03 · What You Need to Know
When Generative AI May Be Used in Research Grant Evaluation
Why Grant Peer Review Is Different From Reviewing Your Own Proposal
Researchers preparing their own grant applications generally have authority to work with their own proposal materials, subject to institutional, contractual, intellectual property, and funding requirements.
Grant reviewers occupy a different position. They receive applications submitted by other researchers for a defined evaluation purpose. The material may contain unpublished discoveries, research strategies, preliminary findings, confidential partnerships, and sensitive institutional information.
Reviewers are not ordinarily authorized to disclose this material to external services simply because doing so might make the review process more convenient.
The distinction is fundamental: permission to read and evaluate an application does not automatically include permission to transmit it to an AI provider.
Consequently, funding agency restrictions on AI use must be interpreted according to the reviewer's role rather than inferred from rules governing applicants.
What Does NIH Prohibit During Grant Peer Review?
In June 2023, NIH issued Notice NOT-OD-23-149, titled The Use of Generative Artificial Intelligence Technologies Is Prohibited for the NIH Peer Review Process.
The notice prohibits NIH scientific peer reviewers from using generative AI technologies to analyze grant applications and formulate peer review critiques.
It also explains that uploading or sharing application content, original concepts, or peer review critiques with online generative AI tools violates NIH confidentiality and integrity requirements.
The notice addresses an accessibility-related exception that requires communication with the designated NIH official before using relevant technology. Reviewers should not assume that an accessibility need independently authorizes the use of an otherwise prohibited service.
NIH reviewers must follow the current confidentiality agreements, reviewer instructions, and applicable authorization procedures. A personal preference for AI assistance does not override the agency's restriction.
What Does NSF Restrict During Grant Peer Review?
In December 2023, NSF issued a notice addressing generative AI in its merit review process.
NSF prohibits reviewers from uploading proposal content, review information, and related records to non-approved generative AI tools. The restriction reflects the confidentiality obligations associated with research proposals submitted for evaluation.
This wording should be interpreted carefully. It identifies prohibited processing through non-approved tools rather than establishing that every conceivable AI technology is prohibited under all circumstances.
However, reviewers should not infer permission merely because a tool is marketed as secure, enterprise-grade, or private. Approval must be established under the applicable NSF procedures.
The NSF notice also distinguishes reviewer obligations from its guidance to applicants concerning AI-assisted proposal preparation.
How Do NIH and NSF Reviewer Policies Differ?
| Issue |
NIH |
NSF |
| Principal restriction |
Prohibits generative AI use for analyzing applications and formulating peer review critiques |
Prohibits uploading proposal and review materials to non-approved generative AI tools |
| Confidential proposal content |
Must not be shared with prohibited AI systems |
Must not be uploaded to non-approved AI tools |
| Reviewer responsibility |
Independent evaluation under NIH peer review requirements |
Independent evaluation under NSF merit review requirements |
| Possible exceptions or authorization |
Follow specified NIH procedures, including applicable accessibility arrangements |
Any approved use must be established through NSF authorization |
This comparison summarizes the cited agency notices. It does not establish the policies of other funding organizations or replace subsequent official instructions.
Which Grant Review Tasks Might AI Technically Assist With?
Generative AI can perform tasks that resemble parts of a research grant review. However, technical capability should not be confused with permission to perform those tasks using confidential applications.
| Potential Task |
What AI Might Do |
Principal Limitation |
| Proposal summarization |
Condense objectives, methods, and expected contributions |
May omit qualifications or misrepresent scientific details |
| Criteria mapping |
Organize proposal content against review criteria |
May misunderstand requirements or overlook relevant evidence |
| Methodological assessment |
Identify apparent design limitations |
May apply inappropriate disciplinary assumptions |
| Statistical review |
Question sample assumptions or analytical procedures |
May produce incorrect technical recommendations |
| Reviewer comments |
Organize or suggest evaluative language |
May introduce unsupported criticism or distort judgment |
| Proposal scoring |
Generate ratings using supplied criteria |
Scores are not necessarily valid, calibrated, or consistent with human review |
For funders that prohibit these activities, their potential usefulness does not make them permissible.
Where a funding organization explicitly authorizes a particular AI-assisted task, its scientific limitations still require attention.
Why Confidentiality Is a Central Concern
Grant applications may contain information that applicants have not made public and that reviewers receive solely for the purposes of assessment.
