03 · What You Need to Know
What Makes an AI-Use Log Useful for Research Integrity and Reproducibility?
What Is a Research AI-Use Log?
A research AI-use log is a structured record of artificial intelligence assistance used during the planning, conduct, analysis, or reporting of a research project.
It may take the form of a spreadsheet, laboratory notebook entry, research management record, or another institutionally approved documentation system.
The log identifies consequential AI-assisted activities and connects them to the relevant research decisions, outputs, and verification records.
For example, a researcher may record that an AI system suggested preliminary qualitative codes, that the research team reviewed those codes, and that several suggestions were rejected because they misrepresented participants' responses.
This is more informative than simply recording that AI was used for qualitative analysis.
The log should complement existing research documentation rather than create an entirely separate and disconnected account of the project.
Is an AI-Use Log Required by Every University or Journal?
No universal requirement obliges all researchers to maintain a dedicated AI-use log.
Universities, funders, journals, and research agreements may establish different documentation and retention requirements.
Some policies emphasize disclosure of AI assistance in publications without prescribing a particular internal logging format.
Others may require records sufficient to demonstrate how AI was used or how consequential outputs were verified.
Researchers should therefore distinguish a recommended documentation practice from a mandatory institutional requirement.
The broader question of whether researchers should keep records of AI use depends on both the research context and applicable policies.
What Information Should an AI-Use Log Contain?
A practical log should capture enough information to identify the activity, reconstruct important decisions, and locate supporting records.
| Log Field |
What to Record |
Why It Matters |
| Date of use |
When the AI interaction or activity occurred. |
Establishes chronology and helps identify the relevant service configuration. |
| AI tool and provider |
The application or service used. |
Identifies the technology involved. |
| Model or version |
The documented model identifier, where available. |
Supports technical identification and methodological evaluation. |
| Research activity |
The task AI assisted with. |
Connects AI use to the research workflow. |
| Prompt or instructions |
Consequential instructions or a reference to their stored location. |
Helps reconstruct how outputs were generated. |
| Input materials |
The types of data or files supplied to the tool. |
Documents data processing and relevant confidentiality considerations. |
| AI output |
The generated response, code, classification, or other relevant result. |
Preserves what the system contributed. |
| Human decision |
Whether the output was accepted, modified, or rejected. |
Documents consequential scholarly judgment. |
| Verification |
How the output was checked and by whom, where relevant. |
Supports evaluation of reliability and accountability. |
| Research record reference |
Links to scripts, drafts, analytical memos, or other supporting materials. |
Connects the AI log to the broader research audit trail. |
Not every field needs the same level of detail for every activity. A minor language correction may require little documentation, while AI-assisted analysis that influences published findings may warrant a more complete record.
Why Should the Date of AI Use Be Recorded?
Commercial AI services can change over time. Providers may update models, interfaces, system instructions, and other technical components.
Recording the date helps establish when a particular interaction occurred.
For example, a researcher may use an AI model to classify survey responses in March and return to the same application several months later, when the service may operate differently.
The date can help explain the historical context of the interaction.
However, the date alone does not establish the exact underlying model version.
The issue of recording AI-use dates should therefore be considered alongside model identification and preservation of actual outputs.
Should the Log Identify the AI Model Version?
Where the information is available and relevant, recording the model identifier can improve documentation.
Different models may produce different classifications, code, summaries, or interpretations from similar instructions.
However, some applications do not display the exact underlying model or may change configurations without providing detailed version information.
Researchers should record what they can substantiate rather than infer a version from the current interface.
Where the precise model is unknown, the log can identify the application and indicate that the model version was unavailable.
False precision may be more misleading than an explicitly documented limitation.
Should Every AI Prompt Be Included?
Not necessarily. The value of preserving prompts depends on whether they influenced the research methodology or consequential decisions.
For example, a prompt defining the coding categories for qualitative data may be important for reconstructing the analysis.
A request to correct a punctuation error is unlikely to have the same methodological significance.
Researchers may record consequential prompts directly or provide a reference to a securely stored interaction record.
Where prompts changed during analysis, important revisions should be documented if they influenced the resulting outputs.
The distinction between saving AI prompts and recording every incidental interaction helps keep documentation proportionate.
Should the Log Preserve AI-Generated Responses?
Consequential AI outputs may need to be preserved to explain how the research developed.
Examples include generated analytical code, preliminary classifications, suggested themes, and outputs used as experimental materials.
Preserving the actual output can be important because repeating the same prompt may not reproduce an identical response.
However, a log does not necessarily need to contain the complete output within each entry.
It may instead reference a securely stored file, code repository, or analytical record.
The relevant requirement is that the record can be located and connected to the research decision it influenced.
Why Should Human Decisions Be Recorded?
AI-generated suggestions do not explain how researchers evaluated or incorporated them.
A useful log should therefore record whether consequential outputs were accepted, revised, or rejected.
For example, an AI system may suggest a statistical test that the researcher rejects after examining the study design.
Another system may generate qualitative codes that are modified because they overlook important contextual information.
Recording these decisions helps explain the researcher's analytical reasoning.
The log should not imply that accepting an output automatically establishes its validity. The basis for consequential decisions may also need to be documented.
Should Verification Procedures Be Included?
Yes, particularly when AI-generated material affects research findings or substantive manuscript claims.
The log may identify the verification performed, the relevant evidence, and the person responsible for checking the output.
For example, a researcher using AI-generated statistical code may record that the script was inspected, tested against known cases, and compared with independently calculated results.
A researcher using AI-assisted literature summaries may record that claims were checked against the original publications.
However, the log should describe what actually occurred.
Statements such as "verified by researcher" may provide limited information if they do not identify the nature of the verification.
