What Funders Have Actually Published About AI — and What It Means for Your Next Proposal

Checked against primary sources on September 18, 2026. Every policy claim links to its source in Section 10; where a number could not be independently verified, it is flagged in the text.

The short version

• Funder positions on applicant AI use cluster into three verified groups: hard-line (NIH), disclosure-leaning (NSF; the UK Research Funders Policy Group, including UKRI, Wellcome, CRUK, and the BHF; and at least one private funder, the Fibrolamellar Cancer Foundation), and the silent majority (most private foundations, with no published policy).
• A fourth position — explicitly permissive in writing — is one I could not verify as of September 2026. That is an absence of evidence, not proof of absence; some funders tell survey researchers they are “agnostic to tools.” If you believe your funder is permissive, get it in writing before you rely on it.
• Detection is weaker than the fear: the best-measured detectors performed near coin-flip accuracy with documented bias against non-native English writers. But detection weakness is not a strategy, because funders who act will act on sourcing checks and human judgment. Traceability is the only defense that works at every funder position.
• The biggest 2026 change is not a restriction. Eight major UK funders now permit themselves to use AI to process applications, and applicant disclosure rates already run 12–25% at major UK charities. Write for reviewers whose workflow may include AI.
• NIH’s six-application cap is fully in force and, for most research institutions, touches more of the pipeline than the AI clause does.

1. The question that actually matters

Consultancy claims of 40–60% time savings on first drafts come from the vendors selling the tools, and even if accurate, they leave the real work untouched: funder research, compliance, the needs statement, the relationships. What changes your risk profile is not your tool stack. It is your funder’s published position — and, as Section 4 shows, your funder’s own toolchain. This piece maps the first and explains the second.

2. The positions that exist

On the fourth position. An explicitly permissive funder — one that states in writing that AI use carries no penalty and requires no disclosure — is one I could not verify. Two honest caveats follow. First, absence of a published policy is not strong evidence that informal permission doesn’t exist: several foundations have told survey researchers they “fund relationships and site visits, not documents” and are agnostic to tools. That may be real permission in practice. Second, informal permission is only as durable as the staff who gave it. If a program officer tells you tools don’t matter, ask one follow-up — “should I note AI assistance anywhere in the application?” — and keep the answer in your proposal file. A funder’s generosity in conversation is not a policy; a written answer to a direct question is the closest private-foundation equivalent you will get.

3. The NIH rulebook

NIH’s AI policy is three notices governing three actors at three lifecycle stages. Confusing them is the most common compliance error:

A rule binding a reviewer does not bind an applicant, and vice versa. If you serve as an NIH reviewer, your obligations under the first row are stricter than anything applicants owe.

The cap, precisely: each PD/PI or MPI is limited to six new, renewal, resubmission, or revision applications per calendar year. Co-investigator, mentor, and subaward-PI roles do not count. An application counts once it passes administrative review; one withdrawn after peer review still counts. T activity codes and R13 conference grants are exempt under the original notice, and the March 2026 NIH Grants Policy Statement also exempts R25 research education grants — verify current activity-code status before routing, because institutional summaries often lag the codified policy. PIs monitor their own counts in eRA Commons. NIH justified the cap by citing PIs who used AI tools to submit more than 40 distinct applications in a single round. It is a volume policy and an originality policy in one notice, which tells you how NIH thinks about AI: not as a writing tool, but as a flood mechanism against a finite reviewer pool.

“Substantially developed by AI” is undefined. No percentage, word count, or prompt threshold exists. Cautious reading: risk rises whenever AI generates or materially shapes the scientific argument, novelty claim, specific aims, rationale, or methodology. The May 2026 notice clarifies the target — unverified AI output entering a submission (a fabricated citation, paraphrased text presented as original). Verified AI assistance is not the violation. That distinction is the whole game, and it is why the AI use log in Section 8 matters.

The certification. Item #8 on NIH’s eProposal routing approval is signed by a named human certifying the application as their own original work containing no substantially AI-developed sections. Draft so that the person signing can sign it honestly.

