03 · What You Need to Know
Understanding Confidentiality Risks in AI-Assisted Research Supervision
Why an Unpublished Thesis May Be Confidential
A thesis draft is not necessarily confidential simply because it has not yet been published. Nevertheless, unpublished research may contain information subject to specific restrictions.
Examples include identifiable participant responses, confidential institutional records, unpublished collaborative findings, commercially sensitive information, and material governed by research agreements.
Some information may also carry intellectual property interests or restrictions imposed by a research sponsor.
Supervisors should therefore evaluate the actual contents and obligations associated with the document rather than assume that all thesis drafts have identical confidentiality requirements.
Research involving publicly available information may present different risks from research involving identifiable participants.
Access to a Manuscript Is Not the Same as Authorization to Upload It
Supervisors commonly receive access to student manuscripts because they have responsibilities for academic guidance and assessment.
That access does not automatically establish authority to transfer the material to an external service.
For example, a supervisor may be authorized to read confidential interview excerpts through an institutional learning platform while being prohibited from submitting the same excerpts to an unapproved AI provider.
The distinction concerns the purpose, recipient, and conditions of processing.
Before using AI, supervisors should establish whether the proposed activity is covered by the institution's technology, data-governance, and research ethics arrangements.
What Happens When Research Material Is Submitted to an AI Service?
The precise processing arrangement depends on the service and its configuration.
Information may be transmitted to provider-operated infrastructure, processed through connected services, retained for specified periods, or made accessible to authorized personnel under particular conditions.
Some services offer contractual restrictions on data use, administrative controls, or deployment arrangements that differ from ordinary consumer accounts.
These differences should not be inferred from a product's name or marketing description alone.
Supervisors should consult the applicable service documentation and institutional agreement to determine what happens to submitted material.
The relevant question is not simply whether the tool is described as secure, but whether its actual processing conditions satisfy the research's obligations.
Classify the Research Information Before Choosing a Tool
Data classification helps supervisors determine which safeguards are appropriate.
| Information category |
Illustrative material |
Primary consideration |
| Public research material |
Published articles and publicly released reports |
Accuracy, copyright, and applicable tool conditions |
| Unpublished scholarly material |
Draft thesis chapters and preliminary findings |
Confidentiality, ownership, and authorized disclosure |
| Internal institutional information |
Restricted administrative or academic records |
Institutional processing authorization |
| Participant-related information |
Interview transcripts and research datasets |
Consent, ethics approval, privacy, and re-identification |
| Highly sensitive research material |
Detailed health records or stigmatizing personal accounts |
Enhanced safeguards and potentially restricted processing |
These categories are illustrative. Actual classifications should follow the institution's established information-governance framework.
A single manuscript may contain information belonging to several categories, and the most restrictive applicable obligations may determine which processing arrangements are permissible.
Why De-Identification May Not Be Enough
Removing names, student numbers, and direct identifiers can reduce privacy risks.
However, participants may remain identifiable through combinations of contextual details, unusual experiences, institutional roles, or demographic characteristics.
For example, an interview excerpt describing the only department chair responsible for a particular institutional initiative may identify the participant without including their name.
Supervisors should therefore distinguish removing direct identifiers from demonstrating that information can no longer reasonably be linked to individuals.
The adequacy of de-identification depends on the data, available contextual information, and intended processing environment.
It should not be treated as automatic permission for external AI processing.
Research Ethics Approval May Restrict AI Processing
Research ethics approvals may establish how participant information can be collected, stored, accessed, analyzed, and shared.
A proposed AI-assisted review may introduce a processing arrangement not considered when the study was approved.
Whether an amendment or additional authorization is required depends on the approved protocol and institutional procedures.
For example, a protocol may authorize analysis by the named research team using approved institutional systems without addressing an external generative AI provider.
The supervisor should consult the responsible research ethics authority before introducing processing that may materially change the approved arrangements.
Approval for the original research does not necessarily cover every subsequent technological method.
Does Student Consent Resolve Confidentiality Concerns?
Student agreement may be appropriate when their unpublished work will be processed through AI.
However, students may not possess authority to authorize disclosure of every element within a manuscript.
A thesis may contain participant data, collaborative intellectual contributions, or institutional information governed by separate obligations.
For example, a student cannot necessarily authorize external processing of confidential interview transcripts merely by agreeing that the supervisor may use AI.
Supervisors should distinguish the student's preference from the permissions required by research ethics, data protection, and collaboration arrangements.
Where the student has a meaningful choice, the supervisory power relationship should also be considered when seeking agreement.
How Should Supervisors Evaluate AI Provider Policies?
Supervisors should examine the actual terms and institutional arrangements applicable to the proposed account or deployment.
Relevant considerations include permitted data uses, retention periods, access controls, subprocessors, geographic processing conditions where relevant, and procedures for deletion or incident reporting.
Features and contractual protections can vary across service tiers and may change over time.
Consequently, a privacy statement applicable to one product configuration should not automatically be assumed to govern another.
Institutional information security or data-protection personnel may be better positioned to evaluate these arrangements than an individual supervisor.
The supervisor should rely on documented authorization rather than informal assumptions about provider practices.
Why Data Minimization Matters
Data minimization involves limiting the information processed to what is necessary for a defined purpose.
The principle is reflected in data-protection frameworks, including the European Union's General Data Protection Regulation, although the regulation's legal applicability depends on the processing context.
For AI-assisted feedback, minimization may mean submitting a short, nonconfidential passage rather than an entire thesis, where doing so is authorized and sufficient.
It may also mean asking AI to evaluate a generic description of a methodological issue without disclosing the underlying participant records.
