03 · What You Need to Know
How to Decide Which AI-Use Records Belong in the Research Archive
What Is a Research Archive?
A research archive is an organized collection of records preserved to document the conduct, evidence, procedures, and outcomes of a research project.
Depending on the discipline and institutional requirements, it may include datasets, research instruments, consent documentation, laboratory records, analytical code, methodological decisions, and other supporting materials.
Not every record must necessarily be stored in the same folder or repository.
Some materials may require restricted access, while others may be suitable for public sharing.
The important consideration is that records remain identifiable, appropriately protected, and available for the period required by the governing policy.
AI-use records may become part of this archive when they document consequential research activities.
When Does an AI Interaction Become a Research Record Worth Retaining?
An AI interaction may warrant archival retention when it contributes materially to the evidence, methods, or decisions underlying the study.
For example, a language model used to classify open-ended survey responses may generate a preliminary coding matrix that influences the final analysis.
The original classifications, coding instructions, and human revisions may help explain how the findings were developed.
By contrast, an incidental request to improve punctuation may have little methodological significance.
The distinction depends on the role of the interaction rather than the application used or the number of messages exchanged.
Researchers should also consider whether applicable policies explicitly require retention of particular categories of AI records.
Which AI-Use Records Are Most Likely to Belong in the Archive?
| AI-Use Record |
Potential Archival Value |
Retention Consideration |
| AI-generated research data |
Constitutes observations or materials analyzed in the study. |
Retain according to applicable research-data requirements. |
| Consequential prompts |
Document instructions that shaped research procedures or outputs. |
Preserve where necessary for methodological reconstruction. |
| AI-generated classifications |
May influence coding decisions and analytical findings. |
Connect original outputs to final reviewed data. |
| AI-generated analytical code |
Documents computational assistance and software development. |
Preserve relevant provenance and final tested scripts. |
| AI-generated experimental stimuli |
Represent materials presented to research participants. |
Retain actual stimuli and relevant generation records. |
| AI-output verification records |
Document checks supporting consequential research claims. |
Retain evidence needed to understand validation. |
| Methodological decision records |
Explain accepted, revised, or rejected AI recommendations. |
Preserve significant analytical reasoning. |
| Routine language-editing interactions |
May have limited relevance to research procedures. |
Follow applicable policy; extensive retention may be unnecessary. |
These categories provide a practical framework rather than a universal retention schedule.
Does Every AI Conversation Need to Be Archived?
No universal requirement obliges researchers to preserve every AI conversation.
Some interactions may have no material relationship to the final research.
For example, a researcher may use ChatGPT to explore several possible article titles without adopting any suggestions.
Such exchanges may not warrant the same archival treatment as AI-generated classifications used in the analysis.
However, researchers should not delete conversations merely because their outputs were revised or excluded from the final manuscript.
Some rejected suggestions may document consequential decisions.
The appropriate retention decision should reflect the interaction's significance and any governing requirements.
What Is the Difference Between Retaining AI Records and Disclosing AI Use?
AI disclosure communicates information about reportable assistance to journals, institutions, or readers.
Archival retention preserves records needed for later examination of the research.
AI disclosure
Identifies and explains AI assistance according to applicable publication or institutional requirements.
AI archival retention
Preserves relevant evidence of AI interactions, outputs, verification, and decisions for the required research-records period.
A published declaration may state that AI assisted with qualitative coding without including the original coding instructions or generated classifications.
Those underlying records may still be important to the research archive.
Conversely, preserving a complete AI-use log does not automatically satisfy a journal's disclosure requirements.
Do ICMJE and COPE Require Every AI Conversation to Be Archived?
The International Committee of Medical Journal Editors (ICMJE) emphasizes disclosure of AI-assisted technologies and human responsibility for submitted work.
The Committee on Publication Ethics (COPE) similarly emphasizes transparency and accountability concerning AI assistance.
These positions do not establish a universal requirement to retain every AI conversation for a fixed period.
Research-record retention is also governed by institutional policies, funder conditions, disciplinary standards, and applicable legal obligations.
Researchers should therefore avoid treating general publication-integrity guidance as a complete AI archival policy.
The appropriate retention requirements must be established for the particular research project.
Should AI-Generated Research Data Be Retained?
