03 · What You Need to Know
Why Records of AI Use Matter for Research Integrity
What Counts as a Record of AI Use?
A record of AI use is documentation showing how an artificial intelligence system contributed to a research activity.
Depending on the activity, it may include the AI tool's name, interaction date, prompts, generated responses, uploaded materials, model information, analytical decisions, and verification procedures.
For example, a researcher who uses AI to classify survey responses might preserve the coding instructions, the generated classification matrix, and notes explaining how the classifications were reviewed.
A researcher who uses AI to generate statistical code might retain the relevant generated script, the final tested version, and evidence of validation.
These records need not always be stored as complete conversation transcripts. They may be incorporated into existing research notebooks, analytical files, or approved recordkeeping systems.
Why Is AI Disclosure Different From AI Documentation?
Disclosure communicates the nature of AI assistance to editors, reviewers, readers, or other relevant parties.
Documentation preserves information about the actual research process.
AI disclosure
Explains what AI assistance was used and where it contributed, according to applicable publication or institutional requirements.
AI documentation
Preserves evidence of consequential AI interactions, outputs, verification procedures, and subsequent research decisions.
For example, a disclosure may state that ChatGPT assisted with generating preliminary R code.
That statement does not necessarily identify which code was generated, whether the researcher modified it, or how the final analysis was tested.
Those details may be preserved in the research records even when they are not reproduced in the manuscript.
Are Researchers Required to Keep Records of Every AI Interaction?
No universal requirement applies to every research project or AI-assisted activity.
The International Committee of Medical Journal Editors (ICMJE) recommends disclosure of AI-assisted technologies and emphasizes that human authors remain responsible for the accuracy and integrity of submitted work.
The Committee on Publication Ethics (COPE) similarly emphasizes transparency concerning AI assistance and human accountability.
However, these positions should not be interpreted as requiring researchers to archive every AI conversation indefinitely.
Specific institutions, funders, journals, or research agreements may impose additional documentation requirements.
Researchers should distinguish formal obligations from recommended practices that support methodological transparency.
Why Might AI-Use Records Be Important for Reproducibility?
Generative AI systems may produce different outputs from similar prompts, even when researchers use the same named application.
Differences can arise from model versions, settings, conversation context, input materials, or provider-side changes.
For example, a researcher who uses AI to classify open-ended responses may obtain different category assignments when repeating the procedure several months later.
Preserving the original prompts and classifications helps establish what occurred during the initial analysis.
However, documentation does not guarantee that the original output can be reproduced exactly.
Researchers should distinguish the ability to examine a historical procedure from the ability to regenerate identical results.
Which AI-Assisted Activities Are Most Important to Document?
The significance of AI assistance should guide the amount of documentation preserved.
| AI-Assisted Activity |
Potential Research Significance |
Documentation Consideration |
| Routine grammar correction |
Usually affects manuscript presentation rather than research procedures. |
Follow applicable policy; detailed interaction records may be unnecessary. |
| Research brainstorming |
May influence research questions or conceptual decisions. |
Document consequential suggestions that were adopted. |
| Literature searching |
May influence which publications are identified and considered. |
Preserve relevant search procedures and verified sources. |
| Study screening |
May determine which evidence enters a review. |
Document screening instructions, decisions, and verification. |
| Statistical code generation |
May affect calculations and reported findings. |
Preserve final tested code, relevant provenance, and validation records. |
| Qualitative coding |
May influence categories, themes, and interpretation. |
Preserve consequential prompts, outputs, and analytical revisions. |
| Research data processing |
May alter datasets or analytical results. |
Record relevant inputs, transformations, and verification. |
| AI-generated research stimuli |
May influence participants' responses. |
Preserve the actual materials and relevant generation procedures. |
These are practical considerations rather than universally mandated documentation categories.
Should Researchers Record the AI Tool and Model Used?
Where AI materially influences the research, identifying the tool and available model information can improve methodological documentation.