4. What actually changed in 2026: funders adopted AI faster than they regulated yours

The June 2026 revision of the UK Research Funders Policy Group joint statement — first published September 2023 by UKRI, Wellcome, Cancer Research UK, the British Heart Foundation, the Royal Society, and others — still prohibits reviewers from putting confidential applications into public AI tools and still promises human final decisions. It now also explicitly permits funders to run applications through “enterprise data protected” (EDP) generative AI tools assessed for security. In plain terms: eight of the UK’s largest funders have green-lit AI on their own side of the desk. Research Professional News reported the change in June 2026, alongside applicant disclosure data presented at the ARMA conference: roughly 25% of Wellcome early-career applicants disclosed AI use, 16% at the British Heart Foundation, and 12% at Cancer Research UK. (These figures come from conference reporting rather than published funder documents; treat them as well-sourced but second-hand.) Wellcome deliberately does not pass disclosure data to grant committees — a design choice against reviewer bias, which tells you the policy problem funders are actually solving.

It is not only the institutional giants. Suffolk Community Foundation documented its internal AI use, including a Plinth-powered applicant portal launched in June 2026. In the US, Candid’s 2025 Foundation Giving Forecast Survey found 97% of foundations do not use AI to screen applicants — but 19% are considering it, 3% expect to start, and 12% aren’t sure. As many as one-third have not ruled it out.

What this means for your drafting practice: whether or not an explicitly permissive funder exists, you should now plan for reviewers whose workflow may include automated summarization of your proposal, extraction of your outcomes, and consistency checks across your budget narrative and work plan. That argues for more careful sourcing, not less. An AI-assisted reviewer surfaces unsupported claims faster than a tired human panelist ever did.

5. Detection: weaker than the fear — in both directions

NIH’s July 2025 notice says it “will continue to employ the latest technology in detection of AI-generated content.” That is a deterrent, not a published measurement; no validation of NIH’s detection pipeline is public. What has been measured:

• OpenAI’s own classifier — built by the company that knows these models best — flagged AI text correctly only 26% of the time, with a 9% false-positive rate on human writing. OpenAI discontinued it on July 20, 2023, citing low accuracy.
• Independent press testing (TechCrunch, February 2023) of seven leading detectors found most wrong more than half the time; the best classified AI text correctly in five of seven tests.
• The peer-reviewed study by [Liang, Yuksekgonul, Mao, Wu, and Zou in Patterns (2023)](https://doi.org/10.1016/j.patter.2023.100779) put 91 human-written TOEFL essays through seven widely used detectors: a 61.22% average false-positive rate, with 18 of 91 essays unanimously misclassified and 89 of 91 flagged by at least one detector. Stanford HAI’s write-up explains the mechanism: detectors score linguistic predictability, which correlates with writing proficiency, not authorship.
• The same study shows both bypass and bias: a single “elevate the text” prompt collapsed detection rates from up to 100% to up to 13%, and enriching the same human essays with native-like vocabulary cut the false-positive rate from 61.22% to 11.77% — with authorship unchanged.
• Vendor marketing is not validation. Turnitin’s accuracy claims contrasted with press testing in which it flagged innocent student work.

Hold both halves at once. If fear is paralyzing you: off-the-shelf detection is closer to a coin flip than a lie detector, and it systematically flags the writers least able to afford being doubted. If confidence is tempting you: funders who act will act on sourcing checks, citation verification, and human judgment — and an unverifiable claim is catchable no matter how it was written. One honest limitation: the strongest detection evidence dates from 2023, tested against GPT-2/3.5-era models. The direction of the findings has replicated, but the specific percentages are 2023 measurements, not current ones.

6. The silent majority, read honestly

Candid’s 2024 survey remains the best public dataset on private foundation attitudes:

Three honest conclusions, none of them scary. First, undecided does not mean hostile — the survey’s undecided respondents mostly haven’t formed a view, and the same survey shows funders continuing to weight relationships and site visits over documents. Second, undecided does mean your proposal may be processed without a script: if something in it looks off, the response will be improvised. Third, improvisation is survivable if your application is easy to clear. A funder with no AI policy still has a fraud policy, a due-diligence process, and a program officer who can Google a citation in five minutes. The behavior that protects you under a written AI policy — disclosure, traceability, verifiable claims — is also the behavior that survives an unwritten one. Default to it, and use the direct question from Section 2 with any program officer you have a real relationship with.

7. Section-by-section risk

Assumes the strictest funder you might plausibly submit to; relax only where a published policy says so.

The operational rule: the lower the risk, the more the AI’s job is reorganizing material you already have rather than generating material you don’t. The moment a draft contains a fact you have never seen in a source document, you have left the low-risk zone — regardless of which section it landed in.

