Direct comparison
NIH vs NSF vs ERC: AI in Grant Writing
How NIH, NSF, and ERC currently treat AI use in grant proposal writing and peer review/evaluation, compared side by side — disclosure, limits, enforcement.
Side-by-side comparison
| Dimension | NIH | NSF | ERC |
|---|---|---|---|
| Governing document(s) | NOT-OD-25-132, Supporting Fairness and Originality in NIH Research Applications (released July 17, 2025); NOT-OD-23-149, generative AI prohibited for NIH peer review (June 23, 2023). | Notice to the Research Community: Use of Generative Artificial Intelligence Technology in the NSF Merit Review Process (Dec 14, 2023); PAPPG 24-1 Supplement 1 / NSF 26-200 (Dec 8, 2025), which revises the research-misconduct definition. | Current position of the ERC Scientific Council on AI (issued Dec 1, 2023; updated March 25, 2026) for applicants; a separate document, The use of AI in grant proposal evaluation (published around March 23-24, 2026), for reviewers. |
| Can AI help write the proposal at all? | Not banned. AI may be appropriate to assist with limited aspects or in specific circumstances, but NIH will not consider an application substantially developed by AI (or containing sections substantially developed by AI) to be the applicant's original idea. | Allowed. Proposers are responsible for the accuracy and authenticity of their proposal submission, including content developed with the assistance of generative AI tools. | Explicitly allowed for brainstorming, literature search, and revising, translating, or summarizing text u2014 the Scientific Council notes researchers regularly seek input from AI technologies to brainstorm or generate ideas. Full authorship responsibility stays with the PI regardless. |
| Mandatory disclosure of AI use? | No mandatory disclosure field identified in NOT-OD-25-132 itself u2014 the operative standard is originality/substantiality, not a disclosure checkbox. | Not mandatory, but proposers are encouraged to indicate in the project description the extent to which, if any, generative AI technology was used. | No standardized applicant disclosure statement identified in the current position u2014 framed around retained authorship responsibility rather than a disclosure form. |
| Where the bright line actually sits | Content substantially developed by AI isn't treated as original; if identified post-award, NIH may refer the matter to the Office of Research Integrity and take enforcement action independently of any misconduct finding. | The Dec 2025 misconduct-definition revision names AI-based tools explicitly: fabrication, falsification, or plagiarism committed directly or through the use or assistance of other persons, entities, or tools u2014 including AI-based tools u2014 is prohibited in proposing, performing, reviewing, or reporting NSF-funded research. | No equivalent applicant-side substantially-developed-by-AI prohibition was identified in the current position statement u2014 the ERC's sharpest bright line is on the evaluator side (see below), not the applicant side. |
| Rules for the funder's own reviewers/evaluators | Generative AI use by NIH peer reviewers (and Advisory Council/Board members) is entirely prohibited for analyzing applications or drafting critiques (NOT-OD-23-149) u2014 uploading application material to an external AI tool violates NIH peer-review confidentiality. | Reviewers are prohibited from uploading any content from proposals, review information, or related records to non-approved generative AI tools u2014 a breach of the confidentiality pledge on NSF Form 1230P. Reviewers may still use approved/non-confidential GenAI for general purposes. | Reviewers may not use AI to summarize proposals, assess scientific merit, or generate draft evaluations (non-delegation principle), and may not upload proposal content to external AI systems (confidentiality principle). Limited permitted use: polishing the language of their own review text or searching general, non-confidential information. |
| Funder-specific quirk | The same notice caps a single PI/PD to six new, renewal, resubmission, or revision applications across all council rounds per calendar year (T-series and R13 conference grants excluded) u2014 bundled with the AI-originality language because both target volume-gaming, including AI-assisted mass application generation. | No comparable numeric application cap tied to AI use; NSF's distinguishing move is folding AI explicitly into its existing, general-purpose research-misconduct definition rather than issuing AI-specific proposal rules. | Distinguishing feature is a formally separate, freshly dated (March 2026) evaluator-facing guidance document, issued apart from u2014 and more detailed than u2014 the applicant-facing position, reflecting that ERC's proposal evaluation is panel-based and confidentiality-sensitive. |
