03 · What You Need to Know
Supervising AI-Assisted Research Without Losing Sight of Student Development
Begin by Establishing What the Student Actually Did
When a supervisor discovers AI assistance, the first task is to understand its role in the research.
AI may have corrected grammar, suggested search terms, generated analytical code, drafted substantive explanations, or recommended methodological decisions.
These activities have different implications.
For example, correcting sentence structure generally raises different concerns from allowing AI to select a statistical model that the student cannot explain.
The supervisor should ask which tasks involved AI, how the outputs were evaluated, and what intellectual decisions the student made independently.
This fact-finding process should be neutral rather than accusatory.
Without understanding the actual assistance, a supervisor cannot reliably determine whether the issue concerns research quality, learning, transparency, or compliance.
Distinguish Legitimate Assistance From Intellectual Outsourcing
Researchers have long used software, reference managers, statistical packages, editorial services, and specialist collaborators.
Generative AI introduces additional forms of assistance, including some that can perform reasoning-intensive tasks.
The relevant distinction is not simply whether a tool contributed to the final work.
Legitimate research assistance
AI supports an authorized activity while the student retains and demonstrates the intellectual responsibilities required by the research project and degree program.
Inappropriate intellectual outsourcing
AI performs substantive reasoning or scholarly work that the student is required to undertake or independently defend, without sufficient understanding or authorization.
The boundary depends partly on the activity's educational purpose and the institution's rules.
For example, AI-generated code may be permissible during thesis research when the student can explain the analysis and verify the implementation.
The same assistance may be prohibited during an assessment intended to measure independent programming competence.
Supervisors should therefore evaluate the intellectual contribution rather than rely solely on the amount of generated material.
Check Applicable Policies Before Making a Judgment
Supervisors should establish which requirements govern the student's work.
Relevant sources may include graduate school regulations, university AI policies, research ethics approvals, course requirements, confidentiality agreements, and funder conditions.
Publication policies may become relevant when thesis findings are prepared for journal submission.
The European Code of Conduct for Research Integrity, revised by ALLEA in 2023, provides general principles of reliability, honesty, respect, and accountability.
However, general research integrity principles should not be misrepresented as universal rules prohibiting particular AI activities.
Where policies are ambiguous, supervisors should consult the responsible institutional authority rather than invent restrictions or assume that their personal preferences constitute binding requirements.
Assess the Student's Understanding Before Judging the Work
A polished AI-assisted thesis chapter does not necessarily establish that the student understands its methodology or conclusions.
Conversely, AI assistance does not establish that the student lacks competence.
Supervisors should ask students to explain consequential research decisions.
For example, a student might be asked why a particular design was selected, how the variables were measured, what assumptions the analysis requires, and which limitations affect interpretation.
The questions should reflect the competencies expected by the degree program.
This supports assessment of the student's own research ability when AI assistance is permitted.
Assessment should be educationally appropriate and should not automatically become an interrogation simply because AI was involved.
Why Inadequate Understanding Does Not Automatically Establish Misconduct
A student may be unable to explain a method because they have not yet mastered it, misunderstood an explanation, received inadequate instruction, or relied excessively on AI.
These possibilities require different responses.
Insufficient competence may justify additional teaching, revision, or reassessment.
Academic misconduct, by contrast, must be evaluated under applicable institutional definitions and procedures.
A supervisor should not infer dishonesty solely from an inability to explain sophisticated material.
Equally, a student should not be allowed to retain scientifically indefensible work simply because the underlying problem is educational rather than disciplinary.
Research quality and academic conduct require separate but appropriately coordinated consideration.
