Responsible AI Use in Academic Research: A Framework for Faculty, Librarians, and Writing Centers

This is the long-form piece we would want every writing-center handbook and every departmental AI-use policy to draw from. It is written for faculty, librarians, and administrators — the people who set the practices that students then follow. Students and researchers who use AI for their own work will also find it useful; the same four principles apply from both sides of the institutional relationship.

It is also the piece where CiteDash explicitly stakes out a position. We have spent two years building a retrieval-first academic AI system and running what is becoming the field's main citation-hallucination benchmark. Our view, earned through that work, is that responsible AI use in academic research is genuinely achievable — but it requires institutional practices that match what the technology can and cannot reliably do. This piece is our attempt to write those practices down, in a form that can be reused.

The Four Principles

The framework that holds up across every institutional policy we have reviewed — and across every disciplinary tradition — reduces to four principles.

Transparency

Disclose what AI tools you used and how. The default is to disclose; the exception is to not disclose. If a use case is so minor that disclosure feels unnecessary, the disclosure can be a single sentence in an author's note. If a use case is significant, disclosure belongs in the methods section or a dedicated statement.

The point of transparency is not to embarrass researchers who used AI. It is to make the research reproducible, to let readers evaluate the work in context, and to build a public record of how AI was actually used across the field. Over-disclosure has almost no downside. Under-disclosure can kill a reviewer's trust in an otherwise solid paper.

Verification

Every load-bearing claim and every citation produced or surfaced by AI must be independently verified before it enters the work. The burden is on the researcher, not the tool. "I trusted the AI" is not a defence when a fabricated citation is caught in peer review or by an instructor.

The specific verification workflow — check DOIs against CrossRef, verify claims against primary sources, confirm quotations exist in the cited paper — is covered in our AI hallucination detection workflow. The principle here is simpler: verification is not optional. It is a precondition for AI use in research.

Attribution

When AI summarises, paraphrases, or surfaces work produced by others, cite the underlying work, not the AI. The AI is a tool, like a library catalogue; it is not a source. Citing AI as a source — or worse, using AI output verbatim without attribution to the real source — misrepresents both the AI and the underlying scholarship.

For verbatim AI output that does not summarise a specific source (for example, AI-generated connective tissue between sections), the AI is cited directly. See how to cite ChatGPT, Claude, and Gemini for the specific formats in APA, MLA, and Chicago.

Oversight

The human researcher bears full responsibility for the final work. If the submitted paper contains a fabricated citation, the researcher is responsible for that fabrication — regardless of which tool produced it. If the argument is wrong, the researcher is responsible for the argument. If the work is superficial, the researcher is responsible for the superficiality.

This principle is the one that often gets lost in discussions of AI ethics. The question is not "who is morally culpable for AI errors" — which leads into tangled arguments about tool intent and emergent behaviour. The question is "who is institutionally accountable for the submitted work," and the answer has been the same since the invention of the academy: the person whose name is on it.

A practice that respects all four principles — transparency, verification, attribution, oversight — is defensible under every institutional policy we are aware of. A practice that violates any one of them is fragile, regardless of which specific tools were or were not involved.

When AI Use Is Appropriate

Applying the four principles leads to a clear distinction between tasks where AI use is routinely appropriate and tasks where it is not.

Appropriate use cases

• Literature scoping. Using AI to identify candidate sources on a topic is a routine aid, analogous to a research librarian pointing you toward relevant databases. Retrieval-first tools (CiteDash's deep research, Elicit, Consensus) are well-suited here because they search real databases and return real papers. The verification step is still required — read the papers before citing them — but the accelerator is legitimate.
• Summarisation of read material. After you have read a paper, using AI to produce a structured summary of it for your notes is a productivity aid. The researcher has the source material and can verify the summary. The work does not depend on the summary being AI-generated.
• Drafting support. Using AI to help with outline, paragraph organisation, or initial drafting is defensible if the final work is substantively shaped by the researcher's own judgment. The test: could the researcher, asked to defend any paragraph in the paper, do so? If yes, the AI contribution is organisational; if no, the AI contribution is substantive and raises oversight concerns.
• Language and style. Non-native English speakers, and in fact any writer, legitimately use AI for grammar, style, and clarity adjustments. This is a modern equivalent of copy-editing and does not generally require disclosure.
• Brainstorming and ideation. Using AI as an interlocutor for thinking through a problem is legitimate. The researcher uses the brainstorming as input to their own reasoning.
• Translation. Using AI to translate source material into a working language, with subsequent verification, is legitimate and standard.

