Using AI in grant proposals
Last updated
527 US foundations · Candid, August 2024
Would accept AI-generated content10%
Would not accept it23%
Undecided67%
In 2024, most had not decided. Several large funders have since published rules.
The short answer
Most large funders that have published a rule allow AI, with conditions. You must check the content, stay responsible for it and, for several funders, say how you used it. Many smaller funders have no rule yet: in a 2024 Candid survey of US foundations, 10% said they would accept applications with AI-generated content, 23% said they would not and 67% were undecided. So read the funder's policy as well as the call, answer truthfully if asked, and never submit a figure or a claim you cannot trace to your own records.
What the evidence says
There is one large survey on whether funders accept AI-written applications: Candid, published on 14 August 2024. It was sent to 2,156 foundations in the United States and 527 replied.
| Answer | Share of foundations that answered |
|---|---|
| Would accept applications with AI-generated content | 10% |
| Would not accept them | 23% |
| Undecided | 67% |
Two limits matter. The survey is from 2024, and it covers one country. Since then several large funders have written their rules down, and those rules are a better guide than the survey.
Readers also ask whether a machine is reading their application. In a later Candid survey, published in November 2025, 97% of 529 foundations said they do not use generative AI to screen applicants or help decide whom to fund.
What funders have published
None of the funders below bans AI outright. Most allow it if you check the content and stay responsible for it, and several ask you to say how you used it. We read each page in November 2026; rules change, so read the funder's own page before you apply.
| Funder | What it tells applicants | Read |
|---|---|---|
| European Commission, Horizon Europe | AI may be used with care. Applicants are fully responsible for the content, must check it and its citations, and must disclose which tools they used and how. | Application form, version 5.1, January 2026 |
| UKRI (UK research councils) | Allowed, but not to generate an application or whole sections without human involvement. Substantive use should be declared; translating into English and improving English need not be. Declaring it does not affect assessment. | Policy, September 2024 |
| Wellcome | Allowed. Any substantive use must be declared on the application form; minor use such as translation or formatting need not be. Do not put confidential information or personal data into AI tools. | Policy, updated April 2026 |
| UK research funders, joint statement | Applicants may use generative AI responsibly. They must not fabricate data or citations, or generate an application without human involvement. | Statement, revised June 2026 |
| US National Science Foundation | Applicants are encouraged to say in the project description whether and how they used generative AI, and are responsible for the accuracy of what they submit. | Notice, updated May 2026 |
| Gates Foundation, Grand Challenges | Calls ask whether AI was used to develop the proposal, to understand how applicants use it, not to penalise them. | Call FAQ, 2026 |
| National Lottery Community Fund (UK) | “We will not reject an application just because AI was used.” It notes that AI can help people for whom English is not a first language, and warns against accepting an AI-written budget unchecked. | Guidance page |
No published rule found. On the same date we found no rule for applicants from the European Commission's external action grants, FCDO, the Global Fund, UN agencies, the Mastercard Foundation or the Ford Foundation. That does not mean anything is allowed: where a funder is silent, the call and the application form decide.
What funders object to
When funders explain their worry, it is rarely the software. It is what tends to come with it:
- Invented facts. A statistic, a baseline or a partner that does not exist. Assessors check, and one false figure discredits the rest.
- Invented or copied sources. A citation that cannot be found, or text lifted from somewhere else without saying so.
- Generic applications. Text that could describe any organisation in any county.
- Volume. Many more applications, each weaker, from organisations that have not read the call.
- Confidentiality. Beneficiaries' details, or the funder's own documents, pasted into a public tool.
- Authorship. A team that cannot explain its own proposal at interview.
Every one of these can also happen without AI. The tool makes them faster.
Where AI helps and where it hurts
| Task | Use AI? | Why |
|---|---|---|
| Reading a long call and listing its requirements | Yes, then check | Fast and usually accurate; verify deadlines and limits against the call itself. |
| Translating your own text into English, or correcting your English | Yes | Funders that have rules treat this as minor use. Check that the meaning has not changed. |
| A first draft from your own documents | Yes | Saves the blank page, if the tool is working from your real records. |
| Tightening a paragraph to a word limit | Yes | Low risk; you can see what changed. |
| Checking a draft against the call's criteria | Yes | A second reader before the deadline. |
| Statistics about your area | No | It will supply numbers that look right. Use a source you can cite. |
| Your track record or staffing | No, unless from your files | It does not know them and can invent. |
| The budget | No | Costs must come from your own budget. An AI-written budget will not add up to your real costs. |
| Deciding whether you are eligible | Only with the call and your records in front of it | A guess here wastes a week. |
| Anything with a person's name, health or case details | No, in a public chatbot | See privacy, below. |
Privacy: what you paste goes somewhere
Before you put your organisation's documents into any AI tool, find out three things: whether what you enter is used to train the model, where it is stored, and who at the provider can read it. Free consumer chatbots often keep what you type, and may use it for training unless you switch that off.
