03 · What You Need to Know
Establishing Defensible Boundaries for AI in Student Research
Why Supervisors Need Explicit Permission Categories
Instructions such as "use AI responsibly" leave substantial room for interpretation. Students may understand responsibility as checking the final output, while supervisors may expect consultation before AI contributes to research design or data analysis.
A practical policy should distinguish three situations: assistance permitted under standing conditions, assistance requiring prior review, and assistance that is prohibited.
These categories should be supported by clear reasons. Restrictions may protect research participants, preserve assessment validity, prevent scientific misrepresentation, or enforce institutional obligations.
The supervisor should also distinguish personal preferences from binding institutional requirements. A supervisory agreement cannot override research ethics approval conditions or university data-security rules.
Which AI Uses Can Usually Be Permitted Under Standing Conditions?
Some activities may be permitted without requiring students to request individual approval every time.
Examples include explaining general methodological terminology, suggesting alternative search keywords, improving the clarity of nonconfidential writing, and generating practice examples using synthetic data.
Permission should still have conditions. Students must verify substantive claims, preserve scientific meaning, and comply with applicable disclosure requirements.
For example, using AI to explain the difference between statistical power and effect size may support learning. Accepting an AI-generated calculation or interpretation without checking its assumptions is a separate activity.
Standing permission should describe what students may do and where that permission ends.
Which Activities Should Require Prior Supervisory Review?
Prior discussion is especially valuable when AI may influence consequential decisions that are difficult to reverse or central to the thesis.
These may include changing the research design, generating substantial measurement instruments, selecting the primary analytical model, automating qualitative interpretation, or modifying data-exclusion procedures.
Supervisory review should examine whether the proposed AI assistance is methodologically appropriate, what information will be processed, and how the output will be verified.
Not every such activity requires formal institutional approval. That depends on the applicable rules and whether the proposed change affects an approved protocol.
The supervisor should identify which decisions can be resolved through ordinary consultation and which must be referred to a research ethics committee or another authorized body.
Which AI Uses Should Be Explicitly Prohibited?
Some activities are incompatible with responsible research because they involve misrepresentation, unauthorized processing, or violation of established requirements.
For example, students should not use AI to fabricate participant responses, invent experimental procedures, create fictitious research findings, or present analyses as performed when they were not conducted.
Students should not upload restricted participant information to an AI service when that processing has not been authorized.
AI assistance should also be prohibited where an applicable independent assessment expressly requires students to perform the relevant intellectual work themselves.
These prohibitions should identify the underlying conduct rather than imply that every AI-generated output is inherently improper.
A Practical Permission Matrix for Thesis Research
| Research activity |
Suggested supervisory category |
Essential condition |
| Explaining general research concepts |
Standing permission |
Verify consequential explanations before applying them |
| Suggesting literature-search keywords |
Standing permission |
Test terms in appropriate databases |
| Editing nonconfidential prose |
Standing permission where policy allows |
Preserve meaning and meet disclosure requirements |
| Generating practice code using synthetic data |
Standing permission |
Inspect and test the implementation |
| Developing questionnaire items |
Prior discussion |
Evaluate content and measurement validity |
| Selecting the primary statistical model |
Prior discussion |
Require independent methodological justification |
| AI-assisted qualitative interpretation |
Prior review where consequential |
Ensure methodological coherence and data authorization |
| Changing research eligibility criteria |
Appropriate protocol and institutional review |
Obtain required authorization before implementation |
| Uploading restricted data without authorization |
Prohibited |
Use only an authorized processing arrangement |
| Fabricating research observations or results |
Prohibited |
Research evidence must represent what actually occurred |
| Using AI contrary to explicit assessment restrictions |
Prohibited during that assessment |
Preserve the competency being independently measured |
This matrix is a recommended starting point, not a universal institutional policy. Supervisors should adapt it to the discipline, student competence, approved research protocol, and governing requirements.
Why Tool Approval Is Not the Same as Permission for Every Task
A university may provide access to an institutionally approved AI system.
That approval does not necessarily authorize all research activities or all categories of information.
For example, an approved system may be suitable for public literature exploration but not for identifiable participant records unless specific processing arrangements have been assessed.
Supervisors should distinguish permission to access a tool from authorization to use it for a particular purpose.
Relevant considerations include provider conditions, data retention, access controls, consent arrangements, and institutional policies.
How Should Confidential Research Data Affect AI Permissions?
Data sensitivity should influence the permitted processing environment and required safeguards.
Interview transcripts, institutional records, unpublished manuscripts, and participant datasets may contain information subject to confidentiality obligations.
Even removing direct identifiers may not eliminate re-identification risks.
Supervisors should establish which information can be processed using particular systems and when approval is necessary.
Where external processing is unauthorized, students may use synthetic teaching materials or approved alternatives instead.
Permission to use AI for an analytical task should never be interpreted as automatic permission to upload the underlying research data.
Why AI-Assisted Literature Searching Needs Clear Boundaries
AI can suggest search terminology, explain unfamiliar concepts, and help organize preliminary research questions.
However, generated references may be inaccurate or fabricated. An AI-generated synthesis may also omit relevant literature or misrepresent original findings.
