Universities must look beyond plagiarism to govern AI research

Universities have spent much of the past three years asking whether researchers and doctoral students may use generative AI.

That was a necessary first question. But AI is already being used for literature searches, coding, drafting, transcription and significant data work. The practical question is now whether those uses produce research that remains ethical, valid, reproducible and transparent.

Most university responses to AI do not yet answer that question. They concentrate on plagiarism, authorship and disclosure.

They sit at the surface of a deeper research integrity problem. A disclosed AI-assisted analysis can still be invalid. A carefully attributed literature review can still contain fabricated citations. A secure enterprise tool can protect confidential data while producing an inference that no researcher can reproduce or defend scientifically.

This was the problem behind my open-access Instats policy report, Responsible AI in Academic Research: A competency framework for research training.

The research examined public material from 38 top-tier doctoral universities across 15 countries and jurisdictions, 14 national research funders, 18 major publishers and three preprint servers, together with 13 AI literacy and researcher development frameworks. Every substantive claim was traced to a public policy, regulatory instrument or research source current to May 2026.

Universities have responded slowly to AI

The response gap was hard to miss.

Major academic publishers converged within 10 weeks of ChatGPT’s public release on the rule that AI cannot be an author because it cannot take responsibility for scholarly work. Research funders responded more slowly and unevenly, mainly in grant applications and assessment.

Universities, which train the researchers who must make these decisions every day, have moved slowest.

Only six of the 38 universities examined had policies that extended beyond research integrity and disclosure into AI literacy, supervision, examination and valid research practice.

Fifteen remained at the level of a generic student conduct rule or an AI-specific plagiarism policy. Seventeen had moved into research integrity and reproducibility, but without codified expectations for researcher competence.

The six institutions in the most developed group were University College London and King’s College London in the United Kingdom, Heidelberg University in Germany, KU Leuven in Belgium, the University of Helsinki in Finland, and Tsinghua University in China.

Germany had the most consistently mature cluster, helped by its tradition of sworn doctoral declarations. The United Kingdom was divided, with UCL and King’s College London at the leading edge while other institutions relied on devolved faculty guidance.

Australia’s leading universities had generally incorporated AI into research integrity procedures, but none had codified AI literacy as a doctoral competency. The central policies of the United States institutions in the sample were the least developed.

Conceptual lag

The reason for this lag is partly conceptual. Large language models mimic human writing, so universities have treated their use mainly as an authorship and plagiarism problem. Yet their underlying research function is closer to that of other task-directed tools or 'agents' in the modern AI vernacular.

No one treats output from a statistical package as plagiarism. A machine transcript is not fraudulent if a researcher checks it against the audio. The relevant question is not simply who produced the words. It is whether the tool was used in a way that preserves the standards on which research depends.

In the report, responsible AI use therefore means uses that are ethical, valid, reproducible and transparent.

Ethical use protects participants, communities and third parties whose data may be processed. Valid use meets the same methodological and evidentiary standards expected of human-produced work.

Reproducible use records the inputs, prompts, parameters, model version and relevant settings well enough for another researcher to repeat the process to a stated tolerance. Transparent use discloses AI’s role with enough specificity for supervisors, thesis/dissertation examiners, reviewers and readers to assess it.

None of these four conditions can substitute for another.

A five-dimension framework

The five-dimension framework translates this definition into institutional practice.

The first dimension is human-in-the-loop discipline. Universities need a clear distinction between research labour that AI may accelerate and research judgement that must remain human.

Formatting, verified transcription and code cleanup may be delegated with checks. Framing a research question, choosing a method, interpreting an unexpected result, deciding what counts as an outlier and weighing conflicting evidence cannot be outsourced without weakening the research itself.

The second dimension specifies responsible practice across four use modes. AI used for search should help scope a literature, but every citation that survives must be checked against the primary source.

AI used as a co-authoring tool may polish or reorganise text that a researcher can independently produce, but it should not substitute for knowledge the researcher does not possess.

AI used as a validator can generate criticism, but its agreement is not evidence and its critique does not complete human verification. AI used as a tutor should be grounded in trusted materials, with researchers still required to demonstrate independent understanding.

The third dimension concerns tooling. Universities should not leave the choice of AI systems entirely to individual laboratories.

Procurement should test whether tools provide verifiable citations, suitable data residency and training controls, meaningful uncertainty information, reproducibility at a known model version and auditable session records.

Responsible use is easier when technology supports it by design rather than relying on warnings that users are expected to remember.

The fourth dimension is AI-literate people. This is not mainly a matter of teaching researchers which buttons to press. Doctoral training now needs to cover citation verification, model and parameter reporting, prompt sensitivity, the use of different model families for adversarial review, detection of sycophantic agreement and structured reporting of failure modes.

These competencies should extend the existing curriculum on validity, reproducibility and scepticism, not replace it. Supervisors and examiners need the same development because candidates cannot demonstrate disciplines that their training environment does not model.

The fifth dimension is institutional benchmarking. Policy is only one part of competency. Universities also need named people responsible for implementation, systems that support responsible choices and processes that bring AI use into supervision, progression reviews, thesis declarations and examinations.

The framework scores each dimension from absent to nascent, established and leading. It is a self-diagnostic grid, not a league table. Its value lies in showing where an institution is constrained and what a credible next step looks like.

What ought university leaders to do?

For senior leaders, the starting point is practical. Appoint an owner for AI in research and research training. Assess the institution against the five dimensions. Publish the distinction between tasks that AI may support and judgements that remain human.

Integrate AI competence into doctoral methods training. Train supervisors and examiners. Put a research-specific procurement gate around the tools researchers actually use. Then review progress at a stated cadence and make the results visible to the university community.

Universities are not passive consumers of AI. They are part of the world’s epistemic infrastructure. They train the people who produce evidence for decisions about health, climate, economic policy and technology.

If AI-supported research is not ethical, valid, reproducible and transparent, the public record absorbs fabricated citations, biased analyses and findings that cannot be checked.

The answer is not to prohibit useful tools or to produce another isolated policy document. It is to build the judgement, systems and accountability needed to use them well.

Publishers have shown that an academic sector can move quickly when it recognises a common problem. Universities now need to act with similar urgency, but with a more ambitious goal: excellent AI-enabled research through excellent research training.

Michael J Zyphur is director of Instats (instats.org) and professor of quantitative methods at the University of Queensland.

He is the author of the open-access report Responsible AI in Academic Research: A competency framework for research training, published earlier this year. Instats provides research training, including AI-related training. The framework was developed independently of commercial activity and is offered as an open-access contribution to discussion about responsible AI in research and research training.

AI disclosure: OpenAI GPT-5.6 (sol xhigh) was used to support the writing and editing of this commentary. The author determined the argument, checked the evidence and source interpretations, approved the final wording and took full responsibility for the article.

This article is a commentary. Commentary articles are the opinions of the author and do not necessarily reflect the views of University World News.