Uses that require caution

• Original synthesis. When a paper makes an original synthesis of a field — the kind of argument that is supposed to be the researcher's own contribution — AI-drafted synthesis sits in a more fraught place. This is not automatically inappropriate, but it does raise oversight concerns. If the argument of the paper is the AI's argument, the paper is not really the researcher's work in the way the academy traditionally means.
• Data analysis. AI-assisted data analysis is a growing practice with specific pitfalls: AI can reproduce statistical errors, mis-apply tests, and confidently produce wrong conclusions. Verification (reproducing the analysis independently, checking the code, confirming the test is appropriate) is essential.
• Peer review. Using AI to generate a peer review of a manuscript is an ethical grey area that is trending toward prohibition. The reviewer's judgment is the value the journal is procuring; AI-generated review subverts that relationship. Several major publishers have prohibited or restricted AI use in peer review as of 2025.

Inappropriate uses

• Ghost-writing. Having AI write the substantive content of a paper and submitting it as one's own work is a category violation. This is not a case of "AI use without disclosure"; it is a case of "no scholarship." The disclosure issue is downstream of the integrity issue.
• Citation fabrication. Submitting AI-produced citations that have not been verified is irresponsible regardless of whether any specific citation turns out to be fabricated. The verification principle forbids this.
• Claim fabrication. Similar to citation fabrication but for claims. Any substantive claim in the paper needs to be supportable from a real source the researcher has engaged with.

The boundary between "original synthesis" and "drafting support" is where most institutional policy disputes happen. A helpful operational test: if the AI were removed from the process entirely, would the resulting work still be the researcher's own contribution, recognisably, at the right level of quality? If yes, the AI contribution is assistance. If the work would be substantially worse or would not exist, the AI contribution is substantive and is likely to raise integrity questions.

Disclosure Language for Methods Sections

Ready-to-use templates for several common patterns. Adapt as appropriate to your specific case and institutional policy.

Template 1: Mechanical AI use only

AI assistance in the preparation of this manuscript was limited to grammar and style checking via [tool name, version]. No AI tools contributed substantive content, research, or analysis. All content was drafted, researched, and verified by the author.

Template 2: Literature scoping with AI

Initial literature scoping for this review used [retrieval-first tool, version], which searches peer-reviewed databases and returns real papers with verifiable metadata. All papers identified by the tool were independently reviewed by the author against the primary source. ChatGPT-4o (OpenAI, 2026) was used to support outlining during the writing phase. All final prose was drafted by the author, and all cited sources were independently verified via CrossRef and Semantic Scholar.

Template 3: AI drafting support

This paper was prepared with drafting support from [tool name, version]. AI-generated drafts were used as starting material for the author's revision and elaboration. No AI-generated text appears verbatim in the final manuscript; all substantive claims and interpretations are the author's. All citations were independently verified via CrossRef and Semantic Scholar.

Template 4: AI summary of read material

Structured summaries of reviewed papers were generated with assistance from [tool name, version] and verified by the author against the primary sources before inclusion in the manuscript's analysis. The summaries were used as internal notes; no summary text appears verbatim in the final manuscript.

Template 5: AI-assisted data analysis

Data analysis for this study was performed using [analysis tools], with AI assistance from [tool name] for code generation and interpretation of results. All statistical methods were reviewed by the author for appropriateness. Key results were independently reproduced using [independent method]. The author takes full responsibility for the analytical choices and conclusions.

Template 6: Full AI drafting with human oversight (most cautious case)

This manuscript was drafted in close collaboration with [tool name, version]. The author provided the research question, source materials, and overall argumentative structure; the AI produced initial drafts of most sections. All drafts were substantively revised by the author, all claims were verified against primary sources, and all citations were independently confirmed. The author accepts full responsibility for the accuracy and interpretation of the final work.

These templates assume good-faith use. Researchers who are uncertain whether their use falls within one of these patterns should consult their department's integrity officer or their graduate advisor.

