03 · What You Need to Know
Responding to AI-Generated Methods That Students Cannot Defend
Why the Supervisor Must Separate Three Questions
When a student cannot explain a submitted method, the supervisor should distinguish scientific validity, student competence, and policy compliance.
Scientific validity
Is the selected method appropriate for the research question, data, assumptions, and intended inference?
Student competence
Can the student understand, apply, evaluate, and defend the methodology at the level required by the degree program?
A third question concerns compliance: whether the student used AI in a manner consistent with applicable academic and research requirements.
These issues may overlap but require different evidence.
A method can be valid even when the student cannot explain it. A student can understand a method while having violated an explicit AI-use restriction. A method can also be invalid despite the student's sincere belief that it is appropriate.
Supervisors should avoid assuming that answering one question resolves the others.
First, Establish What the Student Actually Submitted
The supervisor should identify the method, its role in the research, and whether it has already been implemented.
For example, the problem may concern a proposed analytical plan that has not yet been executed, or a completed analysis supporting the thesis findings.
The consequences differ.
An unsuitable proposed method may be corrected before data analysis begins.
An unsuitable method already used to produce reported findings may require reanalysis and revision of the results and discussion.
The supervisor should also determine whether the methodological description accurately represents procedures that were actually performed.
Research reporting must not describe an analysis as completed when it was merely suggested by AI.
Second, Identify What AI Contributed
AI may have explained an unfamiliar method, recommended the procedure, generated code, or drafted the methodology section.
These contributions have different implications.
For example, a student may independently select an appropriate regression model but use AI to help describe it.
Another may ask AI to select the entire analytical strategy without evaluating alternatives.
The supervisor should ask which decisions were made independently, which suggestions were accepted, and how the final procedure was verified.
Relevant documentation may help reconstruct the process, but complete interaction histories are not automatically necessary.
Third, Identify the Specific Knowledge Gap
A student may struggle to explain a technical term while understanding the broader analytical procedure.
Another may understand the mathematical formulation but fail to recognize that the model is inappropriate for the research design.
A third may be unable to explain the method's purpose at all.
These situations should not receive identical responses.
Supervisors should ask focused questions that identify which knowledge or reasoning is missing.
For example, questions may concern the research question, unit of observation, assumptions, variables, analytical alternatives, and interpretation.
The goal is diagnosis of the educational problem rather than immediate attribution of intent.
What Should Supervisors Ask Students to Explain?
| Area of understanding |
Illustrative supervisory question |
What the answer may reveal |
| Research purpose |
What question is this method intended to answer? |
Understanding of the method's scientific role |
| Method selection |
Why was this method chosen instead of an alternative? |
Methodological justification |
| Data structure |
What are the units of observation and relevant dependencies? |
Familiarity with the dataset |
| Assumptions |
Which assumptions are important for the intended inference? |
Recognition of methodological limitations |
| Implementation |
What did the analytical procedure actually do? |
Understanding of computational operations |
| Results |
How should the principal estimates be interpreted? |
Statistical or analytical reasoning |
| Uncertainty |
What conclusions cannot be justified from these results? |
Critical interpretation |
| AI contribution |
Which recommendations came from AI and how were they checked? |
Verification and intellectual accountability |
Questions should be appropriate to the student's training and the methodology involved.
They should not demand specialist knowledge unrelated to the student's actual research responsibilities.
Why a Polished Methods Section Does Not Establish Understanding
Generative AI can produce technically fluent descriptions of statistical, computational, and qualitative procedures.
However, the presence of accurate terminology does not establish that the student understands the procedure.
A student may submit an explanation containing correct definitions while remaining unable to apply those concepts to the actual dataset.
Similarly, an AI-generated paragraph may be scientifically inaccurate despite appearing sophisticated.
Supervisors should therefore evaluate both the written description and the student's ability to explain its relevance to the study.
This distinction is central to assessing independent research competence when AI assistance is allowed.