How Should Supervisors Respond to Different Forms of AI Assistance?
| Student AI activity |
Primary supervisory question |
Possible response |
| Language editing |
Was meaning preserved and were applicable rules followed? |
Review the final text and any required disclosure |
| Literature exploration |
Were references and attributed findings verified? |
Examine selected original sources and search procedures |
| Research design suggestions |
Can the student justify the selected methodology? |
Discuss assumptions, alternatives, and methodological evidence |
| Generated analytical code |
Was the implementation tested and understood? |
Request relevant code checks and analytical explanations |
| Qualitative interpretation |
Does the student demonstrate appropriate engagement with the data? |
Review analytical reasoning and methodological coherence |
| AI-generated manuscript arguments |
Can the student defend the claims and supporting evidence? |
Require source verification and independent explanation |
| Restricted data uploaded externally |
Was processing authorized? |
Follow applicable data-security and institutional procedures |
| AI use contrary to explicit rules |
What requirement applied and what evidence establishes the conduct? |
Address the concern through appropriate institutional processes |
These responses are illustrative. Actual supervisory authority and procedures vary across institutions.
Why Supervisors Should Evaluate Verification, Not Merely AI Disclosure
A student may openly disclose using AI while failing to verify its outputs.
For example, an AI-generated literature review may contain genuine references that do not support the attributed claims.
Disclosure does not correct these inaccuracies.
Supervisors should examine whether consequential information was checked against original sources, analytical records, or authoritative methodological guidance.
For computational work, successful code execution does not establish that the implementation is scientifically correct.
For qualitative research, a fluent AI-generated interpretation does not necessarily demonstrate methodological coherence or adequate engagement with the original material.
Transparency and verification are complementary responsibilities.
How Much AI-Use Documentation Should Supervisors Request?
Documentation can help explain how automated assistance influenced the research.
For consequential decisions, a record may identify the AI task, relevant recommendation, verification undertaken, and final human judgment.
However, requiring every prompt and response may be unnecessary or disproportionate.
Complete interaction histories can contain unrelated personal information, confidential research material, and preliminary ideas.
Supervisors should distinguish the records necessary for scholarly accountability from information that serves no clear supervisory purpose.
The question of whether supervisors should review AI prompts or interaction logs requires attention to privacy, proportionality, and applicable institutional requirements.
How Should Supervisors Protect Research Confidentiality?
AI-assisted supervision may involve unpublished manuscripts, participant records, institutional datasets, or sensitive interview material.
Supervisors should establish whether the proposed AI system is authorized to process such information.
Approval may depend on institutional arrangements, provider conditions, consent, and research ethics requirements.
De-identification may reduce risks but does not necessarily eliminate them.
If a student has already uploaded restricted material without authorization, the supervisor should assess the situation through the appropriate institutional data-governance or incident-response process.
The response should address actual information-security obligations rather than assume that every external AI interaction creates the same level of risk.
How Should Supervisors Handle AI-Assisted Statistical Analysis?
Statistical assistance requires attention to the research question, data structure, analytical assumptions, and interpretation.
A student may use AI to generate preliminary code for a model selected through independent methodological reasoning.
Alternatively, the student may allow AI to choose the model without understanding the relevant assumptions.
These situations warrant different supervisory responses.
Supervisors should examine whether the student understands the unit of observation, model specification, uncertainty, and limitations.
Where the analysis exceeds the supervisor's expertise, qualified methodological consultation may be appropriate.
Neither the student's nor the supervisor's confidence should substitute for relevant expertise.
How Should Supervisors Handle AI-Assisted Qualitative Research?
Qualitative research requires attention to methodological assumptions and the relationship between researchers and their data.
Some approaches permit structured coding assistance, while others place particular emphasis on the researcher's active interpretive engagement.
For example, Braun and Clarke (2022) emphasize methodological coherence and researcher interpretation in thematic analysis.
A supervisor should examine whether the student's AI-assisted procedure is consistent with the selected approach.
The student should be able to explain how interpretations were developed and substantiated.
Automated coding should not be accepted merely because the resulting categories appear plausible.
How Should Supervisors Respond When Students Cannot Explain AI-Generated Methods?
The supervisor should first identify the specific knowledge gap.