However, submitting excerpts is not automatically safe. A short passage can still contain identifiable or restricted information.
Minimization reduces unnecessary exposure but does not replace authorization.
Can Synthetic Data Support Confidentiality-Preserving Feedback?
Synthetic data may help supervisors demonstrate analytical procedures or evaluate AI-generated code without exposing actual participant records.
For example, a supervisor could construct a fictional dataset illustrating repeated measurements and ask AI to explain possible analytical approaches.
The resulting guidance would still require methodological evaluation before being applied to the real study.
Care is needed when synthetic data are derived from sensitive original records, because some generation methods may preserve information that creates disclosure risks.
Independently constructed fictional examples may be preferable for straightforward teaching purposes.
Synthetic material should not be presented as actual research evidence.
Are Locally Deployed AI Systems Automatically Confidential?
Local deployment may reduce some risks associated with transferring data to external providers.
However, local operation does not guarantee confidentiality.
Security depends on access controls, system administration, storage arrangements, logging, software vulnerabilities, and the handling of generated outputs.
For example, a locally deployed model operating on a shared server may expose material if permissions are configured incorrectly.
Supervisors should use systems assessed and authorized for the relevant information rather than assume that local installation alone establishes compliance.
Why Connected Tools and Document Retrieval Matter
Some AI systems can access cloud documents, institutional repositories, external databases, or connected applications.
These capabilities may introduce additional processing pathways beyond the initial prompt.
For example, a supervisor might ask an AI assistant to examine a thesis stored in a connected document service.
The supervisor should understand which documents the system can access, what information is retrieved, and whether the connection is authorized.
Permission to use a document platform does not automatically establish permission for every connected AI integration.
Institutional access controls and integration settings should be examined before use.
How Should AI-Generated Feedback Be Stored?
AI-generated comments may reproduce confidential passages from the student manuscript.
They may also include methodological observations about unpublished findings or participant information.
Consequently, feedback records may require the same confidentiality protections as the source material.
Supervisors should use authorized storage locations and limit access according to institutional requirements.
Retention should reflect legitimate supervisory, research, and records-management obligations.
Generated feedback should not automatically be treated as unrestricted material simply because it was produced by an AI system.
What If a Supervisor Accidentally Uploads Restricted Information?
The supervisor should follow the institution's applicable information-security and research ethics incident procedures promptly.
Relevant facts may include what information was submitted, which service received it, the account configuration, and whether the information remains accessible.
Institutional personnel may need to determine whether containment, provider contact, notification, or other corrective measures are required.
Deleting a conversation from a user interface does not necessarily establish that all copies of the information have been eliminated.
The appropriate response depends on the provider's arrangements and applicable obligations.
Watch Out
Do not assume that deleting an AI chat reverses an unauthorized disclosure. Follow the institution's incident-response process and obtain authoritative guidance about the information that may have been processed or retained.
Why Confidentiality and Feedback Accuracy Must Be Evaluated Separately
An AI system may be authorized to process a student draft but still generate inaccurate methodological feedback.
Conversely, a technically accurate review may have been produced through unauthorized processing.
These are distinct concerns.
The supervisor must ensure both that the processing arrangement is acceptable and that consequential comments are scientifically defensible.
The broader practice of using AI to review student research drafts requires independent verification and accountable human judgment.
Confidentiality protection does not eliminate the need for scholarly evaluation.
Should Students Be Informed About AI Processing?
Students may have legitimate interests in understanding how their unpublished research is handled.
Transparency is particularly important when AI processing affects confidential material or contributes materially to substantive feedback.
However, notification and authorization are not identical.
The precise requirements depend on institutional policy, the research context, and the nature of the processing.
Supervisors should explain relevant arrangements accurately and avoid implying that notification alone makes otherwise restricted processing permissible.
The question of disclosing AI-assisted review to students should be addressed through clear institutional and supervisory expectations.
How Can Research Teams Establish Consistent Confidentiality Rules?
Research supervision may involve multiple supervisors, laboratory members, external examiners, and collaborating institutions.
Different participants may use different AI systems or operate under different organizational policies.
A shared agreement should identify authorized tools, permitted data categories, access responsibilities, and procedures for resolving uncertainty.
Where institutions have conflicting requirements, collaborators should seek an arrangement that satisfies the applicable obligations before sharing material.
The broader question of research team agreements concerning AI use becomes particularly relevant when unpublished documents move between organizations.
Why a Risk-Based Approach Is Preferable to Assuming All Drafts Are Identical
A chapter discussing publicly available theories may present different confidentiality risks from a chapter containing identifiable clinical records.
Similarly, a short nonconfidential excerpt may require different controls from an entire restricted dataset.
The NIST AI Risk Management Framework emphasizes evaluating AI risks within their actual context.
For supervisory feedback, relevant factors include information sensitivity, processing environment, potential consequences, and available safeguards.
Risk assessment should not be used to override explicit prohibitions.
Where binding requirements restrict processing, supervisors must comply regardless of whether they personally consider the risk acceptable.
What Does a Confidentiality-Preserving Review Workflow Look Like?
A practical workflow begins by identifying the feedback task and determining whether AI is necessary.
The supervisor then classifies the material, checks authorization, selects an appropriate processing environment, and minimizes unnecessary information exposure.
After receiving generated feedback, the supervisor independently verifies consequential recommendations and stores the resulting records appropriately.
Student-facing transparency should follow applicable institutional requirements.
The workflow should be proportionate to the sensitivity of the document and the significance of the review.
Where adequate safeguards cannot be established, manual review or another authorized method remains an appropriate alternative.