| Effective / publication dates | Peer-review AI prohibition: June 23, 2023. Application AI-originality + submission-limit notice: released July 17, 2025, effective for the September 25, 2025 receipt date onward. | Merit-review GenAI notice: December 14, 2023. Research-misconduct definition revision naming AI tools: effective December 8, 2025 (PAPPG 24-1 Supplement 1 / NSF 26-200). | Applicant-facing position: first issued December 1, 2023; updated March 25, 2026. Evaluator-facing guidance: published around March 23-24, 2026. |
| Enforcement mechanism | Office of Research Integrity referral, plus grant-administration remedies: disallowing costs, withholding future awards, suspending the grant wholly or in part, or termination. | NSF's standard research-misconduct process under the revised Chapter XII.C definition u2014 the same institutional-inquiry/ORI-equivalent pathway NSF already uses for any fabrication, falsification, or plagiarism finding. | Handled through the European Commission's/ERCEA's standard grant-agreement and research-integrity procedures u2014 the AI guidance sets the rule; a breach is processed under existing grant-agreement misconduct provisions, not a novel AI-specific channel. |
| Stated rationale | Explicitly framed as Supporting Fairness and Originality u2014 protecting peer-review integrity and the genuine originality of the ideas being funded. | Framed around merit-review confidentiality and preventing fabrication, falsification, and plagiarism, whatever tool is used to commit it. | Framed around two evaluator-side principles u2014 non-delegation of the evaluation task and confidentiality/privacy protection u2014 paired with trust that applicants retain full authorship responsibility for AI-assisted drafting. |
| What's still unsettled | Substantially developed by AI has no published bright-line test (e.g., a percentage or specific AI-output threshold) u2014 institutions and applicants are left to judge case by case. | The full NSF 26-1 PAPPG u2014 the natural place to consolidate AI guidance u2014 has been deferred pending OMB Uniform Guidance changes; NSF is issuing supplemental Policy Notices in the interim rather than a single consolidated AI policy. | Applicant-facing and evaluator-facing AI guidance remain two separate, differently-dated documents rather than one unified policy, and no mandatory applicant AI-disclosure requirement was identified as of this verification. |
Common questions
FAQ
What is NIH's policy on using AI to write a grant application?+
NIH does not prohibit AI assistance in preparing an application outright, but under NOT-OD-25-132 (July 17, 2025) it will not treat an application substantially developed by AI as the applicant's original idea. If AI-substantial development is identified post-award, NIH may refer the matter to the Office of Research Integrity and separately pursue enforcement actions such as disallowing costs, withholding future awards, or terminating the grant.
Is the NIH six-application limit related to its AI policy?+
They're issued in the same notice (NOT-OD-25-132) but are two distinct measures. The application cap limits a PI/PD to six new, renewal, resubmission, or revision applications across all council rounds per calendar year (T-series and R13 conference grants excluded), effective for the September 25, 2025 receipt date onward. NIH frames both the cap and the AI-originality standard as responses to the same underlying risk: gaming application volume, including via AI-assisted mass application generation.
What is NSF's current policy on generative AI in proposals?+
NSF permits GenAI-assisted drafting and encourages u2014 without mandating u2014 disclosure of AI use in the project description. Proposers remain fully responsible for the accuracy and authenticity of their proposal submission, including any content developed with AI assistance. Since a December 8, 2025 revision to the PAPPG's research-misconduct definition (Chapter XII.C, via PAPPG 24-1 Supplement 1 / NSF 26-200), fabrication, falsification, or plagiarism committed through the use of AI-based tools is explicitly prohibited, on the same footing as if committed directly.
Can I use AI tools like ChatGPT to help write my grant application?+
Under all three funders' current guidance, limited AI assistance u2014 brainstorming, literature searching, editing, translating u2014 is not banned. None of NIH, NSF, or ERC prohibits AI as a writing aid. What each funder restricts is delegating the proposal's actual intellectual substance to AI: NIH won't credit an application substantially developed by AI as original, NSF holds the proposer fully accountable for AI-assisted content under its misconduct definition, and ERC keeps full authorship responsibility with the PI regardless of what AI assistance was used.
Do NIH, NSF, or ERC require me to disclose AI use in my proposal?+
Only NSF has an explicit disclosure recommendation, and even that is non-mandatory (encouraged, not required) u2014 indicating the extent of GenAI use in the project description. Neither NIH's nor the ERC's current applicant-facing guidance specifies a mandatory AI-use disclosure statement or field, as verified directly against their published positions.