For example, a student may understand the research design but be unfamiliar with a technical detail of the generated code.
Another may be unable to explain why the entire analytical approach was selected.
The educational significance differs.
The supervisor may request a methodological explanation, guided revision, consultation with a specialist, or an independent demonstration of the relevant competency.
Where the work cannot be defended, it should not be accepted as scientifically adequate until the problem is resolved.
However, inability to explain the work should not automatically be treated as proof of a policy violation.
What Should Happen When AI Rules Were Not Clearly Communicated?
Ambiguous expectations complicate fair supervisory responses.
A student may reasonably have understood that AI assistance was permitted when no explicit restriction existed.
However, absence of a specific AI prohibition does not eliminate general research integrity, confidentiality, or methodological obligations.
Supervisors should distinguish previously established requirements from new expectations they wish to introduce.
Where a new rule is appropriate, it should be communicated clearly and applied prospectively unless an existing binding requirement provides otherwise.
Supervisors should avoid retroactively presenting personal preferences as formal institutional rules.
This is one reason clear expectations for student AI use should be established early.
Why Supervisors Should Avoid AI-Detection-Based Accusations
Automated AI-text detectors do not directly establish whether a student understands their research or whether an institutional rule was violated.
Such systems may produce false positives and false negatives.
Liang and colleagues (2023) reported bias against non-native English writing in several GPT detectors under the conditions they examined.
These findings do not establish that every detector performs identically, but they illustrate the risks of relying on automated classification as decisive evidence.
Supervisors should focus on verifiable research concerns and follow institutional procedures where misconduct is suspected.
A detection score should not replace examination of the work, relevant evidence, and the student's opportunity to respond.
How Should Supervisors Respond to Explicit AI-Use Violations?
If a student appears to have used AI contrary to an applicable restriction, the supervisor should identify the specific rule and establish the relevant facts.
The response should follow institutional procedures and respect the student's opportunity to explain the circumstances.
For example, unauthorized use during an independent assessment may be governed by academic integrity regulations.
Unauthorized disclosure of participant information may require a separate data-security response.
These concerns should not automatically be treated as interchangeable.
Supervisors should avoid promising sanctions or outcomes beyond their authority.
The aim is a fair, evidence-based response that protects both the student and the integrity of the research.
Why Supervisors Must Also Examine Their Own AI Practices
Supervisors may use AI to summarize student drafts, generate feedback, or evaluate methodological alternatives.
These practices raise similar questions about accuracy, confidentiality, and human responsibility.
A supervisor should not upload a student's confidential manuscript to an unauthorized service merely because the intended purpose is educational feedback.
Nor should AI-generated criticism be presented as an authoritative evaluation without appropriate review.
Responsible supervision requires consistency between the expectations imposed on students and the standards applied to supervisory AI use.
This does not mean supervisors and students necessarily have identical permissions. Their responsibilities and institutional authority may differ.
How Can Supervisors Encourage Appropriate Student Autonomy?
Research supervision should progressively develop independent scholarly judgment.
Excessive control over every AI interaction may discourage students from developing responsible decision-making.
Conversely, unrestricted delegation may leave important competencies insufficiently developed.
A proportionate approach distinguishes routine permitted assistance from consequential activities requiring consultation.
Students should know when they can proceed independently and when methodological, ethical, or institutional approval is necessary.
Supervisors can use discussions of AI-assisted decisions to develop critical reasoning rather than merely monitor compliance.
How Should Supervisory Expectations Evolve?
AI use may change as a research project moves from proposal development to data collection, analysis, thesis examination, and publication.
For example, exploratory use involving public literature may present different concerns from processing confidential participant data.
Publication preparation may introduce additional disclosure requirements.
Supervisors should review expectations at relevant research milestones.
Changes should be documented where appropriate and communicated clearly.
The goal is consistent scholarly accountability without assuming that a single rule adequately addresses every research activity.
