The introduction to a resubmission addressing reviewer concerns is structurally similar to a response-to-reviewers letter. The substantive responses must be yours. AI can help tighten the language; it can't make the technical case.

Agency-specific disclosure rules

This is where careful attention pays off. Each funder treats AI use differently, and the rules have changed multiple times in the past two years. The following is current as of mid-2026.

NIH. Current guidance allows AI tools for editing and language refinement. Use must be disclosed in the proposal cover letter, with specificity about which tool and which sections. AI cannot be used to generate scientific content, hypotheses, or interpretation. Peer reviewers are prohibited from using AI tools to evaluate proposals (this affects you indirectly — your proposal is read by humans).

NSF. Current guidance allows AI for editing assistance. Disclosure required in the project description if AI was used substantively. AI-generated text is not prohibited but must be disclosed. NSF is more permissive than NIH on AI use in proposal preparation but stricter on disclosure format.

Horizon Europe. The strictest current framework. Disclosure required in a dedicated section of the proposal. AI use in idea generation is restricted; AI use in editing is allowed with disclosure. The evaluation panels are explicitly instructed to assess "the genuine contribution of the applicants" — heavy AI use that masks the applicants' own thinking can hurt scores even when properly disclosed.

Wellcome Trust, ERC, and other major funders. Each has its own guidance, and most published or updated their policies in 2025. Always check the current funder guidance for the cycle you're submitting to — these change faster than journal guidance does.

Common requirements across all major funders. AI cannot be listed as a contributor or author of the proposal. The applicants take full responsibility for the content. Any AI use must be disclosed; failure to disclose is treated as a research integrity issue.

For the broader picture of AI disclosure across academic publishing, see our AI-use disclosure statement guide — many of the templates adapt for grant context.

A workflow that respects the constraints

A sequence that produces better proposals without creating disclosure or integrity issues.

Step 1: Draft the entire proposal yourself first. Specific Aims, Significance, Innovation, Approach (NIH); Project Description with Intellectual Merit and Broader Impacts (NSF); Excellence, Impact, Implementation (Horizon Europe). Draft completely before any AI involvement. The substance of the proposal — what you're proposing, why, and how — comes from you.

Step 2: Run a self-edit pass. Read aloud. Cut padding. Verify page or character count. Most first drafts are 15-25% over the limit and need structural cuts. Apply the same techniques as the cutting words guide.

Step 3: AI editing pass on the Specific Aims (NIH) or first two pages (NSF/Horizon). These pages carry disproportionate weight. Paste them into the proofreader, run a Standard editing pass, review tracked changes individually. Accept changes that tighten or clarify; reject changes that flatten your voice or remove specificity.

Step 4: Section-by-section editing pass. Proceed through the proposal in subsection-sized chunks (500-1,000 words). Run the same Standard pass. Watch for two failure modes: edits that introduce generic phrasing, and edits that subtly change technical claims. Reject both.

Step 5: Consistency pass across the full document. Run a Comprehensive editing pass on the assembled document to catch acronym inconsistencies, tense shifts, and cross-section terminology drift. Accept consistency edits; verify any substantive changes.

Step 6: Final read-aloud. Read the proposal start to finish, out loud if possible. AI-edited text sometimes sounds smooth in isolation but uniform across paragraphs. If multiple sections sound identical in rhythm, restore some variation manually.

Step 7: Write the disclosure. Use the template appropriate to your agency. Be specific about which tool, which sections, and what role it played. Generic disclosure is worse than detailed disclosure.

Step 8: Check page or character count one more time. Editing sometimes adds words. Verify you're still within the limit, and if you've gained margin, decide deliberately whether to add substance or leave the margin as breathing room.

Step 9: Co-investigator sign-off. Every co-PI and key personnel should read the final proposal, know about the AI use, and approve. Surprise disclosure during a discussion with a program officer is much worse than disclosed disclosure.

Step 10: Final formatting check. Font size, margins, line spacing, page count — verify all against the call. Agencies have rejected proposals for 10.5 point font in a footnote. Don't lose your proposal to formatting.

Common pitfalls

Specific Aims sounding generic. AI-edited Specific Aims pages have a recognizable rhythm: short topic sentence, two-sentence justification, transition sentence. Reviewers see this pattern often enough that it now reads as a flag for over-editing. If your Specific Aims sound like every other AI-edited Specific Aims you've read this year, restore your own voice on the third and fourth sentence of each aim.

Innovation section that doesn't innovate. Innovation sections (NIH) and Excellence sections (Horizon Europe) require you to make a strong, specific claim about what's new. AI editing tends to soften claims into hedged language. Push back. The strong specific claim is what wins.

Broader Impacts as boilerplate. NSF reviewers see thousands of Broader Impacts sections. AI editing tends to produce versions that read like every other Broader Impacts section. The strongest Broader Impacts sections describe specific, concrete activities tied to specific, named populations and outcomes. Edit toward specificity, not smoothness.

Forgetting to update the cover letter. A proposal cover letter that's identical to last cycle's after you've substantially revised the proposal flags carelessness. Edit the cover letter to reflect the current proposal's framing and any responses to prior reviews.

Letters of support drift. If you've revised your aims, the letters of support written for the previous version may not match. AI editing of your proposal doesn't fix this — you need to coordinate with letter writers.

Disclosure mismatch. Your cover letter says you used AI for editing only. Your proposal contains a section that's clearly AI-generated. Reviewers and program officers catch this kind of mismatch more often than you'd expect. Match your disclosure to your actual use.

See the Full AI Proofreader

Tracked-changes editing for grants and proposals. Free tier includes every feature.

Frequently asked questions

Q: Will using AI to edit my proposal hurt my fundability score?

The Nature analysis we cited at the start showed AI-edited proposals had higher scores on average, but a subset showed lower scores when the editing produced uniform or templated prose. The deciding factor is the depth of editing. AI used to polish prose you wrote — fixing awkward sentences, tightening padding, catching typos — generally helps. AI used to generate substantive content, draft Significance arguments, or smooth proposals into a generic rhythm hurts. Stay on the editing side of the line, and AI helps.

Q: How explicit does my AI-use disclosure need to be?

Match the agency's specificity requirements. NIH currently accepts a brief cover-letter mention identifying the tool and the role it played. NSF wants slightly more detail in the project description if AI was used substantively. Horizon Europe requires a dedicated section with specifics about which tool, which sections, and what role it played. Vague disclosure ("AI tools were used in preparation") is treated as non-compliant by Horizon Europe and increasingly by NIH. Specificity protects you.

Q: Can I use AI to translate a Horizon Europe proposal from my native language to English?

Yes, with disclosure. AI translation is now considered standard practice for non-native-English applicants to English-language calls. Disclose the tool used (DeepL, our AI translator, or other), confirm that all translated content was reviewed by authors fluent in English, and confirm that any substantive editing of the translated text is also disclosed. Horizon Europe explicitly does not penalize non-native English applicants for using translation tools when properly disclosed.

Q: My collaborators don't know I used AI to edit the proposal. Do they need to?

Yes. Every co-investigator and key personnel listed on the proposal is responsible for the content, which means they need to know what tools were used in its preparation. This isn't a courtesy; most disclosure templates require all listed personnel to be aware. Surprise disclosure during program officer discussions or post-award questioning damages the team's credibility and can affect future funding decisions. Have the AI-use conversation with co-PIs before submission, not after.