Should Scientists Be Allowed to Use AI in Grant Proposals?

Getting your Trinity Audio player ready...

Yes. Scientists should generally be allowed to use AI in grant proposals, provided that AI remains an assistive tool rather than a substitute for scientific authorship, judgment, and responsibility.

Using AI to improve grammar, translate text, organize a budget narrative, identify unclear passages, or reduce a proposal to the required word count is comparable to using professional editing software. Using AI to invent hypotheses, fabricate citations, generate preliminary results, or disguise a researcher’s lack of understanding is fundamentally different.

The correct policy is therefore neither an unrestricted permission nor a total prohibition. Funding agencies should regulate how AI is used, require accountability for the resulting proposal, protect confidential information, and distinguish minor editorial assistance from substantive intellectual generation.

Why AI Use in Grant Proposals Is Controversial

A research proposal is not merely a writing exercise. It represents several claims:

• the applicants understand the scientific problem;
• the proposed methodology is technically defensible;
• the cited literature exists and is represented accurately;
• the team can perform the proposed work;
• the ideas attributed to the applicants are genuinely theirs;
• preliminary data, budgets, timelines, and institutional commitments are authentic.

Generative AI can help scientists communicate these claims. It can also produce persuasive text that conceals weak reasoning, nonexistent references, unrealistic methods, or uncertain authorship.

This creates a central distinction:

AI-assisted writing can improve the presentation of scientific judgment. AI-generated judgment can create the appearance of expertise without the underlying competence.

Grant policies should preserve that distinction.

Legitimate Uses of AI in Grant Writing

Scientists should be permitted to use AI for tasks that improve accessibility, clarity, and administrative efficiency without transferring responsibility for the research idea to the model.

Reasonable uses include:

Language Editing and Translation

Researchers who are not native English speakers frequently compete against applicants with access to professional editors, experienced grant offices, and institutional communications teams. AI can reduce this linguistic disadvantage by:

• correcting grammar;
• translating a draft;
• improving sentence structure;
• suggesting clearer terminology;
• shortening text to meet character limits.

UK Research and Innovation specifically identifies translation, improvement of English, formatting, and word-count reduction as examples of minimal AI use that applicants do not need to disclose. UKRI also notes that generative AI may reduce language barriers and support neurodivergent applicants.

Prohibiting these uses would not eliminate inequality. It would preserve the advantage enjoyed by wealthy universities that can employ human grant writers and editors.

Structural and Administrative Assistance

AI may also help applicants:

• convert funder instructions into a checklist;
• identify missing proposal sections;
• reorganize text according to application criteria;
• compare a draft with publicly available submission rules;
• prepare non-scientific summaries for general audiences;
• detect inconsistent terminology;
• generate preliminary project-management tables for human review.

These uses can reduce the substantial administrative burden associated with grant applications. The applicant must nevertheless check every output against the official call and the actual research plan.

Critical Review of an Existing Draft

A researcher may ask an AI system to identify:

• unsupported assertions;
• unclear transitions;
• unexplained technical terms;
• missing limitations;
• inconsistencies between aims and methods;
• passages that a non-specialist reviewer may misunderstand.

This is similar to asking a colleague to critique a draft. The researcher, however, must decide whether the criticism is valid.

Uses That Should Be Restricted or Prohibited

The acceptability of AI use changes when the model begins supplying substantive content that the applicant cannot independently defend.

Fabricated Citations, Results, or Credentials

No proposal should include AI-generated references that have not been verified against the original publications. The same applies to:

• invented preliminary data;
• false claims about prior work;
• nonexistent collaborators;
• fabricated quotations;
• inaccurate institutional resources;
• unsupported performance estimates.

The US National Institutes of Health warns that AI use can produce plagiarism, fabricated citations, and other forms of research misconduct. NIH states that applications substantially developed by AI—or containing sections substantially developed by AI—will not be treated as the applicants’ original ideas.

Generating an Application the Scientist Does Not Understand

A principal investigator should be able to explain and defend every major component of the proposal:

• the research question;
• the assumptions;
• the experimental or theoretical method;
• the statistical analysis;
• the risks;
• the interpretation of possible results.

Submitting technically impressive text that the applicant does not understand misrepresents the team’s capacity to perform the project.

This is especially serious when AI is used to produce specialist material outside the team’s competence. Fluent prose does not establish scientific validity.

Producing Entire Proposals at Industrial Scale

AI makes it inexpensive to generate many superficially different applications. This can flood review systems, increase reviewer workloads, and reduce the attention available for serious proposals.

NIH reported evidence that AI tools had enabled some principal investigators to submit more than 40 distinct applications during one submission round. In response, NIH introduced a limit of six specified application types per principal investigator per calendar year, while allowing only limited and appropriate AI assistance.

The underlying problem is not merely that AI wrote text. It is that automated proposal production can impose costs on reviewers and competing researchers while requiring little effort from the submitter.

Uploading Confidential Material to Public AI Systems

A grant proposal may contain:

• unpublished discoveries;
• patentable ideas;
• personal information;
• sensitive datasets;
• commercial plans;
• security-relevant research;
• confidential contributions from collaborators.