Where AI-generated outputs constitute research data, they should be managed according to the applicable data-management and retention requirements.
For example, a study may systematically evaluate the accuracy of AI-generated educational feedback.
The generated feedback is part of the observations analyzed in the study.
Preserving the actual responses may be necessary because repeating the same prompts may not reproduce identical outputs.
The archive may also need relevant model information, prompts, settings, and collection dates.
The researcher should distinguish AI-generated research data from incidental assistance that does not form part of the study's evidence.
Should Original AI Outputs Be Retained After Human Revision?
Where the original output materially influenced the research, retaining it alongside the revised version may be useful or required.
For example, an AI-generated coding matrix may contain classifications that the researcher subsequently corrects.
The final matrix documents the data used for analysis, while the original output and revision records help explain how the classifications developed.
Similarly, a generated statistical script may be substantially modified before execution.
The final tested script is central to computational reproducibility, but the original generated code may provide relevant provenance.
The appropriate retention should reflect the importance of the development history and applicable research-records requirements.
Should Researchers Retain Rejected AI Suggestions?
Consequential rejected suggestions may form part of the research decision record.
For example, a researcher may reject an AI-generated causal interpretation because the study design supports only an association.
Preserving the suggestion and the reason for rejecting it may help explain the final interpretation.
However, not every unused wording suggestion or unsuccessful exploratory prompt requires archival retention.
The question of documenting rejected AI suggestions concerns whether the rejection contributes meaningfully to reconstructing the research process.
Should AI-Use Logs Be Retained With the Research Archive?
Where an AI-use log documents consequential research activities, retaining it with the archive or maintaining a reliable link to it may support long-term traceability.
The log can identify the AI tool, task, date, outputs, human decisions, and supporting records.
However, a log does not necessarily need to reproduce every prompt or generated response.
It may instead reference securely preserved materials.
The broader research AI-use log should remain interpretable and connected to the final research files.
A log that refers to deleted or inaccessible materials may provide limited value for later verification.
How Long Must AI-Use Records Be Retained?
There is no universal retention period that applies to every AI-use record across all research institutions and jurisdictions.
Some research-records policies specify minimum retention periods following publication, project completion, or another defined event.
Other obligations may require longer retention for particular categories of research data, regulated studies, or contractual arrangements.
Researchers should determine which policy governs the project and when the retention period begins.
They should not assume that publication automatically authorizes deletion of supporting AI records.
Where multiple requirements apply, the institution's research administration or records-management office may need to clarify how they interact.
Does Publication Mean AI Records Can Be Deleted?
Not automatically. Publication is not necessarily the end of the applicable research-record retention period.
Records may need to remain available for verification, corrections, institutional review, or other authorized purposes.
For example, a journal may ask for additional methodological information after publication if questions arise about AI-assisted analysis.
If the underlying prompts, outputs, and verification records were deleted prematurely, the researcher may be unable to reconstruct the procedure.
However, indefinite retention is not universally required.
Disposal should follow the applicable retention schedule and any relevant legal or institutional restrictions.
Should AI Records Be Retained for the Same Period as Research Data?
Often, consequential AI records may appropriately follow the retention requirements applicable to the research data or methodological records they support.
However, not every AI-related record necessarily belongs to the same retention category.
For example, AI-generated experimental stimuli may constitute research materials, while routine manuscript-editing conversations may be treated differently.
Records containing personal information may also be subject to specific data-protection and disposal requirements.
Researchers should classify records according to their function and the governing policies rather than automatically assign one retention period to every AI interaction.
What If Several Retention Requirements Apply?
A project may be subject to institutional, funder, contractual, journal, and legal requirements.
These obligations may differ in scope, duration, or permitted handling of records.
Researchers should not assume that one policy automatically overrides all others.
For example, a funder may require preservation of certain research data, while privacy obligations restrict unnecessary retention of identifiable information.
The appropriate arrangement may involve controlled access, de-identification where permitted, or separate retention schedules for different record categories.
Where requirements appear inconsistent, researchers should seek guidance from the appropriate institutional offices.
What If AI Records Contain Personal or Confidential Information?
Records containing participant data, unpublished manuscripts, institutional information, or other restricted materials require appropriate safeguards.
Researchers should determine whether the information may be retained and under what conditions.
