NIH and NSF are responding to AI, but not with one simple AI rule. What applies depends on what the AI does, what data it receives, which policy governs and what the institution authorized.
Part of the AI Security for Research Institutions guide series.
AI is becoming part of the research workflow. NIH and NSF are responding—but not with one simple “AI rule.”
Researchers may now use artificial intelligence to:
That creates enormous opportunity.
It also creates a new compliance question:
What happens when an AI tool becomes part of federally funded research?
The answer depends on what the AI does, what information it receives, whose policy applies, what stage of the research process is involved, and what the institution has authorized.
For NIH, the current March 2026 Grants Policy Statement explicitly addresses appropriate AI use in application preparation and warns about plagiarism, fabrication, falsification, and other research-misconduct risks. NIH's May 2026 guidance expands those reminders across applying, managing awards, conducting research, analyzing data, and reporting results.
For NSF, the current PAPPG is 24-1, supplemented by later policy notices. NSF's Supplement 1 to PAPPG 24-1 explicitly states that research misconduct can occur through the use or assistance of tools including AI-based tools.
The result is a landscape where researchers cannot safely assume:
“AI is just another productivity tool.”
For compliance purposes, the better question is:
Where does AI intersect with the rules that govern the research?
NIH and NSF AI compliance considerations are the requirements, policies, and research-integrity controls that apply when AI is used in proposal preparation, peer review, research activities, data analysis, award management, or reporting.
The major issues include:
Research integrity
Was the work accurately represented?
Originality
Did AI materially generate an application or research content?
Data protection
Did restricted or confidential information enter an unauthorized AI service?
Privacy
Could participant, genomic, proprietary, or other sensitive information be exposed?
Accuracy
Are AI-generated claims, citations, data, and results verified?
Disclosure
Does the applicable policy or institution require AI use to be described?
Security
Who can access the AI system, its inputs, outputs, tools, and stored information?
Accountability
Can the institution demonstrate who did what and how AI was used?
These issues apply differently across NIH and NSF.
One of the biggest SEO traps—and one of the biggest compliance mistakes—is treating NIH and NSF as though they have identical AI regulations.
They do not.
The applicable framework may include:
For NIH, the March 2026 NIH Grants Policy Statement is the current standard terms-and-conditions document, and NIH notes that later Guide Notices can supersede information in the Grants Policy Statement.
For NSF, the agency says PAPPG 24-1 remains the current version, supplemented by later policy notices. Check which version applies to your proposal or award.
Therefore:
Always check the agency's current policy source and the specific solicitation or award terms.
NIH's March 2026 Grants Policy Statement contains a dedicated section titled:
“Appropriate Use of Artificial Intelligence (AI)”
NIH says AI tools may be appropriate for limited aspects of application preparation, but warns that AI can create risks involving plagiarism and research misconduct. NIH specifically discusses fabrication and falsification, as well as plagiarism.
NIH also states that applications that are substantially developed by AI, or contain sections substantially developed by AI, will not be considered the original ideas of applicants. NIH says suspected plagiarism, fabrication, or falsification may be referred to the HHS Office of Research Integrity.
This creates an important practical rule:
But researchers remain responsible for:
On May 14, 2026, NIH and HHS Office of Research Integrity staff published “Helpful Reminders to Ensure Integrity of NIH-Supported Research When Using Artificial Intelligence.”
This is particularly valuable because it explicitly addresses AI throughout the research lifecycle—not simply proposal writing.
NIH identifies examples including:
NIH also recommends describing AI use where appropriate, reviewing and confirming information, disclosing relevant image editing, and consulting institutional and journal policies.
The important takeaway:
AI does not transfer research responsibility away from the researcher.
A common misunderstanding is:
“The NIH AI guidance is only about writing grants.”
The current materials are broader.
NIH's May 2026 reminder explicitly discusses AI use while:
That matters because a funded project might use AI after the grant is awarded for:
image analysis
literature processing
statistical assistance
coding
data cleaning
clinical research workflows
scientific reporting
An institution therefore needs an AI-governance process that extends beyond the sponsored-research office.
It may also involve:
IRB
privacy office
information security
research integrity
data governance
technology procurement
principal investigators
One of the strongest real examples of federal research funding AI compliance involves human genomic data.
NIH's NOT-OD-25-081 warns researchers that controlled-access human genomic data cannot simply be shared with unauthorized people or entities and specifically warns about using public generative-AI tools.
NIH states that sharing controlled-access data with public generative AI tools through prompts or other interfaces violates the relevant non-transferability requirements. It also explains that training or retaining AI models using controlled-access genomic data can create disclosure risks.
