03 · What You Need to Know
What Meaningful AI-Use Documentation Requires
Why AI-Use Documentation Matters in Research Education
Research documentation helps establish how scholarly decisions were made, which procedures were followed, and what evidence supports the reported findings.
When AI contributes to research activities, additional information may be necessary to explain the relationship between automated suggestions and human decisions.
For example, a student who uses AI to generate statistical code should be able to identify what the code was intended to accomplish, how it was evaluated, and whether it was modified before use.
A student who uses AI to suggest interpretations should be able to explain which suggestions were accepted and why they were supported by the actual findings.
Documentation can help supervisors distinguish the student's intellectual contribution from automated assistance.
However, the existence of a record does not establish that the research is correct. Documentation supports scrutiny; it does not replace verification.
Documentation, Disclosure, and Reproducibility Are Different
These concepts overlap but should not be treated as interchangeable.
Documentation
Preserving relevant information about how AI contributed to research activities and decisions.
Disclosure
Communicating AI assistance to supervisors, institutions, publishers, or readers when appropriate or required.
Reproducibility concerns whether research procedures and results can be reconstructed or repeated under appropriately specified conditions.
An AI-use statement may disclose that a language model helped generate analytical code without providing enough information to reproduce the analysis.
Conversely, a detailed computational record may support reproducibility while failing to satisfy a publisher's disclosure requirements.
Students should understand which purpose each record serves.
What Do Research Integrity Principles Suggest?
The European Code of Conduct for Research Integrity, revised by ALLEA in 2023, identifies reliability, honesty, respect, and accountability as fundamental principles.
These principles support accurate reporting of research procedures and responsible management of research records.
However, the Code should not be interpreted as imposing a universal requirement that every student preserve every generative AI interaction.
Specific documentation requirements may arise from institutional policies, research protocols, funder conditions, or publication guidelines.
Students should consult the requirements applicable to their projects rather than assume that a single documentation standard governs all research.
Which AI-Assisted Activities Most Clearly Warrant Documentation?
The strongest case for documentation arises when AI materially influences consequential research work.
| AI-assisted activity |
Documentation consideration |
Illustrative record |
| Routine grammar editing |
Follow applicable disclosure rules; extensive logs may not be necessary |
Brief statement of language assistance where required |
| Literature-search development |
Record material changes to search strategies |
AI contribution, final search expression, databases, and verification |
| Research design advice |
Document consequential recommendations and final rationale |
Proposed alternatives, evidence consulted, and selected design |
| Instrument development |
Record substantive AI-generated items and validation decisions |
Item-development history and human review |
| Statistical code generation |
Preserve verified code and relevant implementation details |
Scripts, dependencies, tests, and analytical decisions |
| Qualitative coding assistance |
Document the role of AI consistent with the methodology |
AI-supported procedures, researcher interpretations, and relevant memos |
| Results interpretation |
Record material AI suggestions and independent evaluation |
Original output, proposed interpretation, corrections, and final justification |
| Manuscript drafting |
Follow institutional and publication requirements |
Relevant AI assistance and author verification |
These are practical recommendations rather than universally binding requirements.
The appropriate record depends on whether the assistance affects scientific validity, intellectual attribution, reproducibility, or compliance with applicable rules.
Should Every Prompt and Response Be Saved?
Not necessarily.
Complete interaction logs may contain substantial amounts of information unrelated to consequential research decisions.
They may also include preliminary ideas, confidential material, or personal information that does not need to be retained or disclosed.
Furthermore, retaining every interaction does not guarantee that a supervisor can understand how the research was conducted.
A large collection of prompts may be less informative than a concise record identifying the consequential recommendation, the evidence used to evaluate it, and the final human decision.
Complete logs may be appropriate in particular circumstances, such as a study investigating human-AI interaction, an approved audit, or a reproducibility protocol that explicitly requires them.
Outside such contexts, documentation should be proportionate to its purpose.
What Should a Meaningful AI-Use Record Contain?
For consequential assistance, a useful record may identify the date, AI system, task, nature of the assistance, relevant output, verification procedure, and final decision.
It may also explain whether the researcher accepted, modified, or rejected the recommendation.
Where model configuration or version information is available and materially relevant, recording it can help explain differences in system behavior.
However, model identifiers and prompts alone do not guarantee reproducibility because outputs may depend on system updates, configuration, retrieval context, and nondeterministic behavior.