Why Inability to Explain Does Not Automatically Prove Misconduct
A student may lack understanding because of insufficient prior knowledge, inadequate instruction, confusion, or inappropriate reliance on automated assistance.
These explanations are not equivalent to deliberate misrepresentation.
Academic misconduct requires evaluation under applicable institutional definitions and procedures.
The European Code of Conduct for Research Integrity emphasizes honesty and accountability, but it does not establish that every instance of inadequate methodological understanding constitutes misconduct.
Supervisors should therefore avoid making disciplinary conclusions from a single unsuccessful explanation.
However, the absence of proven misconduct does not mean that scientifically indefensible work should be accepted.
How Should Supervisors Determine Whether the Method Is Valid?
The supervisor should examine the relationship between the research question, design, data structure, analytical procedure, and intended inference.
For example, an AI-recommended statistical test may be inappropriate because observations are not independent.
Another method may be suitable but described inaccurately.
Verification may involve examining the original data, analytical code, methodological literature, and actual outputs.
Where the supervisor lacks the necessary expertise, qualified methodological consultation should be obtained.
AI-generated confirmation should not be treated as sufficient independent verification.
What If the AI-Generated Method Is Correct?
A scientifically appropriate method does not automatically establish that the student has demonstrated the competencies required by the degree program.
For example, the analytical procedure may be valid, but the student may still be unable to explain why it was selected.
The supervisor should identify the required level of understanding and provide appropriate opportunities for learning or reassessment.
Where the student subsequently demonstrates competence, the original AI assistance may not prevent the work from being accepted, provided that applicable rules were followed.
The educational concern is whether the student can defend the method, not whether the method originated from an automated suggestion.
What If the Method Is Scientifically Incorrect?
The affected research should be corrected before its conclusions are accepted.
If the analysis has already been performed, the student may need to revise the model, rerun the analysis, and reconsider the findings.
Changes should be documented accurately.
If the method was only proposed, the student may revise the analytical plan before implementation.
The response should reflect the significance of the error and the stage of the research.
Supervisors should avoid treating correction as optional simply because the original recommendation came from AI.
What If the Student Cannot Explain AI-Generated Statistical Code?
Generated code may implement operations that are not obvious from its surface structure.
A student may understand the statistical model but be unfamiliar with particular programming functions.
Alternatively, the student may lack understanding of the analysis itself.
Supervisors should distinguish programming syntax from scientific reasoning.
The student should understand the intended analytical procedure and ensure that the implementation is appropriately tested.
Where necessary, a qualified analyst may review the code.
Wilson and colleagues (2017) discuss scientific computing practices involving testing, documentation, and reproducibility.
These principles remain relevant when AI generates analytical scripts.
What If the Student Cannot Explain AI-Generated Qualitative Methods?
Qualitative methods require attention to methodological coherence and interpretive responsibility.
For example, a student may submit AI-generated themes without understanding how they were developed from the original interview material.
In reflexive thematic analysis, Braun and Clarke (2022) emphasize the researcher's active role in interpretation.
A supervisor should examine whether the student engaged adequately with the data and whether the analytical procedure matches the selected methodology.
However, qualitative traditions differ, and expectations should reflect the actual approach.
The supervisor should not impose one universal coding procedure across all qualitative research.
Should the Student Be Required to Rewrite the Methods Section?
Rewriting may be appropriate when the submitted description is inaccurate, unclear, or inconsistent with the procedures actually performed.
However, rewriting alone may not resolve inadequate understanding.
A student may generate another polished explanation without addressing the underlying knowledge gap.
Supervisors should therefore identify the specific competency requiring development.
The student may need to explain the method independently, demonstrate its application, or justify it using appropriate methodological evidence.
The revised text should accurately represent the final research procedure.
How Can Supervisors Provide Educational Remediation?
Remediation should address the specific knowledge or skill deficiency.
For example, a student unfamiliar with regression assumptions may review relevant methodological materials and complete guided interpretation exercises.