8. The protocol

1. Find the funder’s position before drafting. Check the funder’s own site (“AI policy,” “responsible AI,” application FAQ). Find nothing? Record the URL and date checked in your proposal file — “we checked and found no published policy on [date]” is worth more in a dispute than any drafting decision.
2. Keep boilerplate traceable. For background, capacity, and methodology sections, draft from your own prior materials with a simple source log: claim → document → page. This is the one habit that satisfies every funder position simultaneously.
3. Outline the needs statement; don’t fill it. AI has nothing to draw on but your local numbers, so its most confident output there is its most dangerous. Leave placeholders for program staff.
4. Disclose proactively wherever disclosure is mentioned. One sentence — “Generative AI tools assisted with [formatting / language editing / literature organization]; all content was developed and verified by the named applicants” — costs nothing at Positions 1 and 2 and protects you at Position 3. Never disclose a use you didn’t make; never omit one you did.
5. Run a compliance pass against the actual RFP. Word limits, formatting, attachments — then read the evaluation criteria and confirm every scored criterion maps to human-verified content.
6. Keep an AI use log for every submission. Date, tool, sections touched, verification steps, verifier. Under NIH’s May 2026 framing, the log is what separates assistance from the unverified output that constitutes the violation.

9. Vendor claims: one test

Grant platforms often claim that funders evaluate substance, not whether AI was involved. That describes the silent majority’s stated behavior — and stops being true the moment a funder publishes anything, because NIH, NSF, the UK funders, and FCF all evaluate process and disclosure, not just substance. Apply one test to every such claim: does the vendor name a funder, a policy document, and a date? If not, it is a generalization about the least-regulated corner of the market, sold as a universal rule.

Are purpose-built tools worth it over general chatbots? Mainly if you manage a high-volume pipeline. A tool that grounds drafts in your own source documents structurally prevents the one failure mode every funder position can catch: unsourced claims. A general-purpose chatbot has no such anchor. If a tool cannot show you which source document each sentence came from, it has not removed the risk — it has accelerated your path to it.

10. Limitations of this piece

• Detection evidence is 2023-era, as flagged above. Percentages are historical measurements.
• Candid’s acceptance figure is nearly two years stale — directional, not current.
• “No published policy” is partly structural: 97% not screening with AI is not 97% having no written policy, and no clean public dataset exists on written-policy prevalence among private foundations.
• UK disclosure rates (25% / 16% / 12%) come from conference reporting, not published funder datasets.
• The permissive-category claim is an absence-of-evidence statement and should be read as one.

If you have read a foundation policy that unambiguously qualifies as explicitly permissive, send it and this taxonomy will be updated.

11. Pre-submission checklist

• Funder’s published AI position located and saved (URL + date checked), or program officer’s written answer on file
• AI use log completed: tool, sections, dates, verifiers
• Disclosure statement drafted where required or gestured at
• Every factual claim traceable to a supplied source, or marked as staff-supplied placeholder
• All needs-statement statistics verified against primary local data
• Every citation checked against the original source
• Specific aims and scientific narrative written without AI-generated substance
• RFP compliance pass: word limits, formatting, attachments, fonts
• NIH submissions: PI under the six-application cap; eRA Commons checked; signing PI has read the certification language
• Final read-through: could a program officer verify every claim in this application in five minutes?

12. Sources

1. NIH, Notice NOT-OD-25-132, “Supporting Fairness and Originality in NIH Research Applications” (July 2025)
2. NIH, Notice NOT-OD-23-149, generative AI prohibition in peer review (June 2023)
3. NIH, Extramural Nexus, May 2026 notice on AI and research misconduct standards
4. NIH, Grants Policy Statement (March 2026; R25 cap exception)
5. NSF, Notice to Research Community: Use of Generative Artificial Intelligence Technology in the NSF Merit Review Process
6. Research Funders Policy Group, joint statement on AI in funding applications (September 2023; revised June 2026; hosted copy via Cancer Research UK)
7. Fibrolamellar Cancer Foundation, AI Use Policy
8. Suffolk Community Foundation, internal AI use and Plinth applicant portal (June 2026), suffolkcf.org.uk
9. Candid, foundation attitudes toward AI-generated proposals (2024 cycle) and 2025 Foundation Giving Forecast Survey
10. Research Professional News, reporting on the June 2026 UK funder revision and ARMA disclosure data, researchprofessionalnews.com
11. OpenAI, AI classifier disclosure and discontinuation (2023)
12. Liang, Yuksekgonul, Mao, Wu & Zou, [“GPT detectors are biased against non-native English writers,” Patterns](https://doi.org/10.1016/j.patter.2023.100779) (2023); see also Stanford HAI’s summary
13. Grantable, “Can AI Write a Grant Proposal?” (marketing claim assessed against Candid data; cited, not endorsed)