Uploading such material to an external AI service may constitute unauthorized disclosure or processing, depending on the governing agreement and service arrangements.
The concern extends beyond uploading an entire PDF. Copying a confidential abstract, preliminary finding, unpublished hypothesis, or reviewer critique into a chatbot may also involve restricted information.
Even a paraphrased description may reveal confidential research concepts.
Reviewers should therefore evaluate whether confidential grant proposal material can be entered into an AI tool before attempting any AI-assisted analysis.
Does an Enterprise AI Tool Make Grant Review Permissible?
Not automatically.
Enterprise AI systems may provide contractual protections, access controls, and commitments concerning model training or data retention. These safeguards may be relevant when a funder considers authorizing particular processing activities.
However, a secure service is not necessarily an approved service.
If the funding organization's policy prohibits a task or restricts processing to specifically authorized environments, a reviewer cannot independently override that requirement by selecting a commercially available enterprise product.
Authorization must come from the responsible organization under its applicable procedures.
Can AI Evaluate Scientific Merit Reliably?
AI can generate plausible assessments of research significance, methodological rigor, feasibility, and expected impact. Yet such assessments may be inaccurate, incomplete, or influenced by patterns that do not reflect the merits of the particular proposal.
A model might criticize an exploratory qualitative study for lacking a statistical power calculation, overlook an important limitation in a specialized experimental design, or overvalue familiar terminology associated with successful grant applications.
Its assessment may also depend on prompt wording and the information supplied.
Consequently, AI-generated judgments should not be treated as equivalent to independent expert evaluation. This concern remains relevant even in settings where the funding organization permits AI assistance.
Why AI-Generated Scores Are Particularly Problematic
Grant review scores can influence consequential decisions about which projects receive funding.
An AI model may generate numerical scores that appear precise, but those numbers are not necessarily calibrated to the funder's evaluation process.
Moreover, funding organizations may use different scoring systems and criteria. NIH's simplified review framework, for example, distinguishes scored scientific merit factors from the assessment of expertise and resources for sufficiency.
An AI system that applies an invented scoring rubric may produce an apparently objective evaluation that misrepresents the actual review framework.
Even when a scoring rubric is supplied, the model's numerical output does not establish that the evaluation is scientifically valid or free from bias.
Can AI Bias Affect the Fairness of Grant Review?
AI systems may reproduce patterns or biases present in their training information or introduced through their design and evaluation procedures.
In grant review, this could influence assessments of unconventional research questions, interdisciplinary approaches, institutional contexts, or methodological traditions.
A model might favor familiar proposal structures or established research areas without adequately considering whether an unusual approach is scientifically justified.
The possibility of AI bias influencing research grant scoring is particularly consequential because review outcomes affect access to limited research resources.
Human review is not free from bias either. Nevertheless, introducing AI requires evidence that the system is suitable for the task and that its use does not undermine the fairness or accountability of the evaluation process.
Which Parts of Grant Review Require Human Judgment?
Scientific merit evaluation involves more than identifying keywords or checking whether required sections are present.
Reviewers must consider whether the proposed work addresses a meaningful research problem, whether the methodology is defensible, how uncertainty affects feasibility, and whether the anticipated contribution justifies investment.
These assessments require disciplinary expertise, interpretation of evidence, and accountability for evaluative judgments.
Where AI assistance is permitted, the reviewer should remain responsible for conclusions and should not endorse criticisms or scores that cannot be independently explained and defended.
However, human oversight alone does not authorize a prohibited AI activity. The first requirement remains compliance with the funder's rules.
What About AI Used by Funding Agencies for Administrative Screening?
Administrative screening and scientific merit review are related but distinct functions.
A funding organization might consider authorized automated tools for checking missing documents, identifying incomplete fields, or organizing applications. Such tasks do not necessarily require the same scientific judgments involved in evaluating research quality.
Nevertheless, administrative automation can still affect applicants and requires appropriate governance, accuracy, confidentiality, and procedural safeguards.
Permission for agency-operated administrative tools should not be interpreted as permission for external peer reviewers to use their own AI services.
Watch Out
Do not upload confidential grant applications, reviewer notes, or unpublished proposal details to a generative AI service without explicit authorization under the funding organization's policies. A tool's convenience, privacy settings, or no-training commitments do not independently establish permission.