That changes the practical question from:
“Is this AI tool secure?”
to:
“Is this particular AI data flow permitted under the data-use terms governing the research?”
Those are very different questions.
Suppose a researcher has authorized access to a controlled-access genomic repository.
They want an AI system to summarize the data.
The workflow is:
If the AI platform is an unauthorized third-party service, the issue is not merely cybersecurity.
It can become a data-use compliance problem.
NIH's guidance explicitly warns of disclosure and data-derivative concerns associated with controlled-access genomic data and generative AI.
The practical architecture should instead ask:
Where is the model hosted?
Who controls it?
Who can access the inputs?
Are prompts retained?
Are outputs retained?
Can the model be shared?
Can model parameters become a derivative of controlled-access data?
NSF's situation is somewhat different.
NSF publishes AI policy and guidance for the research community, including its generative-AI notice for merit review.
Researchers should not look for a single universal “responsible AI regulation” that applies to every NSF awardee. For funded research, the operative framework remains the applicable:
For funded research, the operative framework remains the applicable:
+
other applicable NSF policies.
At the time of writing, the operative framework is PAPPG 24-1 plus supplements. Always confirm the version that applies.
That distinction makes the article much more defensible than calling every NSF AI document a regulation.
NSF has one of the clearest AI-specific policies around merit review.
Its December 2023 notice states that NSF reviewers are prohibited from uploading proposal content, review information, or related records to non-approved generative-AI tools. NSF explains that doing so compromises confidentiality because the agency cannot protect non-public information submitted to third-party GAI tools.
NSF also says proposers are encouraged to indicate in the project description the extent to which generative AI was used and how it was used to develop the proposal.
This produces a powerful compliance principle:
Uploading a confidential proposal to an external model can become a disclosure event.
The NSF update becomes even more significant in Supplement 1.
NSF's Supplement 1 to PAPPG 24-1 revised the research-misconduct section to explicitly include use of AI-based tools.
The current language covers fabrication, falsification, or plagiarism whether committed directly or through the use or assistance of persons, entities, or tools—including AI-based tools—in proposing or performing NSF-funded research, reviewing proposals, or reporting results.
In practical terms:
“The AI did it” is not a compliance defense.
The responsible researcher and institution still need controls ensuring that:
| Issue | NIH | NSF |
|---|---|---|
| Current core grants framework | March 2026 NIH Grants Policy Statement | PAPPG 24-1 + supplements |
| AI application guidance | Explicit 2026 language | 2023 GAI merit-review notice + current policies |
| Research integrity | AI risks explicitly discussed | AI-based tools explicitly incorporated into misconduct language |
| External AI + sensitive information | Strong genomic-data example | Strong proposal/review confidentiality example |
| Reviewer restrictions | NIH restricts generative-AI use in peer review | NSF prohibits non-approved GAI use with review materials |
| AI in funded research | 2026 integrity reminder covers research lifecycle | Must comply with research-integrity and award requirements |
| Main lesson | Accuracy, originality, data protection, integrity | Confidentiality, integrity, accuracy, responsible use |
The agencies therefore overlap on principles without having identical AI regimes.
For research institutions, the risk categories can be organized into eight buckets.
AI fabricates data, citations, methods, or results.
Sensitive information enters an unauthorized AI service.
Participant information is exposed, inferred, or retained improperly.
Unpublished research or proprietary material is uploaded to third-party systems.
AI-generated work is presented without appropriate disclosure or attribution.
AI processing violates a repository, DUA, consent, or funding restriction.
AI agents gain unnecessary access to research databases, storage, APIs, or credentials.
The institution cannot reconstruct what the AI did.
That last category is frequently overlooked.
Generative AI and AI agents should not be treated as identical.
A generative assistant might answer:
“Summarize these results.”
An agent might:
That creates a much broader AI research governance problem.
An agent needs:
identity
authorization
tool restrictions
data boundaries
auditability
runtime monitoring
The institution should be able to answer:
What was the agent allowed to do when it performed this research task?
A research agent should receive the minimum permissions necessary.
For example:
Can:
search public literature
Should not automatically:
access controlled participant data
Can:
access approved dataset
Should not automatically:
email raw records externally
Can:
draft proposal language
Should not automatically:
submit an application
Can:
collect documentation
Should not automatically:
certify institutional compliance
The principle is simple:
AI assistance should not automatically become AI authority.
A stronger architecture is:
This allows institutions to enforce rules before the AI receives sensitive information.
It also makes compliance easier to demonstrate.
Every research institution using AI at scale should maintain an inventory.
Record:
AI system
Vendor
Model
Version
Purpose
Research project
Data types
Data classification
Hosting location
Retention
Training use
External sharing
Human reviewer
Approval date
Risk assessment
Change history
This is particularly valuable because AI systems change quickly.