Researchers should preserve the actual verified research procedures and outputs needed to reconstruct their work.
Why Recording the Final Decision Matters More Than Recording the Prompt Alone
A prompt documents what the student requested from AI.
It does not necessarily establish how the student evaluated the answer or why a recommendation was adopted.
For example, a student may ask AI to recommend a statistical model and receive several alternatives.
The research record should explain which model was ultimately selected, why it was appropriate, and what evidence supported the choice.
This makes the student's methodological reasoning visible.
Documentation should therefore focus on the relationship between automated assistance and accountable human judgment.
Can AI-Use Logs Help Supervisors Assess Learning?
Potentially, but their educational value depends on how they are interpreted.
A supervisor may use a concise AI-use record to identify where the student sought assistance and which decisions require further explanation.
For example, extensive AI assistance with statistical interpretation may indicate a need to assess the student's understanding of the analysis.
However, the number of prompts does not establish the student's competence.
A student may use AI frequently for practice and feedback while retaining strong independent understanding.
Another may use AI only once but delegate a consequential intellectual decision.
Supervisors should therefore assess independently demonstrable research competencies rather than treating interaction frequency as a measure of learning.
What Privacy Risks Can Documentation Create?
AI interaction logs may contain unpublished research ideas, participant information, institutional documents, or other confidential material.
Requiring students to submit complete logs could expose information beyond what is necessary for supervision.
Researchers should consider data minimization, access controls, retention periods, and applicable privacy requirements.
Students should not be encouraged to preserve sensitive prompts in insecure locations merely to demonstrate compliance.
Where detailed logs are required, institutions should provide appropriate procedures for secure storage, access, and eventual disposal.
Documentation obligations should not undermine existing research confidentiality commitments.
Should Students Record AI Use That Did Not Influence the Final Work?
The answer depends on the purpose of documentation.
For routine disclosure, an exploratory interaction that had no substantive influence may not require detailed reporting unless applicable policy says otherwise.
For a methodological study of AI-assisted decision-making, rejected recommendations may be important evidence.
Similarly, a student may benefit educationally from recording why an apparently convincing recommendation was rejected.
Such records can reveal the development of critical judgment.
However, documenting every discarded idea is not necessarily proportionate.
The relevant question is whether the interaction materially informs the research process, its accountability, or the intended learning outcome.
How Does Documentation Relate to AI Disclosure in Publications?
Publication policies vary among journals and publishers.
The International Committee of Medical Journal Editors states that authors should disclose relevant use of AI-assisted technologies and remain responsible for the integrity of their work.
Other publishers may distinguish among language editing, content generation, and analytical assistance.
Students should check the current policy of the intended publication venue.
Internal documentation can support accurate disclosure, but the information included in a manuscript should follow the applicable requirements.
A complete prompt history is not necessarily an appropriate or required publication disclosure.
How Can Documentation Support Research Reproducibility?
Reproducibility depends on preserving the procedures and information necessary to reconstruct research outputs.
For AI-generated code, this may include the final scripts, input-data descriptions, software dependencies, and tests.
For literature searches, it may include the actual databases, search expressions, dates, and screening procedures.
For qualitative analysis, appropriate records depend on the methodological approach and confidentiality constraints.
Sandve and colleagues (2013) emphasize practical principles for reproducible computational research.
AI-use documentation can complement these records but cannot substitute for them.
Saving a prompt that requested an analysis does not establish what analysis was actually performed.
Should Documentation Requirements Differ by Research Stage?
Introductory coursework, thesis research, and published scholarship may require different documentation practices.
In coursework, an instructor may request a short AI-use reflection to assess students' verification and decision-making.
In thesis research, a more systematic record may help supervisors evaluate consequential methodological choices.
For publication, disclosure should follow journal requirements and accurately represent the assistance received.
These differences are not inconsistencies. They reflect distinct educational and scholarly purposes.
Requirements should nevertheless be communicated clearly before the relevant work begins.
What Would a Proportionate Institutional Policy Look Like?
An institutional policy could distinguish routine assistance from substantive AI contributions and specify the records expected for each.
It should clarify which activities require disclosure, whether prompts must be retained, who may access records, and how confidential information is protected.
It should also establish reasonable procedures for correcting incomplete documentation.
The policy should avoid equating the absence of a complete interaction history with proof of misconduct.
Documentation concerns should be evaluated under applicable institutional rules and evidence.
The broader objective is meaningful human oversight rather than ceremonial record collection.