For example, a researcher may record ChatGPT as the application and OpenAI as the provider, together with the model identifier displayed at the time of use.
However, an application name does not necessarily identify the exact underlying model.
Commercial services may provide access to multiple models or change their implementations over time.
Researchers should record the information available during the interaction rather than infer historical details from the current interface.
If the exact model version is unavailable, the limitation should be documented honestly.
The separate question of recording AI model versions becomes particularly important when the system materially influences research findings.
Should Researchers Record the Date of AI Use?
Recording the date may help establish the chronological context of consequential AI interactions.
For example, an AI-assisted analysis conducted in January may have used a different service configuration from one repeated in September.
The date can help researchers interpret historical records and investigate discrepancies.
However, the date alone does not establish which model or technical configuration was used.
Researchers should preserve available model information and relevant outputs where appropriate.
Not every minor writing-assistance interaction requires an exact timestamp unless an applicable policy demands it.
The importance of recording AI-use dates depends on the role of the interaction in the research.
Should Researchers Save the Prompts They Used?
Consequential prompts may be important because they define the instructions supplied to an AI system.
For example, a prompt used to classify interview excerpts may contain the coding framework, category definitions, and rules for ambiguous responses.
These instructions can influence the resulting classifications.
Preserving them may help another researcher understand how the procedure was conducted.
However, not every exploratory or language-editing prompt necessarily requires preservation.
Researchers should prioritize prompts that materially shape research procedures or outputs.
The question of saving AI prompts used in research should be considered alongside the significance of the generated material.
Should AI-Generated Responses Be Preserved?
Where generated responses materially influence the research, preserving the original outputs may be valuable.
For example, an AI-generated coding matrix may contain classifications subsequently corrected by the researcher.
Retaining both versions can help explain how the final analysis developed.
Likewise, an AI-generated statistical script may differ substantially from the code ultimately executed.
Preserving the relevant original output may support provenance documentation, while the final tested code supports computational reproducibility.
The broader issue of archiving AI-generated research responses concerns preserving evidence of what the system actually produced.
Should Researchers Record Which Files Were Uploaded to AI?
Where AI processes research files, recording the input materials can support methodological transparency and data governance.
For example, a researcher may need to identify which version of a survey dataset was submitted for AI-assisted analysis.
Different versions may contain different observations or corrections, affecting the resulting outputs.
File records may also help establish whether confidential materials were transferred to an authorized service.
However, documentation of an upload does not establish that the transfer was permitted.
Researchers must comply with applicable institutional, ethics, confidentiality, and data-protection requirements.
The question of recording files uploaded to AI tools is particularly important when participant or institutional data are involved.
Should Researchers Record How AI Outputs Were Verified?
Yes, particularly when AI-generated material influences substantive claims, analysis, or findings.
A record of verification should explain what was checked and how the researcher evaluated the output.
For example, an AI-generated reference may be checked against the original publication to confirm both its existence and its relevance to the claim.
AI-generated statistical code may be inspected and tested against known results.
AI-assisted qualitative classifications may be compared with original participant responses.
However, statements such as "all outputs were verified" should not be used unless the claimed verification actually occurred.
The importance of documenting AI-output verification lies in preserving evidence of the checks performed rather than merely asserting human oversight.
Should Researchers Record AI Suggestions They Rejected?
Consequential rejected suggestions may be worth documenting when they explain important methodological or interpretive decisions.
For example, an AI system may recommend an unsuitable statistical test that the researcher rejects after examining the study design.
Preserving the recommendation and the reason for rejection may help explain the final analytical choice.
Similarly, a researcher may reject an AI-generated qualitative theme because it misrepresents participants' experiences.
However, not every rejected wording suggestion or unsuccessful exploratory interaction needs a permanent record.
The question of documenting rejected AI suggestions concerns consequential scholarly reasoning rather than exhaustive archiving.
What Is an AI-Use Log?