Teaching Students AI Literacy

The instructional side of responsible AI use is, in our view, where most institutional effort should be concentrated. A policy is a document; literacy is what students carry with them into their careers.

Three literacies to teach

Technical literacy. Students need a basic working model of what an LLM actually does: token prediction, training data, the absence of a fact database, the reason fabrication happens. They do not need to become ML researchers. They do need to understand why ChatGPT fabricates citations, because once they understand the architectural reason, they stop treating AI output as authoritative by default. Our piece on why ChatGPT makes up citations is pitched at students who need this foundation.

Workflow literacy. Students need a practical verification workflow that they can run in under two minutes per claim. Checking a DOI, searching a title in Semantic Scholar, confirming a claim against the primary source — these are mechanical skills, easy to teach, directly applicable. Our hallucination detection workflow is the version we recommend teaching from.

Ethical literacy. The hardest of the three. Students need a framework for deciding when AI use is appropriate for a specific task. Rote rules ("no AI in research papers" or "AI allowed for grammar only") do not transfer across tasks or across careers. The four-principles framework in this piece is one option. Institutions will develop others. The point is that ethical judgment about tool use is itself a skill, and one that generalises beyond any specific AI tool.

Where to teach it

The short answer is: early, often, and across the curriculum. AI literacy taught only in one course does not stick. AI literacy taught in a one-off orientation session does not stick either. What works is integration across the curriculum:

• First-year composition: basic verification workflow on AI output.
• Research-methods courses: discipline-specific patterns of appropriate and inappropriate use.
• Discipline-specific courses: how AI is being used in the field, what the specific integrity risks are.
• Thesis and capstone courses: the oversight principle in depth, including how to defend AI-assisted work in a thesis defence.

Libraries are the natural integrators of this material because they already run information-literacy instruction across the curriculum. Writing centres are the natural partners for workflow literacy. Teaching-and-learning centres are the natural partners for pedagogical design.

Institutional Policy Templates

Below are template sections that departments can adapt for formal AI-use policies. They are not a substitute for consulting your institution's legal and policy offices, but they are a starting point that reflects current best practice.

Permitted uses (template)

The following AI uses are permitted in coursework and research in [department] without special disclosure:

— Grammar, style, and clarity checking (e.g., Grammarly, ChatGPT for editing passes)

— Retrieval-first academic search tools that return real, verifiable sources (e.g., Semantic Scholar, Elicit, CiteDash, Consensus). Sources must be independently verified and cited directly.

— Translation, with subsequent verification.

— Summarisation of material the student has themselves read, for note-taking purposes.

— Brainstorming and ideation support where the resulting ideas are the student's own to develop.

Permitted with disclosure (template)

The following AI uses are permitted with a disclosure statement in the methods section or author's note:

— AI-supported outlining or drafting, where the final prose is substantively the student's work.

— Structured summarisation of read material where the AI-generated summary will be referenced in later analysis.

— Data analysis assistance, including code generation, with verification.

— Any AI use whose contribution is substantive enough that a reasonable reader would want to know about it.

Prohibited uses (template)

The following AI uses are not permitted:

— Submitting AI-generated text as one's own work without substantive human shaping and verification.

— Including citations in a submitted paper that have not been independently verified against a primary source.

— Using AI to generate peer reviews or other confidential evaluative material.

— Using AI to bypass process requirements (e.g., submitting AI-generated drafts when drafts are supposed to reflect the student's own work-in-progress).

Disclosure requirements (template)

All submitted work that used AI tools beyond the permitted-without-disclosure list must include a disclosure statement in the methods section or a dedicated author's note. The disclosure must include:

— The name and version of each AI tool used.

— The specific tasks for which the tool was used.

— The verification steps applied to AI-generated output before inclusion in the final work.

Disclosure statements must be written in plain language and must be specific enough that a reviewer could evaluate the degree of AI involvement.

Assessment consequences (template)

AI use that exceeds the permitted-without-disclosure list and is not disclosed will be treated as an academic integrity violation under [specific policy reference]. AI use that is disclosed will be evaluated on the merits: disclosed mechanical use will not affect grading; disclosed substantive use may affect grading depending on the assignment's pedagogical goals, as specified in the course syllabus.

These templates should be read, revised, and ratified by departmental faculty. A policy that faculty have not discussed and agreed to is a policy that will not be applied consistently.

