University AI policies must address data, not only cheating

A university can investigate a student who submits an AI-written essay without acknowledging the assistance. The assessment rules specify the potential offence, the evidence to review and the person authorised to reach a decision.

Now imagine a researcher uploading a confidential interview transcript to a commercial AI service for thematic analysis. The transcript includes testimony about political persecution and details that could identify participants. Who approved that transfer?

The researcher may have obtained ethics clearance to conduct the interviews, and the university may have licensed AI software. Neither fact necessarily establishes that this material can be processed by that service.

The answer may depend on participant consent, the provider’s terms, data protection law and the safeguards available under the relevant account. These considerations often fall under different offices, even though the upload itself takes only a few seconds.

Universities have devoted considerable attention to detecting and declaring AI assistance in student work. Their rules governing employees' use of AI with institutional data are less consistently aligned.

A lecturer can seek help drafting feedback, an administrator can summarise a student’s welfare complaint, or a supervisor can ask a chatbot to review an unpublished paper. Each action raises questions that an academic misconduct policy was never designed to address.

AI adoption has outpaced some safeguards

The United Kingdom's Cyber Security Breaches Survey 2025-26 offers a useful measure of the problem. Of 49 higher education institutions surveyed, 63% reported already adopting some AI tools. Among institutions using or considering AI, 49% reported having specific cybersecurity processes for AI-related risks.

These percentages have different denominators and should not be treated as a direct numerical gap. They nonetheless describe a sector where AI use is common and dedicated safeguards are far from universal. The survey is a UK snapshot, not a global estimate.

Senior leaders cannot simply be accused of indifference. Every higher education institution in the survey identified a board member, governor or senior manager responsible for cybersecurity. Eighty percent said they had reviewed risks associated with immediate suppliers, although only 37% had examined the wider supply chain.

The survey's qualitative interviews also described AI being used informally in universities to an extent that IT departments could not fully monitor. That last finding gets closer to the practical problem: responsibility exists at the top, but the institution may not see the everyday decisions that expose information (UK survey).

Cambridge and Oxford have taken different approaches to the problem. The University of Cambridge advises staff performing administrative tasks to avoid entering confidential, sensitive or personal information into generative AI unless the use is justified and the information is protected.

It favours licensed institutional services over unlicensed versions and calls for formal privacy or information security assessments when appropriate. Cambridge explicitly limits this guidance to administrative use, leaving research and direct teaching to other processes.

Oxford’s research policy focuses more directly on scholarly work.

It addresses unpublished intellectual property, confidential research data and the risks associated with cloud-hosted services and recommends enterprise or locally deployed tools when stronger assurances are needed.

It also states that uploading personal data as part of research constitutes processing that must comply with data protection law and requires ethics approval.

These examples raise the question university executives should be asking: how do separate rules for research, teaching, administration and IT become a single decision staff can follow?

A guidance page can be accurate yet still leave a researcher uncertain about who can authorise an exceptional use or who must be consulted when an approved tool is connected to a new database.

Research consent does not settle AI use

Research ethics approval typically sets the conditions for collecting, storing and analysing participant information. These conditions may require reassessment when researchers introduce a new processing arrangement, particularly one involving an external AI provider.

A researcher may be authorised to hold an interview transcript without being authorised to send it to a new processor under a different contract.

Removing names offers limited protection when a participant is recognisable through a rare experience, a job title, or the events described. A small qualitative study can make this especially difficult. Translation, thematic coding and summarisation may still expose personal information to an external provider.

The risk is particularly acute for testimonies about political violence, displacement, health or abuse, where an apparently minor detail could identify a person to someone familiar with the setting.

Consider a research ethics committee that approved interviews on the understanding that recordings and transcripts would remain in an institutional repository. The investigator later discovers an AI tool that promises faster analysis.

Does the original approval extend to a third-party service? Should the committee review the change, or should the principal investigator and the procurement office decide? Universities need a procedure to resolve that question before uploading the material, including a route to amend approval when the processing environment changes.

The same issue arises in scholarly publishing. A reviewer might upload a confidential manuscript to obtain a quick summary or critique.

The International Committee of Medical Journal Editors warns that manuscripts submitted for review are confidential communications. Editors and reviewers should not upload them to AI systems unless confidentiality can be assured and only with the authors’ explicit permission.

The difficulty is deciding where a researcher's discretion ends and the university's responsibility begins. The investigator decides how to analyse evidence; the university must establish the conditions under which others' information may be used. Academic discretion remains important, but it cannot replace an agreed account of disclosure and consent.

Suppliers may not be the only ones handling data

An October 2026 disclosure by the University of Osaka underscores why universities need to look beyond the company named in a contract. On 21 October, the university reported that personal information for 77,851 people had been leaked through arrangements with NEC, a contractor operating the university's computer system.

According to the university, an unreported subcontractor was involved; files were shared via cloud storage without a contractual basis; and copies that should have been deleted were retained. The university reported no confirmed misuse at the time of its notice. This was a conventional supplier-related incident, not an AI breach.