An AI-use log is a structured record connecting AI-assisted activities to relevant outputs, verification, and research decisions.
It may be maintained in a spreadsheet, electronic research notebook, or institutionally approved documentation system.
A typical entry may identify the date, AI tool, research task, consequential output, and human decision.
Additional fields may include model information, prompts, input files, and verification references.
However, a separate AI-use log is not universally required.
Existing research notebooks, analytical memos, code repositories, or data-management systems may already preserve the relevant information.
The practical details of what an AI-use log should contain depend on the methodology and applicable requirements.
Can AI Conversations Become Part of the Research Audit Trail?
Yes. AI conversations may preserve consequential prompts, responses, and revisions.
However, a conversation history does not necessarily explain which suggestions were adopted or how they were verified outside the AI application.
For example, a chat transcript may show that AI generated statistical code without establishing whether the researcher executed that code or modified it before analysis.
A useful audit trail connects the interaction to the final tested script and relevant verification records.
The broader issue of AI conversations as part of a research audit trail concerns this connection between technological assistance and actual scholarly decisions.
Is Saving ChatGPT Conversation History Enough?
Not necessarily. A saved conversation may document what the AI system generated but omit important information about the researcher's subsequent actions.
For example, the researcher may reject an AI-generated interpretation after examining the original data.
If that decision occurs outside the AI conversation, the chat history alone may not explain the final analytical reasoning.
Likewise, generated code may be revised and tested in a separate programming environment.
Researchers should therefore connect relevant conversations to other research records.
The goal is a coherent account of the research process rather than a collection of disconnected transcripts.
Can AI Providers Delete or Change Conversation Records?
Access to AI conversation histories depends on the provider's service, account settings, retention policies, and technical arrangements.
Some services may change interfaces, remove access to models, or provide temporary interaction modes.
Researchers should not assume that provider-hosted conversations will remain available indefinitely.
Where an interaction materially influences the research, an authorized independent copy may provide more reliable preservation.
However, exporting a conversation does not necessarily preserve every hidden instruction or technical detail affecting the original output.
The question of preserving AI interactions when providers may delete or change them is relevant to long-term research documentation.
Should AI-Use Records Be Created During the Research or Afterward?
Contemporaneous documentation is generally more reliable than reconstructing interactions months later.
For example, a researcher may be able to record the exact model label, prompt, and output immediately after an AI-assisted analysis.
Later, the interface may have changed or the original conversation may no longer be accessible.
Researchers may still reconstruct missing information from saved drafts, code histories, or other records.
However, retrospective documentation should distinguish verified information from uncertain recollections.
Missing details should not be invented merely to make the record appear complete.
Can Researchers Create an AI-Use Record Retrospectively?
Yes, where existing evidence supports reconstruction of the activity.
For example, a researcher may identify an AI-generated script in a code repository and locate the corresponding prompt in an exported conversation.
These materials may support a retrospective record explaining the interaction.
However, the researcher should not claim that the record was maintained contemporaneously if it was created later.
Nor should an unavailable model version be inferred solely from the model currently offered by the application.
Accurate acknowledgment of documentation limitations is preferable to unsupported precision.
Do AI-Use Records Need to Be Publicly Available?
Not universally. Internal documentation and public disclosure are separate considerations.
Some AI records may be suitable for supplementary materials or research repositories, particularly when they constitute research data or are necessary for methodological evaluation.
Others may contain confidential participant information, unpublished manuscripts, or proprietary material.
Researchers should follow applicable ethics, institutional, legal, and journal requirements before sharing them.
Controlled-access arrangements or appropriately documented redacted records may be necessary in some circumstances.
Preserving a record does not automatically authorize its public release.
Can AI-Use Records Contain Confidential Information?
Yes. Prompts and outputs may include participant quotations, unpublished findings, institutional records, or other restricted materials.
Researchers should determine whether the original AI processing was authorized and whether the resulting records may be retained.