The separate issue of documenting AI-output verification concerns the evidence supporting these claims.
Should Rejected AI Suggestions Be Logged?
Not every rejected suggestion needs a formal entry.
However, consequential rejected recommendations may be worth documenting when they help explain the development of the research.
For example, a researcher may reject an AI-generated causal interpretation because the study uses an observational design.
Recording the decision can demonstrate why the final manuscript adopted a more limited conclusion.
Similarly, rejecting an AI-proposed analytical procedure may help explain methodological choices.
The relevant question is whether the rejection forms part of the research's defensible decision history.
The log should not become an exhaustive archive of every unhelpful response.
Should Uploaded Files Be Recorded?
Where AI processes research materials, recording the types of files or datasets supplied can support data governance and methodological transparency.
For example, a log may identify that de-identified survey responses were processed for preliminary coding.
However, the log should not unnecessarily reproduce sensitive participant information.
Researchers should record sufficient information to identify the relevant input materials while complying with confidentiality and access restrictions.
Where files contain personal or restricted data, the record may also need to identify the authorization or approved processing arrangement.
The question of recording files uploaded to AI tools is particularly important when external services process research data.
Should the Log Include AI Tool Settings?
Settings may matter when they influence outputs or are necessary to understand the computational procedure.
Depending on the tool, relevant information may include temperature, output limits, sampling parameters, or other configurable options.
However, consumer AI applications may not expose every technical setting.
Researchers should document available and relevant settings without inventing hidden values.
Where the system configuration is central to the methodology, unavailable settings may represent a limitation that should be acknowledged.
What Is the Difference Between an AI-Use Log and an AI Disclosure Statement?
An AI-use log is primarily a research record. An AI disclosure statement communicates information required by a journal, institution, or other authority.
AI-use log
Documents relevant interactions, outputs, decisions, verification, and supporting records across the research workflow.
AI disclosure statement
Summarizes reportable AI assistance according to the applicable publication or institutional requirements.
A log may contain more detail than the published disclosure, particularly when prompts or outputs include confidential information.
Conversely, the existence of a log does not automatically satisfy a journal's requirement to disclose AI assistance.
The two serve related but distinct purposes.
Can an AI Conversation History Replace an AI-Use Log?
Not necessarily. Conversation histories preserve some prompts and responses but may omit important research context.
For example, a chat transcript may show that an AI system suggested a statistical procedure without documenting whether the researcher accepted it, modified it, or used it in the final analysis.
It may also omit independent verification performed outside the AI application.
Some services may delete, modify, or limit access to conversation records over time.
A structured log can connect relevant interactions to the research decisions and supporting evidence.
Conversation histories may therefore complement the log rather than replace it.
Should Researchers Use a Spreadsheet for the Log?
A spreadsheet can be a practical format because it allows researchers to record activities consistently and link entries to supporting materials.
However, a spreadsheet is not inherently superior to an electronic laboratory notebook, institutional research system, or other approved recordkeeping arrangement.
The best format depends on the project, team size, data sensitivity, and existing research workflow.
Researchers should prioritize consistency, accessibility to authorized team members, and appropriate preservation.
Where records are sensitive, access controls and institutional storage requirements may be more important than the particular file format.
How Detailed Should Each Log Entry Be?
The level of detail should reflect the significance of the AI assistance.
A minor language-editing activity may need only a brief record where documentation is required.
AI-generated code used to produce the primary statistical findings may require references to prompts, scripts, tests, and analytical outputs.
Similarly, AI-assisted qualitative coding may require documentation of the coding framework, generated suggestions, human revisions, and relevant analytical memos.
The goal is to preserve enough information to reconstruct consequential decisions without documenting every interaction at the same level.
A useful log should be proportionate rather than exhaustive.
Can an AI-Use Log Help During a Research Integrity Investigation?
Potentially. Contemporaneous records may help establish what AI assistance occurred and how researchers evaluated its outputs.
For example, a log may show that an AI-generated reference was checked against the original publication before submission.
It may also help identify when a problematic interpretation entered the manuscript.
However, the existence of a log does not automatically prove that the research was conducted responsibly.
Records must be accurate, authentic, and consistent with other available evidence.
A retrospective log reconstructed from memory should not be presented as though it was maintained contemporaneously.
Can Researchers Create the Log After Completing the Study?
A retrospective record may still be useful, but its limitations should be recognized.
Researchers may reconstruct AI activities from saved conversations, manuscript drafts, code repositories, and other records.
However, missing information should not be invented or presented with unsupported precision.
For example, a researcher should not assign a model version to an earlier interaction solely because that version is currently available.
Where entries are reconstructed retrospectively, the record should distinguish documented facts from recollections or estimates.
Maintaining the log during the research process is generally more reliable than attempting to reconstruct every consequential interaction months later.
How Should AI-Use Logs Be Protected?
AI-use logs may contain prompts, outputs, file references, or descriptions of confidential research activities.
Researchers should therefore apply appropriate access controls, storage arrangements, and retention practices.
Where logs reference participant information, institutional policies and research ethics requirements may impose additional safeguards.
Researchers should avoid placing confidential materials in an unrestricted shared spreadsheet merely for convenience.
The log should preserve sufficient information for accountability without creating unnecessary privacy or security risks.
How Long Should an AI-Use Log Be Retained?
Retention depends on institutional policies, funder requirements, research agreements, and the significance of the AI records.
Where AI assistance materially contributes to published findings, relevant records may need to be retained with the broader research archive.
However, there is no universal retention period specifically applicable to every AI-use log.
Researchers should follow the governing research-records policy and avoid assuming that all AI interactions must be retained indefinitely.
The separate question of retaining AI-use records with the research archive concerns both preservation and appropriate disposal.
