When Using AI in Research, Which Steps Must Be Documented?
Both researchers stated publicly that they “used AI to assist with data cleaning.” One had the model write scripts, which were then validated using fixed rules and test samples; the other fed the samples to the model, which determined which ones should be excluded. The first approach was primarily intended to demonstrateCorrect Implementation of Rules in Code, and for the second one, we still need to proveChanges in the rules governing model participation and in the sample will not distort the study’s conclusions.. Such a vague statement does not tell readers whether the AI helped write the code or influenced which samples were included in the analysis.
Cornell University Interim Policy on “Responsible and Ethical Use of AI in Research and Academia,” Section 1.8Published on November 2, 2026; a non-substantive clarification to Section 3.1 was made on November 19. It applies to individuals conducting research at Cornell’s various campuses. For other institutions, it may be more helpful to consider the following three issues separately:Can the data be imported? How are the results verified? What information is disclosed to whom?
Can the data be included? Approval is granted for the tools and their intended uses, not for all materials.
Section 3.2 requires the use of approved tools when available; if none are available, the IT governance process must be followed. Data classified by the school as medium- or high-risk must not be provided to AI tools that have not undergone appropriate internal review. Section 4.3.6 places the focus on control: even if a vendor promises that data “will not be used for training,” it is necessary to verify whether the information will be transferred outside the scope of the project’s control.
If a study uses interview transcripts, removing names does not necessarily ensure anonymity. The combination of job titles, organizations, and specific experiences may still point to an individual. A team’s decision record should state “why this type of data, for this purpose, and in this context, can be used,” rather than simply stating “the company purchased the Enterprise Edition.” When processing only a portion of a text, there is no need to submit the entire file; project logs should record the location of controlled documents and the basis for approval, rather than including a copy of the original text.
How to Verify Results: Documentation Should Follow the Decision-Making Process
Sections 3.3, 4.3.1, and 4.3.3 distinguish between verification and responsibility: Users should verify the output, while the Principal Investigator (PI) bears ultimate responsibility for team protocols and sufficiently detailed project records, which must be retained for a period consistent with project data requirements. Outputs that affect experimental design, risk identification, or regulatory compliance should also be reviewed in conjunction with the appropriate compliance or safety office. The policy does not prescribe a single, universal log form for all projects; the following distinctions can be applied in practice:
| At what stage does AI come into play? | What Should Be Prioritized for Preservation in the Project? | Who or how will the review be conducted? |
|---|---|---|
| Help write a cleaning script | Script version, actual code used, input version, boundary tests | Have a colleague review the discrepancies and rerun the analysis using a fixed sample. |
| Classify the samples and determine whether to include them | Rule or hint version, original output index, manual re-evaluation and rationale, number of samples before and after | Without relying on the model, first conduct a sample review to assess testing errors across different populations; compare the main conclusions before and after the reevaluation. |
| Generate substantive text, figures, or tables | Records of adopted sections, original sources, manual rewrites, and proofreading | Verify citations, numerical values, and attributions item by item; the author is responsible for this. |
If the online model version cannot be locked, simply recording the system name does not guarantee reproducibility. You can record the date of the call, the version displayed by the system at the time, and the configuration (if available). Within the scope of the license, retain the output and the manually finalized version, and then create a standalone script of the statistical steps used to reach the conclusion. While it cannot guarantee that the exact same sentence will be generated in the future, this does not prevent a review of why the decision was made at the time or the extent to which the results were affected.
Disclosure isn't just saying, “We used AI.”
Section 3.4 specifies that if AI generates substantial research outputs or academic writing, or serves as an integral part of the research methodology, this must be disclosed. The policy also encourages truthful disclosure of broader auxiliary uses.Internal project records, disclosure of methods to readers, and informed consent from participants are distinct matters.: If the Institutional Review Board (IRB) determines that the use of AI will affect research participants, this must be disclosed during the informed consent process; it cannot be substituted with an acknowledgment at the end of the paper. When submitting a manuscript, authors should also verify the requirements of their funding agreements and the target journal, and should not treat Cornell University’s regulations as a universal template.
One suggested template that can be adapted is: “We provide recommendations on the [steps] for [material] using [tools and versions]; [role] reviews them according to predetermined [rules]; discrepancies are handled by [role or process]; the impact on key results is discussed in [sensitivity analysis or appendix].” Only include in the square brackets what was actually completed. If the team could only change one thing right now, I would move the record toAI Transforms Research Decision-Making for the First TimeThe key point is this: at that stage, the person responsible for data entry, the person responsible for data output, and the person responsible for verification can still be located; these are precisely the details most likely to be overlooked when the manuscript is submitted.
© Copyright notes
The copyright of the article belongs to the author, please do not reprint without permission.
Related articles
No comments...