The Osaka case illustrates the exposure that arises when a university cannot account for subcontracting, onwards transfers and deletion. A commercial AI product may involve hosting companies, external integrations and connected repositories, in addition to the provider visible to the user.

Contracts and technical controls determine which parties can access which information, for what purpose and for how long. Brand familiarity can’t infer any of this.

Not all AI products are governed by the same conditions. Consumer accounts, university-contracted enterprise services, and locally operated models may differ substantially.

OpenAI, for example, states that it does not use inputs and outputs from specified business and education offerings to train its models by default and describes separate retention controls for qualifying organisations.

A no-training commitment answers an important question, but it does not, by itself, establish how long it retains content, who may access it under a particular integration, or how it will delete records. Institutions need to examine the specific service and contractual arrangement they are purchasing.

Procurement should therefore test a scenario, not merely tick a security box. If a research team uploads restricted data to a licensed AI service, can the university identify the processors involved? Can it honour a participant's withdrawal request when legally and technically applicable? Can it retrieve records and analytical outputs if the contract ends?

A supplier’s answer may determine whether the proposed academic use is permissible.

Staff use of AI deserves scrutiny, too

The most consequential university use of AI may be difficult for a student to observe.

An adviser drafting a reply based on a student’s financial hardship statement could expose information the student shared in confidence.

A lecturer generating feedback on unmarked essays might inadvertently reveal identifiable student work. An admissions office might use AI to rank applicants whose records were originally collected for human assessment, without informing them that automated profiling would be part of the selection process.

These examples are hypothetical, but each raises questions about purpose, notice and authority.

The problem extends to judgements drawn from the data. A system might infer that a student is at risk of failure or that an applicant has specific personal characteristics. Such inferences can shape decisions even when the original records have never been disclosed.

The UK's Information Commissioner's Office has identified repurposing, novel inferences and diminished transparency as data protection risks associated with agentic AI. That analysis flags regulatory concerns; it should not be mistaken for evidence that every university has adopted such systems.

When AI materially influences decisions about admission, assessment or student support, affected individuals should be informed of its role and have a meaningful opportunity to challenge any errors. The exact legal rights will depend on the jurisdiction and the decision.

Who has authority to approve the upload?

The first task for university leaders is to define permissions. Published course descriptions and public research abstracts differ from unpublished examination materials, identifiable student records and confidential interviews.

A useful policy would classify information by sensitivity and intended use, then specify the tools and accounts approved for each category. For some material, de-identification and a controlled environment may suffice. For other material, external transfer should be prohibited or require a fresh ethical or legal review.

Classifying information and approving AI tools are only the starting points. More difficult questions arise when proposed uses fall outside established permissions or when academic, legal and technical offices have overlapping responsibilities.

Research ethics committees can assess obligations to participants, data protection officers can assess lawful processing and security teams can evaluate technical risks. Procurement teams can negotiate supplier obligations. When those assessments conflict, who makes the final decision?

Leaving that decision to whoever receives the request invites inconsistent practice. Major AI uses involving restricted data need a designated academic or administrative owner with a documented escalation and refusal process.

University senates and academic boards also have a legitimate role. A service that meets IT security standards may still be unacceptable for confidential peer review or incompatible with the conditions promised to research participants.

Likewise, security specialists must be able to contain an active incident without waiting for a prolonged academic debate. Policies should define both responsibilities rather than asking one group to stand in for the other.

A practical test is whether a staff member can answer three questions before submitting material: Is this data permitted in this tool? What approvals or notices are required? Who will help if the answer is unclear?

Those answers should be available within standard workflows, not scattered across separate webpages for ethics, procurement and information security. An inventory of approved services is useful only when it also clarifies the conditions of use.

International guidance can support that work.

UNESCO’s Guidance for Generative AI in education and research calls for protecting data privacy and establishing coherent institutional frameworks.

The US National Institute of Standards and Technology’s Generative Artificial Intelligence Profile places privacy and intellectual property alongside organisational risk management. Neither document determines who may upload a particular transcript at a particular university; that responsibility remains local.

Restrictions will fail if the approved route is so cumbersome that staff resort to personal accounts. Universities must provide usable alternatives, including licensed services, secure research environments and case-based training. Smaller institutions may need shared procurement arrangements or sector-wide guidance to achieve comparable protection.

AI literacy for university employees should cover the consequences of uploading data, not only the accuracy of the answers they receive.

Accountability before the first prompt

Academic integrity rules have helped universities clarify what students must disclose about AI-assisted work. Institutional accountability now requires a comparable explanation of what professors, advisers and administrators may disclose to AI systems. Those whose records are involved have a stake in those decisions, whether or not they ever use a chatbot.

Before an interview transcript, student welfare record or unpublished manuscript enters an AI service, the university should know who is authorised to send it, which protections apply and who is responsible if those conditions are breached.

Such decisions require institutional oversight, with academic, legal and technical responsibilities clearly assigned. They cannot be left to the default settings of a software account.

James Yoonil Auh is a former vice-president of the National Labor College in Maryland, United States, and former vice-president of Kyung Hee Cyber University in Seoul. He currently teaches at Judson University in Illinois, US. His work examines how technological and social change reshape institutions, authority and human agency.

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