Supervisors should permit appropriate exploratory assistance while requiring students to verify bibliographic records and examine original sources.
When a review requires systematic searching, the student must follow the methodology and reporting requirements appropriate to that review.
A chatbot-generated list of publications should not be represented as a comprehensive database search unless the actual retrieval process supports that claim.
Should AI Be Permitted to Generate Statistical Code?
AI-generated code may be legitimate research assistance when students understand the intended procedure and the implementation is appropriately tested.
However, successful execution does not guarantee analytical correctness.
A script may run while incorrectly merging datasets, excluding valid observations, or applying an unsuitable model.
Supervisors should require students to explain the analysis and preserve verified computational records.
Where the code affects consequential research findings, qualified review may be appropriate.
The central concern is not whether AI produced the syntax, but whether the final procedure is scientifically defensible and attributable to accountable researchers.
Should AI Be Permitted to Generate Research Instruments?
AI can suggest questionnaire items, interview questions, or alternative formulations.
Such suggestions may support preliminary development, but they do not establish validity or reliability.
Supervisors should require appropriate evaluation of item relevance, clarity, construct representation, and other measurement properties.
Depending on the project, substantive changes to approved instruments may also require protocol review.
Students should understand that an AI-generated instrument is not automatically a validated instrument.
Prior discussion is particularly useful when generated content may materially influence how the study measures its principal constructs.
How Should Qualitative Research Be Treated?
AI-assisted qualitative analysis requires attention to the selected methodological approach.
For example, reflexive thematic analysis emphasizes the researcher's interpretive engagement with the data, as discussed by Braun and Clarke (2022).
Automatically accepting generated themes without examining the original material may be inconsistent with the intended analytical approach.
Other qualitative methods may support different forms of structured computational assistance.
Supervisors should evaluate methodological coherence rather than impose one universal rule for qualitative research.
Confidentiality requirements must also be addressed before research material is processed.
Why Independent Research Competence Still Matters
Permitting AI does not eliminate the student's obligation to demonstrate the competencies expected by the degree program.
A student may use AI to suggest a methodology while remaining responsible for explaining why the final design is appropriate.
Similarly, generated interpretations must be evaluated against the actual findings.
Supervisors should identify the research competencies students must demonstrate independently.
This does not require every operation to be performed manually. It requires sufficient understanding to evaluate and defend consequential decisions.
Why Permission Categories Should Reflect Student Competence
Two students may need different levels of supervision for the same technical task.
A student with demonstrated programming expertise may competently inspect AI-generated code, while another may need guided instruction and independent practice.
However, supervisors should avoid treating academic seniority as proof of technical competence.
Permissions should reflect the activity's risks and the student's ability to evaluate its outputs.
Where students lack relevant expertise, the appropriate response may be additional training or qualified review rather than an unconditional prohibition.
How Should Disclosure Requirements Be Attached to Permissions?
Permitted AI assistance may still require documentation or disclosure.
For example, routine grammar correction might be covered by standing permission, while AI-generated analytical code may require a record of the generated contribution, tests, and final decisions.
Students should understand the difference between authorization, documentation, and disclosure.
The question of which AI interactions students should disclose to supervisors should be addressed explicitly rather than left to personal interpretation.
Applicable institutional and publication requirements may impose additional obligations.
Should Supervisors Require Every Prompt to Be Reviewed?
Not necessarily.
Complete prompt histories may contain sensitive information, unrelated conversations, and preliminary ideas that are not needed for research oversight.
For consequential assistance, a concise decision record may be more useful than an extensive interaction history.
Such a record can identify the AI contribution, verification undertaken, and final human decision.
However, detailed logs may be justified in specific research protocols or institutional processes.
Supervisors should establish requirements that are proportionate to their purpose and consistent with confidentiality obligations.
How Should Supervisors Communicate Restrictions?
Rules should be communicated before the relevant activity begins.
For each restriction, students should know which assistance is prohibited, why the restriction exists, and whether any alternatives are permitted.
For example, an instructor may prohibit AI-generated statistical interpretation during an independent assessment while permitting a calculator and ordinary reference materials.
The rule should identify the intellectual competency being protected.
Clear boundaries are part of establishing explicit supervisory AI expectations.
New supervisory preferences should not be retrospectively presented as previously binding institutional rules.
What If Institutional and Supervisory Rules Conflict?
A supervisor cannot ordinarily authorize conduct prohibited by a binding institutional requirement.
Conversely, institutions may allow supervisors to establish additional task-specific expectations within their authority.
Where a proposed restriction or permission conflicts with a formal policy, the relevant graduate school, research ethics committee, or institutional governance office should be consulted.
Students should not be expected to resolve ambiguous institutional authority independently.
Written clarification is particularly useful when research involves multiple institutions or supervisors.
Why a Prohibition Should Have a Defensible Purpose
Prohibitions should address identifiable concerns rather than general discomfort with AI.
For example, preventing unauthorized processing of confidential data protects existing research obligations.
Restricting AI during an independent methodological assessment protects the validity of the assessment.
Prohibiting fabricated research evidence protects scientific integrity.
These rationales differ, but each can be communicated clearly.
Supervisors should periodically examine whether restrictions remain appropriate as the research progresses and available safeguards change.