Applicants should not place such material into an external AI service unless the institution has verified the system’s privacy, retention, contractual, and security conditions.

UKRI states that confidentiality of information entered into generative AI tools is not guaranteed and prohibits entering another person’s sensitive or personal data without formal consent.

The US National Science Foundation similarly warns that information uploaded to unapproved public AI systems may leave the proposer’s control. NSF places particularly strict restrictions on reviewers because proposals contain confidential, non-public information.

AI Assistance Is Not Entirely New

Grant proposals have rarely been written by principal investigators alone. Depending on the institution, a proposal may be revised by:

• co-investigators;
• laboratory members;
• professional grant consultants;
• institutional research offices;
• statisticians;
• technical editors;
• translators;
• communications specialists.

It would be inconsistent to permit extensive human editorial assistance while prohibiting every use of AI.

The relevant question is not simply “Was AI involved?” It is:

Did the proposal accurately represent the applicants’ own scientific ideas, evidence, capabilities, and commitments?

A wealthy institution should not be allowed to purchase unlimited human proposal assistance while an independent researcher is penalized for using an inexpensive language tool. Regulation should focus on integrity, not on preserving institutional advantages.

Current Funding-Agency Approaches

Major funding agencies have not adopted one universal rule.

NIH: Limited Assistance, but Not Substantial AI Development

NIH allows that AI may assist with limited aspects of application preparation. However, it does not consider applications or sections substantially developed by AI to represent the applicants’ original ideas. The agency also emphasizes risks involving plagiarism, fabricated citations, and misconduct.

This is a relatively restrictive model. Its main unresolved difficulty is defining what “substantially developed” means in a consistent and enforceable way.

NSF: Applicant Accountability and Encouraged Disclosure

NSF encourages proposers to describe the extent and manner of generative-AI use. Regardless of whether AI was used, applicants remain responsible for the accuracy and authenticity of the submission. NSF connects this responsibility to existing standards concerning fabrication, falsification, and plagiarism.

This approach focuses more directly on responsibility for the final product.

UKRI: Permitted Use with Proportional Disclosure

UKRI distinguishes between minimal assistance and substantive use. Minimal functions such as translation, English improvement, formatting, and word-count reduction do not require disclosure. More substantive activities—including hypothesis development, literature comparison, data interpretation, code generation, and abstract generation—are expected to be declared.

UKRI also prohibits generating entire applications or sections without human involvement.

Of the three approaches, this is the clearest starting point for a practical policy because it recognizes that not all AI use has the same significance.

Why a Total Ban Would Fail

A complete prohibition would be difficult to justify and nearly impossible to enforce.

Detection Is Unreliable

AI-text detectors do not directly observe authorship. They estimate whether text resembles patterns associated with machine-generated language. Formal scientific prose, writing by non-native speakers, and heavily edited text may be incorrectly classified.

A funding decision should not depend on an opaque probability score produced by another AI system.

AI Is Becoming Part of Ordinary Software

Generative features are increasingly integrated into:

• word processors;
• grammar checkers;
• search engines;
• reference managers;
• coding environments;
• translation software;
• institutional research platforms.

A rule stating “no AI” becomes ambiguous when ordinary editing tools contain embedded machine-learning functions.

Bans Favor Institutions with Human Support

Elite universities can provide professional editing, administrative support, mock review panels, and experienced grant-development teams. Independent scientists and researchers from less wealthy institutions may have none of these resources.

Allowing controlled AI assistance can make proposal preparation more accessible. A ban may therefore reinforce, rather than reduce, structural inequality.

Why Unrestricted AI Use Would Also Fail

Unrestricted use would create different problems:

• proposal spam;
• fabricated references;
• homogenized applications;
• concealed lack of competence;
• leakage of confidential research;
• reduced confidence in authorship;
• strategic optimization for reviewer preferences;
• expansion of grant-writing competition without corresponding scientific output.

AI may also make mediocre proposals sound unusually polished. Reviewers could then spend more time separating rhetorical quality from scientific quality.

The result would be an arms race: applicants use AI to produce longer and more optimized proposals, while reviewers use AI to summarize them. The funding system would generate increasing quantities of machine-mediated text without necessarily improving scientific decisions.

A Practical Policy for Responsible AI Use

Funding agencies should adopt a tiered policy rather than a binary ban.

Level of AI use	Example	Recommended rule
Minimal editorial use	Grammar, spelling, formatting, translation	Permit without disclosure
Organizational assistance	Outlining, checklist creation, shortening	Permit; applicant verifies output
Analytical assistance	Literature comparison, methodological critique, code assistance	Permit with disclosure and verification
Substantive intellectual generation	Hypotheses, methods, interpretation, major narrative sections	Require detailed disclosure and direct human authorship
Deceptive or fabricated use	False citations, data, credentials, or commitments	Prohibit and treat under misconduct rules
Autonomous mass production	Large numbers of minimally supervised proposals	Restrict through submission limits or screening

Every proposal should also include a short certification:

“The applicants have reviewed and approved the complete application and accept responsibility for its accuracy, originality, citations, methods, data, and representations.”

