Is it appropriate to use AI on grant proposals?

Using AI to draft a grant proposal is appropriate when three conditions hold: the draft is grounded in your organization's real data rather than the model's guesses, a person reviews and owns every claim before submission, and the funder's own guidelines allow it. Where any of the three fails, the answer is no, and the third one is the one most applicants forget to check.

Funders are not silent on this anymore. NIH's notice NOT-OD-25-132, effective for applications submitted on or after September 25, 2025, states that NIH will not consider applications substantially developed by AI to be the original ideas of applicants. MacArthur's published AI policy tells its staff not to upload confidential, proprietary, personal, or sensitive information about the Foundation, grantees, or vendors into AI tools, and extends disclosure expectations to the consultants, vendors, and grantees it works with. Whether either applies to a given application depends entirely on who you are applying to; the point is that funder policies now exist, they differ, and reading them is part of eligibility review.

Why grounding is the whole question

The sector's top concern about AI is accuracy. In the Center for Effective Philanthropy's 2025 survey of 451 nonprofit and 215 foundation leaders, misinformation and inaccurate results ranked first among nonprofit leaders at 73% and second among foundation leaders at 77%. A proposal is a document made of claims: this many people served, this outcome, this budget, this partner. A general-purpose model asked to write a proposal will produce fluent claims. It has no way to know which ones are true of your organization.

The appropriate use, then, is a model that writes from your records: your programs as you have described them, your outcomes as you have measured them, your budget as it exists in the accounting system, your staff as they are listed. The writing is assisted. The facts are yours.

What review has to mean

Review is not a read-through. For each section, the person who signs the proposal confirms that every number traces to a source, every named partner has agreed to be named, every outcome claim matches what was actually measured, and the narrative describes the organization the funder will meet if they call. The reviewer is the author of record; the model is a drafting tool.

A practical test: could you defend every sentence in a site visit? If a paragraph reads well and you are not sure where its claim came from, it comes out.

What not to put into a consumer AI tool

Client records, donor lists, unpublished financials, board deliberations, or any document under a confidentiality obligation. This is the line MacArthur's policy draws for its own staff, and it is a reasonable line for anyone. Data security and privacy is the largest AI concern among foundation leaders and the second largest among nonprofit leaders: 83% of foundation leaders and 67% of nonprofit leaders in the CEP survey named it. A consumer chat tool with default settings is not the place for the information that makes a proposal specific.

The workable alternative is a tool that operates inside your own data with a contractual commitment that your prompts and outputs are not retained or used for training.

Disclosure

Some funders ask whether AI was used. Answer accurately. A truthful description ("AI-assisted drafting from our own program and financial records, reviewed and edited by staff") is a stronger position than a denial that a program officer can disprove by reading the prose, and it is the position a written organizational AI use policy makes easy to state.

How TappHQ approaches it

TappHQ's grant writer drafts from the organization's own profile: programs, personnel, financials, and outcomes already recorded in the platform. Every section is editable, any section can be regenerated with instructions, the proposal exports to DOCX for the applicant's own review, and an approval step sits between the draft and anything final. Prompts and responses are processed under a zero-retention agreement with the model provider and are not used for training. AI-authored content carries a record of its provenance at the moment it is written, so the disclosure question has a factual answer rather than a guess. The judgment about whether a given funder allows it stays where it belongs, with the applicant.