For example, a conversation containing identifiable interview excerpts may require restricted storage and access controls.
Removing names does not necessarily eliminate indirect identification risks.
Researchers should avoid creating unnecessary copies of sensitive information merely to demonstrate transparency.
Recordkeeping should remain consistent with confidentiality, research ethics, and data-protection obligations.
Does Keeping an AI-Use Record Prove That the AI Use Was Ethical?
No. Documentation records what occurred but does not independently establish that the activity was permitted.
For example, a researcher may accurately document uploading participant transcripts to an external AI service without having obtained the necessary authorization.
The record does not retrospectively justify the transfer.
Similarly, preserving an AI-generated image does not establish that the intended journal permits its publication.
Researchers should distinguish accurate documentation from compliance with ethical, institutional, contractual, and publisher requirements.
Where an unauthorized activity may have occurred, the appropriate institutional procedures should be followed.
Can AI-Use Records Help During Peer Review?
Potentially. Reviewers may request clarification about AI-assisted methods, analytical code, or generated research materials.
For example, a reviewer may ask how AI-assisted coding was performed and whether the classifications were checked against the original responses.
Preserved prompts, outputs, and verification notes can help the researcher provide an accurate explanation.
However, the existence of records does not establish that the methodology was valid.
The researcher must still justify the analytical approach and demonstrate that the procedures were appropriate for the research question.
Can AI-Use Records Support a Research Integrity Investigation?
Potentially. Contemporaneous records may help establish which AI tools were used, what outputs were generated, and how researchers evaluated them.
For example, a preserved interaction may show that an AI-generated citation was identified as incorrect and excluded before submission.
However, records must be assessed for authenticity, completeness, and consistency with other evidence.
A saved conversation does not automatically prove responsible conduct, and missing records do not automatically establish misconduct.
Research integrity investigations should consider the applicable standards and the broader evidentiary record.
How Long Should AI-Use Records Be Kept?
Retention depends on the applicable institutional research-records policy, funder requirements, research agreements, and legal obligations.
Where AI materially contributes to published findings, relevant records may need to be retained with the broader research archive.
However, no universal retention period applies specifically to every AI interaction.
Researchers should distinguish consequential records from incidental exchanges and follow authorized disposal procedures.
The separate question of retaining AI-use records with the research archive
How Much AI Documentation Is Enough?
The appropriate amount depends on whether the records adequately explain consequential AI assistance and satisfy applicable requirements.
A researcher who used AI only for minor language refinement may not need the same level of documentation as one who used AI to classify research data.
For consequential activities, the records should identify what the system received, what it generated, and how the researcher evaluated the output.
However, preserving every conversation does not necessarily produce a complete or useful audit trail.
Researchers should prioritize meaningful evidence and maintain connections between AI interactions and final research decisions.
The broader question of how much documentation of AI use is enough should be answered in relation to methodological significance rather than the number of saved messages.
What If a Journal Does Not Require an AI-Use Log?
The absence of a journal-specific logging requirement does not necessarily eliminate the value of documentation.
Research-records obligations may arise independently from institutional policies, funder conditions, or methodological standards.
For example, AI-assisted qualitative coding may require adequate analytical documentation even when the journal asks only for a concise AI disclosure.
However, researchers should not invent a universal obligation to preserve every interaction.
The appropriate approach is to document consequential assistance according to the research design and applicable requirements.
What Should Researchers Do if They Have No AI-Use Records?
Researchers should first identify whether the missing information concerns consequential research activities.
Where relevant, they may examine saved conversations, manuscript drafts, code repositories, analytical files, and other contemporaneous records.
These materials may allow portions of the workflow to be reconstructed.
However, missing prompts, outputs, or model identifiers should not be invented.
If important information cannot be recovered, the researcher should acknowledge the limitation where it affects methodological reporting or verification.
Future projects may benefit from documenting consequential AI assistance while the interactions remain accessible.