For substantive AI use, an additional declaration could identify:

• the tool or model used;
• the relevant proposal sections;
• the purpose of its use;
• whether confidential material was processed;
• the human verification performed.

Funders do not need complete prompt logs in ordinary cases. Such a requirement would create excessive bureaucracy and could expose confidential reasoning or unpublished ideas. More detailed records could be requested only during an integrity investigation.

Reviewers Should Face Stricter Rules Than Applicants

AI use by reviewers presents a different risk profile.

Applicants are processing their own material and can decide what information to disclose to a tool. Reviewers receive confidential work belonging to other researchers. They normally lack permission to upload that material to an external service.

NSF prohibits reviewers from placing proposal content, review records, or related information into non-approved generative-AI tools. It treats such disclosure as a violation of confidentiality obligations.

UKRI likewise prohibits assessors from using generative AI to understand, summarize, or evaluate an application. Limited language refinement may be permitted only when no application content or personal information is entered into the system.

This stricter standard is justified. A reviewer should not disclose another scientist’s unpublished proposal merely to save time.

Secure, funder-controlled AI systems may eventually assist with conflict checks, administrative validation, citation verification, or proposal routing. But they should be audited, protected by appropriate data controls, and used to support—not secretly replace—expert judgment.

The Deeper Problem: Grant-Writing Skill Is Not Scientific Merit

The controversy exposes a weakness that existed before generative AI.

Traditional grant systems often reward a mixture of:

• scientific quality;
• institutional reputation;
• persuasive writing;
• familiarity with funder terminology;
• administrative capacity;
• prediction of reviewer preferences;
• access to preliminary resources.

These factors are not identical to scientific merit.

A brilliant researcher may be a weak proposal writer. A polished proposal may describe a project that produces little value. AI intensifies this mismatch because it makes polished language easier to obtain, but it did not create the mismatch.

The long-term solution is to reduce the amount of funding determined by speculative documents and increase funding based on verifiable research outputs, reproducibility, dependencies, public utility, and demonstrated contribution.

AI Internet-Meritocracy proposes one such model. AIIM is designed to evaluate published scientific and open-source contributions and allocate funding according to measurable merit and impact rather than primarily according to grant-writing performance.

This does not mean that all prospective grants should disappear. Expensive laboratories, clinical studies, field research, and infrastructure projects often require funding before results can exist. However, traditional grants can be complemented by continuous or retroactive scientific funding that rewards work after evidence of value emerges.

The comparison between AIIM and Horizon Europe illustrates the distinction: conventional grants attempt to predict future scientific value through proposals and committees, while AIIM aims to measure contribution dynamically through public outputs and dependencies.

Could AI Evaluate AI-Written Proposals?

Funding agencies may be tempted to answer AI-generated proposals with AI-generated reviews. That would be risky.

An AI system may help with narrow, auditable tasks such as:

• confirming whether cited papers exist;
• detecting inconsistent budgets;
• checking compliance with formatting rules;
• identifying duplicated passages;
• comparing claimed outputs with public records;
• routing applications to appropriate experts.

It should not independently determine which scientific ideas deserve funding without transparent criteria, appeal mechanisms, adversarial testing, and human or community oversight.

The danger is not “AI versus humans” in the abstract. Both humans and machines can introduce bias and error. The relevant questions are:

• What evidence does the system evaluate?
• Can its reasoning or scoring be audited?
• Who can contest a decision?
• Can applicants manipulate it?
• Does it measure scientific value or merely writing style?
• Who controls the model and its objectives?

An AI funding system should be judged by these governance properties, not by the mere presence of artificial intelligence.

Recommended Rule

Scientists should be allowed to use AI in grant proposals under five conditions:

1. Human ownership: The scientific ideas, decisions, and commitments must remain attributable to the applicant team.
2. Full responsibility: Applicants must verify every citation, factual statement, method, calculation, and claim.
3. Proportional disclosure: Minor language assistance need not be declared, but substantive intellectual or analytical use should be disclosed.
4. Confidentiality: Sensitive proposal material must not be uploaded to unapproved external systems.
5. No deception: AI must not be used to fabricate evidence, conceal lack of competence, impersonate collaborators, or mass-produce unsupervised applications.

Conclusion

Scientists should not be prohibited from using AI merely because it helps them write more clearly or navigate an unnecessarily burdensome application process. Such a ban would be difficult to enforce and would advantage institutions that can purchase equivalent human assistance.

But AI must not become a mechanism for manufacturing scientific authority, generating false evidence, or overwhelming review systems with automated proposals.

The defensible principle is straightforward:

Scientists may use AI to help express and examine their work, but they must not use it to misrepresent the origin, validity, or feasibility of that work.

Ultimately, the debate should encourage funders to reconsider why so much scientific money depends on proposal-writing performance in the first place. The more funding systems reward demonstrated contribution and verifiable utility, the less important it becomes whether the prose describing a future project was polished by a colleague, an editor, or an AI.