A student who does not understand data transformations may inspect the analytical script and compare input and output records.
Instructional scaffolding may be useful when the student is learning unfamiliar procedures.
Wood, Bruner, and Ross (1976) describe how tutoring support can help learners perform tasks beyond their current independent capabilities.
Assistance should be reduced as understanding develops.
The objective is to establish competence rather than simply produce a corrected manuscript.
How Should Independent Reassessment Be Conducted?
Reassessment should align with the competency that was initially unclear.
For example, a student may be asked to interpret an unfamiliar statistical output, compare methodological alternatives, or explain the assumptions of the revised model.
The task should be appropriately scoped and consistent with degree requirements.
Oral explanations may be useful, but they should not automatically be treated as perfect measures of competence.
Communication anxiety, language proficiency, and accessibility considerations may influence performance.
Written or practical demonstrations may provide complementary evidence.
Why AI-Text Detection Is Not the Appropriate Response
The immediate concern is whether the student understands and can defend the submitted method.
AI-text detection tools do not directly measure this competence.
Even an accurate classification of generated text would not establish whether the methodology is valid or whether the assistance was permitted.
Liang and colleagues (2023) reported limitations involving bias against non-native English writing in several GPT detectors.
Supervisors should therefore focus on methodological evidence and applicable institutional procedures rather than rely on detection scores.
What If AI Use Violated an Explicit Restriction?
If the student used AI contrary to a clearly applicable requirement, the supervisor should identify the specific rule and relevant evidence.
The concern should be addressed through the institution's established procedures.
The student should receive the opportunity to respond required by those procedures.
Scientific correction and disciplinary review may both be necessary, but they serve different purposes.
Supervisors should avoid imposing outcomes beyond their authority or treating an allegation as an established finding.
The broader question of responding to student AI-use violations requires procedural fairness and evidence-based judgment.
What If the AI Rules Were Ambiguous?
Ambiguous AI expectations can complicate supervisory responses.
A student may have understood that AI-assisted methodological advice was permitted because no specific restriction had been communicated.
However, the absence of an AI-specific prohibition does not eliminate general obligations concerning accurate reporting, research ethics, or scientific validity.
Supervisors should distinguish previously established requirements from new expectations they wish to introduce.
Where clarification is needed, the responsible institutional authority should be consulted.
New supervisory preferences should not be retrospectively presented as binding rules that students were already required to follow.
Why Supervisors Should Avoid Taking Over the Research
When a student struggles with an AI-generated method, a supervisor may be tempted to correct the entire analysis personally.
Specialist assistance may be appropriate, particularly when errors could materially affect the research.
However, the educational objective remains the student's development of the competencies required by the degree program.
Supervisors should provide sufficient support to resolve the methodological problem while preserving opportunities for the student to understand and defend the work.
Otherwise, replacing AI-generated reasoning with supervisor-generated reasoning may leave the underlying competence problem unresolved.
How Should the Incident Be Documented?
Supervisors should preserve appropriate records of consequential methodological concerns, corrective actions, and reassessment outcomes.
For example, the record may identify the initial analytical problem, the revised procedure, the evidence supporting the correction, and the student's demonstrated understanding.
Documentation should be accurate and proportionate.
It should not unnecessarily include unrelated AI conversations or personal information.
Where a formal institutional process is initiated, records should be handled according to applicable procedures.
How Can Similar Problems Be Prevented?
Prevention begins with explicit expectations about the student's intellectual responsibilities.
Supervisors should identify which AI-assisted activities are permitted, which require consultation, and what verification is expected.
Students should know that generated methodological explanations must correspond to procedures they understand and can defend.
Research meetings can include discussion of analytical decisions before they become embedded in completed chapters.
The distinction between legitimate AI assistance and intellectual outsourcing should be addressed as part of research training.
Clear expectations may reduce misunderstandings, although they cannot eliminate every methodological error or learning difficulty.
