NSF AI Policy

NSF AI policy is best understood as three separate, evolving strands rather than one document — a distinction worth keeping straight because they carry different weight and different consequences. This page lays out what is confirmed as of mid-2026, and flags explicitly where NSF has signaled the position is likely to keep changing.

1. Generative-AI disclosure in proposal preparation — encouraged, not (yet) mandatory

NSF’s starting position on this dates to its Notice to the Research Community on AI, issued December 14, 2023. That notice encourages proposers to indicate in the Project Description the extent to which, if any, generative AI technology was used and how it was used to develop the proposal — it stops short of requiring a checkbox or a mandatory certification. The notice also stated NSF’s intent to update the PAPPG to align with this guidance going forward. As of this writing, disclosure of proposal-preparation AI use remains framed as encouraged practice rather than an enforced submission requirement; proposers should not assume a missing disclosure statement alone triggers a compliance flag, though the underlying content of an AI-assisted proposal is still fully subject to NSF’s accuracy and originality expectations (see below). Because NSF has already stated an intent to formalize this further, re-check the current PAPPG chapter on proposal preparation before relying on ‘encouraged, not required’ as a permanent description — this is exactly the kind of provision likely to tighten in a future PAPPG reissue.

2. AI-assisted misconduct is explicitly covered by NSF’s research-misconduct definition

Separately from the disclosure guidance, NSF used a PAPPG update — PAPPG 24-1, Supplement 1, effective December 8, 2025 — to amend its research-misconduct definition (PAPPG Chapter XII.C) to explicitly name AI. The amended definition covers fabrication, falsification, or plagiarism ‘whether committed by an individual directly or through the use or assistance of other persons, entities, or tools, including artificial intelligence (AI)-based tools, in proposing or performing research funded by NSF.’ In practice, this means a proposer or awardee cannot treat an AI tool’s output as a shield against responsibility: fabricated data, falsified results, or plagiarized text produced or laundered through a generative-AI tool is evaluated the same way as if a person had done it directly, and a qualified human is expected to verify the accuracy and authenticity of any AI-assisted content before it goes into a submission. This is the strand of NSF’s AI policy with real enforcement teeth — it sits inside the same misconduct-investigation framework that governs any other fabrication, falsification, or plagiarism allegation at NSF, not a separate lighter-touch AI process.

3. NAIRR and the AI Institutes — infrastructure and funding policy, not proposal-conduct policy

The third strand is really about what NSF funds and provides access to, not what a proposer must disclose or avoid. The National AI Research Resource (NAIRR) launched as a pilot in 2024 to give researchers, educators, and students access to computing, data, models, and AI expertise that most individual institutions cannot provide on their own. By 2026 NAIRR had grown to support several hundred research projects and thousands of students, and NSF has been moving it from a pilot into a more permanent structure: a solicitation for a NAIRR Operations Center (NAIRR-OC) to run the transition closed in early 2026, with NSF planning a single award of up to $35 million over as long as five years to the organization selected to operate it. Separately, the NSF AI Institutes program (formally the National Artificial Intelligence Research Institutes) funds large, multi-institution centers doing foundational AI research paired with application domains — a funding-mechanism decision, not a rule about how any given proposal should discuss AI use.

Why ‘NSF AI policy’ is a fast-moving target

Every strand above has already moved at least once in the last two years, and NSF has signaled more change is likely: the 2023 notice explicitly previewed a future PAPPG alignment; the research-misconduct language only took its current AI-explicit form in the December 2025 supplement; and NAIRR’s operating model is actively transitioning from pilot to a centrally-run operations center as of 2026. Treat any specific effective date or dollar figure here as accurate as of verification, not as a permanently fixed policy — always check the current PAPPG on nsf.gov directly before a submission deadline rather than relying on a secondary summary, including this one.

How NSF compares to other funders

NSF’s approach — encourage disclosure, treat AI-assisted misconduct under the existing misconduct framework, fund infrastructure separately — is one of several distinct funder positions on generative AI in grant writing; NIH and the European Research Council (ERC) each frame the same underlying concerns differently. See NIH vs NSF vs ERC: AI Policies for Grant Writing for a side-by-side comparison, and generative-AI disclosure statement for how this kind of disclosure works across funders and publishers more generally.

Related terms

• NSF PAPPG (Proposal & Award Policies & Procedures Guide)
• National AI Research Resource (NAIRR)
• NSF AI Institutes (National AI Research Institutes)
• Generative-AI disclosure statement
• National Science Foundation (NSF)

Frequently Asked Questions

Does NSF require researchers to disclose their use of generative AI in a proposal?

No — as of this writing, NSF only encourages disclosure, it does not require it. Its December 14, 2023 Notice to the Research Community on AI asks proposers to indicate in the Project Description whether and how generative AI was used, but stops short of a mandatory checkbox or certification. NSF has stated it intends to update the PAPPG to align with this guidance, so proposers should re-check the current PAPPG rather than assume disclosure will stay optional indefinitely.

Can using AI tools in an NSF proposal count as research misconduct?

Yes, if the AI use results in fabrication, falsification, or plagiarism. PAPPG 24-1, Supplement 1, effective December 8, 2025, amended NSF’s research-misconduct definition to explicitly cover fabrication, falsification, or plagiarism committed through the use or assistance of AI-based tools. A proposer cannot treat an AI tool’s output as a shield against responsibility — a qualified human is expected to verify the accuracy and authenticity of any AI-assisted content before it goes into a submission.

Is the NAIRR part of NSF’s rules on AI use in proposals?

No. The National AI Research Resource (NAIRR) is an NSF-funded infrastructure program, not a proposal-conduct rule. Launched as a pilot in 2024, it gives researchers, educators, and students shared access to computing, data, models, and AI expertise that most individual institutions cannot provide on their own — a funding and access decision, separate from NSF’s disclosure guidance and its research-misconduct definition.

Does NSF’s AI policy apply the same way to NIH and the European Research Council?

No — NSF, NIH, and the European Research Council (ERC) each take a distinct approach to generative AI in grant writing, even though they address similar underlying concerns. NSF’s approach is to encourage disclosure while handling AI-assisted misconduct under its existing misconduct framework and funding AI infrastructure separately, rather than issuing one unified AI rule. See NIH vs NSF vs ERC: AI Policies for Grant Writing for a side-by-side comparison.

Is NSF’s current AI guidance likely to change?

Yes. Every part of NSF’s AI policy has already changed at least once in the last two years, and NSF has signaled more change is coming: the 2023 notice explicitly previewed a future PAPPG alignment, the research-misconduct language only took its current AI-explicit form in the December 2025 supplement, and NAIRR’s operating model is actively transitioning from a pilot to a centrally-run operations center. Proposers should check the current PAPPG on nsf.gov directly before a submission deadline rather than relying on a secondary summary.