Never paste names, phone numbers, identity numbers, health information or case notes about the people you serve. The same goes for your staff: CVs, salaries and identity numbers in a budget or staffing annex are personal data too.
This is a legal matter, not only good manners. Under Kenya's Data Protection Act 2019, health details are sensitive personal data and may be sent outside Kenya only with the person's consent; a cloud service hosted abroad counts as sending them abroad. South Africa's POPIA treats health information as special personal information and limits transfers abroad. Nigeria's Data Protection Act 2023 has rules of the same kind. Your funders' safeguarding rules apply as well. This is general information, not legal advice.
Whether to disclose
The rule is often not in the call itself. It may be in the funder's general policy, in the instructions of the application form, or a question inside the form.
- The funder forbids AI-generated content: do not use it for that application.
- The call, the form or the funder's website asks whether you used AI: answer truthfully, and say how. “We used an AI tool to draft from our own records; staff checked every figure” is a strong answer.
- You have looked in all three places and found nothing: you are not required to declare it. You may still add one line. You remain responsible for every statement, exactly as if a consultant had written it.
Several funders separate small uses from large ones. UKRI and Wellcome do not ask you to declare translation into English, improving your English or formatting. They do ask about larger uses, such as generating ideas or drafting text.
A safe way to work
- Read the funder's rule first. Look in the call, the application form and the funder's policy pages.
- Start from your own documents, not from a prompt. Registration, annual report, budget, past proposals.
- Make every figure traceable. For each number in the draft, know the document and page.
- Leave gaps visible. A marked gap is honest. A filled gap may be invented.
- Keep people's details out. Remove or mask them before any text reaches an AI tool.
- Have a person read the whole draft against the call, and be able to explain every paragraph.
- Keep a record of which tool you used and for what, and keep your drafts. Some funders ask, and it shows the work is yours.
It also helps to agree inside your organisation who may use which tool, and for what.
How Afrant handles it
Afrant was built around these rules. Drafts are written only from the facts in your own documents, each figure links to the page it came from, missing facts are left as gaps, and a claim nothing supports is flagged. Personal details it recognises are masked before text is sent to the AI, and any file can be kept away from it. Details are on security and privacy and in the comparison with ChatGPT.
It does not make an application acceptable to a funder that forbids AI. Where a funder does, write that one by hand. And where a funder asks how you used AI, say so: you used a drafting tool on your own records, and your staff checked the result.
Questions people ask
Do funders reject AI-written proposals?
A few do. Most large funders that have published a rule allow AI if you check the content, stay responsible for it and say how you used it. Many smaller funders have no rule yet: in a 2024 Candid survey of foundations in the United States, 10% said they would accept applications with AI-generated content, 23% said they would not and 67% were undecided. Read the funder's own policy as well as the call.
Is it OK to use ChatGPT for grant writing?
It is acceptable where the funder allows it and you stay responsible for every word. The danger is not the tool; it is submitting figures and claims that are not true of your organisation, or pasting beneficiaries' personal details into a public chatbot.
Should we tell the funder we used AI?
Look in three places: the call, the application form and the funder's policy pages. If any of them asks, answer truthfully and say how you used it. If the funder forbids AI-generated content, do not use it for that application. If you find nothing, you are not required to declare it, but your organisation answers for the accuracy of everything submitted either way.
Do we have to declare using AI to correct our English?
Usually not. Funders that separate small uses from large ones, such as UKRI and Wellcome, do not ask you to declare translation into English or improving your English. They do ask about larger uses, such as drafting text. Check that the meaning has not changed.
Can funders tell that a proposal was written by AI?
Not reliably with detection software, which makes mistakes in both directions and wrongly flags writing by people whose first language is not English. But assessors recognise generic text: no local detail, no real numbers, the same phrases as twenty other applications. That is what loses marks. Keep your drafts and notes; they show the work is yours.
Can AI write a whole grant proposal?
It can produce a complete-looking document in minutes. Whether it is true is another matter. A general chatbot knows only what you give it, so it can fill the gaps with plausible text. A draft is only as good as the facts it was written from, and several funders say an application must not be generated without a person's involvement.
Will AI replace grant writers?
It changes the work more than it removes it. Deciding which calls to pursue, knowing the funder, checking facts and standing behind the application remain a person's job. The first draft and the cross-checking are where AI saves time.
Read next
- Afrant and ChatGPTWhen a general chatbot is enough for grant writing, and where it stops.
- Proposal writingDrafts written from your own documents. Every figure links to the page it came from, and a claim with no source is flagged.
- How to write a concept noteWhat funders expect in a few pages, the usual sections, the fixed EU form, and how a concept note differs from a full proposal.
Use AI on your own facts, with the sources showing.
Afrant is in pilot and opens one organisation at a time. It is free during the pilot, and there is a sample workspace to explore before you upload anything.