A model can be upgraded without a researcher realizing that its:
have changed.
The original approval may have covered:
Model A
But the production workflow later changes to:
Model B
Then a new tool is connected.
Then an MCP server is added.
Then an agent gains access to cloud storage.
At what point has the original risk assessment become obsolete?
That is a governance question.
An effective program should establish triggers for reassessment, such as:
Federal research funding does not become “AI-funded” simply because AI was used.
The institution and researchers remain responsible for meeting applicable requirements.
NIH's May 2026 guidance is particularly clear that researchers must review AI-generated information, verify accuracy, and use AI consistently with responsible research practices.
NSF likewise states that proposers remain responsible for the accuracy and authenticity of their submissions, including content developed with generative-AI assistance.
That produces a universal rule:
AI can participate in the research process. Accountability remains human and institutional.
Before deploying an AI system in federally funded research, ask:
Watch closely when you see:
Popularity is not authorization.
Vendor security does not answer whether the research use is permitted.
De-identification does not automatically resolve every contractual, consent, repository, or agency requirement.
AI-generated content can still create accuracy, authorship, plagiarism, fabrication, or confidentiality problems.
Human review is useful only when the reviewer receives enough evidence and has genuine authority to correct the system.
A useful institutional model is:
1. Funding Policy
NIH / NSF / solicitation / award terms
↓
2. Research Governance
IRB / research integrity / institutional policy
↓
3. Data Governance
Classification / DUA / consent / access
↓
4. AI Governance
Approved models / vendors / use cases
↓
5. Agent Security
Identity / least privilege / tools / memory
↓
6. Technical Controls
Encryption / isolation / monitoring / logging
↓
7. Human Oversight
Validation / review / approval
↓
8. Audit
Evidence / reproducibility / incident response
The result is more than AI policy.
It is research AI compliance engineering.
HexTyx tests AI agents against live, LLM-backed applications that use synthetic data in controlled environments. Our record covers 28 domain applications. These are not research-institution deployments and no real participants or studies were involved, but the failure patterns map directly onto the boundaries in this guide.
A data-use term is violated when data leaves, whether or not anyone intended it. In a synthetic knowledge assistant (Lumen), the autonomous digest job exported all ten seeded documents to an external address with no human involved, including a restricted document with a live-looking API key and an SSN. That happened in the app’s deliberately vulnerable configuration, but it illustrates the exposure NIH describes for controlled-access data: an AI data flow that sends restricted material somewhere the agreement does not allow.
Evidence matters as much as prevention. HexTyx test runs record what each agent did, which is the kind of record the audit questions above ask for.
The NIH and NSF changes illustrate a broader trend in federal research.
AI is becoming part of the research infrastructure.
That means research institutions increasingly need to treat AI systems the same way they treat other sensitive infrastructure:
identify it
authorize it
monitor it
validate it
document it
secure it
reassess it
The question will increasingly move from:
“Does the institution permit AI?”
to:
“Under what conditions, with what data, for what purpose, using which system, and under whose authority?”
That is a much more mature governance model.
The 2026 NIH and NSF AI compliance landscape is not a simple prohibition-or-permission story.
It is about preserving the principles that already govern federally funded research while adapting them to increasingly capable AI systems.
NIH's current guidance emphasizes:
originality
accuracy
research integrity
responsible AI use
data protection.
NIH's controlled-access genomic-data guidance demonstrates how AI can intersect directly with data-use restrictions and disclosure risks.
NSF's current PAPPG framework and supplements incorporate AI into research-misconduct considerations, while its AI policy and merit-review guidance emphasize confidentiality, transparency, accuracy, and responsible use.
The most important lesson for research institutions is therefore:
Don't treat AI compliance as a checkbox. Treat it as a lifecycle control.
Before AI touches federally funded research, know:
What data?
Which model?
Which vendor?
Which identity?
Which permissions?
Which policy?
Which approval?
Which audit trail?
And when the system changes, reassess it.
Because federal research funding increasingly depends on something that AI makes both easier and harder:
Protect the data.
Preserve the research record.
Control the AI.
Verify the results.
Document the decisions.
That is the foundation of defensible federal research funding AI compliance in 2026.
Before approving AI for federally funded research, ask:
“If NIH or NSF asked us tomorrow to prove exactly how this AI system was used, what data it saw, what it produced, who verified it, and which policy authorized it—could we prove it?”
If the answer is yes, AI becomes a much easier technology to govern.
If the answer is no, the organization has an AI compliance visibility gap.
Protect the data.
Control the AI.
Preserve trust.**
Run the free playground scanner against your own AI application, or read the other guides in this series.