The Role of Academic Librarians

We think academic librarians are the most underutilised institutional asset in responsible AI-use work. Three reasons.

Existing information-literacy infrastructure

Libraries already run information-literacy instruction — how to evaluate sources, distinguish peer-reviewed work, search databases effectively. AI literacy is a natural extension of this existing infrastructure, not a parallel track. Adding AI-specific modules to existing library instruction is much more efficient than building a separate AI-literacy programme from scratch.

Trusted relationships across disciplines

Subject librarians already work with faculty across every discipline in the institution. They know which departments have strict AI policies, which are more permissive, which faculty are willing to experiment, and which are conservative. That cross-disciplinary network is exactly what institutional AI-literacy work needs, and it exists nowhere else on campus.

Expertise in source verification

Librarians are professional source-verifiers. The verification workflow for AI output — checking DOIs, confirming claims against primary sources, distinguishing reliable from unreliable metadata — is an extension of what librarians already do every day. The teaching of this workflow fits naturally into librarian skillsets and is often a better pedagogical fit than having faculty do it.

Institutions that want to invest in AI literacy should start by funding their libraries to build AI-specific instructional programmes, and by creating formal partnerships between libraries, writing centres, and teaching-and-learning centres. This is where the most durable institutional progress is happening in 2026.

CiteDash's Stance and Why We Built Hallucination Detection

A transparent note about our own position.

CiteDash is an academic AI platform. We build retrieval-first tools — deep research, citation verification, literature review support — and we compete with general-purpose chatbots in academic settings. We have a commercial interest in institutions adopting responsible AI use, because responsible AI use favours tools built for that purpose.

We also have a direct interest in the citation hallucination benchmark we are running: tools like ours perform well on it, and tools built on different architectures perform badly. We are aware of the self-promotional risk of running a benchmark that our own tool wins, and we have committed to open data, external scoring, and OSF preregistration specifically to address that risk.

Our view — stated here so readers can weigh it with appropriate scepticism — is that the architectural distinction between retrieval-first academic tools and general-purpose chatbots is real, large, and not going to be closed by better chatbots alone. The responsible-use framework above is designed to reflect that architectural reality: it does not privilege any specific tool, but it does require practices (verification, attribution, oversight) that are much easier to implement with retrieval-first tools than with token-prediction chatbots.

We built our hallucination-detection pipeline because we think the commercial AI market, left to itself, was going to produce a lot of academic-integrity disasters before building the kind of verification layer the technology actually needs. That view has been validated by the last three years of published research on fabrication rates, detector false positives, and institutional struggles to write defensible AI policies.

The framework in this piece is the framework we hope becomes standard regardless of which specific tools win. The principles — transparency, verification, attribution, oversight — do not depend on CiteDash existing. What CiteDash offers is an implementation of those principles as an integrated product. Other implementations are possible and welcome. What the academy needs is the principles themselves, applied consistently, across institutions.

Final Thoughts

Responsible AI use in academic research is not a radical proposal. It is the straightforward application of four principles that academic research has always operated on — be transparent, verify your claims, attribute your sources, take responsibility for your work — to a new category of tool.

The difficulty is implementation. The specific workflows, the specific policies, the specific instructional modules, the specific cultural practices that make these principles durable across a large institution. That implementation is the work of the next few years, and it is happening already in scattered pockets across higher education.

If you are a faculty member: apply the principles to your own work first, then to your teaching, then to your department's policy discussions. A policy without practice is decorative.

If you are a librarian: your profession is the natural home of this work. Claim the territory. Partner with writing centres and teaching-and-learning centres. Build the instructional programmes that do not exist yet.

If you are an administrator: fund the library, fund the writing centre, fund the teaching-and-learning centre. Adopt the four-principles framework as the basis for institutional policy. Resist the temptation to rely on automated detection tools as a shortcut; they do not work well enough to justify the risk.

And if you are a student: the four principles are yours to uphold personally. Transparency means you say what you did. Verification means you check what you wrote. Attribution means you credit who you drew from. Oversight means you own the final work. That is the whole job. Everything else in AI ethics, at a personal level, is commentary.

Academic research has always required practices of honesty that went beyond what any external policy could enforce. That tradition is what we are continuing. The tools change; the practices do not.