For example, an AI conversation containing interview excerpts may need to remain in restricted institutional storage rather than a publicly accessible repository.
Removing names alone may not eliminate indirect identification risks.
Retention should therefore be consistent with research ethics approval, consent arrangements, institutional policies, and applicable data-protection law.
Archival preservation does not justify keeping personal information longer than authorized.
Does the Data Privacy Act Affect AI Record Retention in the Philippines?
Potentially. Republic Act No. 10173, the Data Privacy Act of 2012, establishes requirements relevant to the processing of personal information.
Where AI-use records contain personal information, researchers and institutions should consider applicable principles concerning legitimate purpose, proportionality, security, and retention.
For example, retaining identifiable participant transcripts within an AI conversation archive may require a different assessment from preserving a de-identified analytical summary.
Researchers should consult institutional data-protection personnel and relevant National Privacy Commission guidance where necessary.
The existence of a research-record retention requirement does not eliminate obligations concerning appropriate access, security, and lawful processing.
Should AI Conversations Be Stored in the Same Folder as Research Data?
Not necessarily. The appropriate storage arrangement depends on the content and sensitivity of the records.
Some AI interactions may contain confidential participant information requiring restricted access.
Other records, such as prompt templates or non-sensitive generated code, may be suitable for broader sharing.
The archive should maintain reliable references between related materials even when they are stored separately.
For example, an AI-use log may link to a restricted conversation archive and a separately preserved final analytical script.
Logical connection is more important than physical storage in one folder.
Can AI Conversation Links Be Used Instead of Archiving the Records?
Provider-hosted conversation links may be useful references but should not automatically be treated as durable research preservation.
Access may depend on the provider's retention policies, account arrangements, and continued availability of the service.
Consequential records may therefore require independent preservation under authorized institutional arrangements.
The question of preserving AI interactions when providers may change or delete them is particularly important for long-term research archives.
Researchers should avoid assuming that a currently accessible link will remain available throughout the required retention period.
Should Researchers Preserve the Exact Model Version?
Where AI materially contributes to the methodology or findings, model identification may be important to the research record.
Different model versions can produce different outputs from similar prompts.
However, some commercial services do not disclose precise underlying model identifiers.
Researchers should preserve the information available at the time of use without inventing missing technical details.
The model record should remain connected to the relevant prompts, outputs, and research activities.
Preserving a model name alone does not guarantee reproducibility.
What About AI Settings and Configuration Files?
Relevant settings may need preservation when they influence consequential AI-generated outputs.
For example, an API-based research procedure may use specified sampling parameters, output limits, or structured-response configurations.
These settings can help explain how the outputs were produced.
However, not every AI application exposes the same controls.
Researchers should preserve known settings and acknowledge unavailable configurations where relevant.
Configuration files should remain connected to the corresponding model, prompts, and outputs.
Should Verification Records Be Retained?
Where AI-generated material materially influences research findings, verification records may be important supporting evidence.
For example, an AI-generated citation may be linked to notes confirming that the original publication was examined and that the cited claim was supported.
An AI-generated statistical script may be accompanied by tests and methodological checks.
AI-assisted qualitative classifications may be linked to human-reviewed coding matrices and analytical memos.
These records help explain how the researcher evaluated the outputs.
The question of documenting AI-output verification is therefore closely connected to archival retention.
What If AI Records Are Needed During a Research Integrity Investigation?
Relevant records may need to be preserved when a research-integrity inquiry, investigation, litigation hold, or other authorized preservation obligation is pending.
Researchers should not dispose of potentially relevant records merely because an ordinary retention period has expired if a valid hold applies.
The applicable institution or authority should determine the scope and duration of the preservation obligation.
AI conversations, analytical outputs, code, and decision records may be relevant depending on the allegations or questions under examination.
However, the presence or absence of a particular AI record does not automatically establish misconduct.
Research-integrity assessments require consideration of the broader evidence and governing standards.
Can AI Records Be Deleted After the Retention Period Ends?
Potentially, subject to the applicable records-management, legal, ethical, and institutional requirements.
Researchers should confirm that the required retention period has ended and that no preservation hold or other obligation remains.
Where disposal is authorized, it should be performed through appropriate procedures, particularly for records containing personal or confidential information.
Deletion from a local folder may not remove copies held in backups, repositories, or external services.
Researchers should follow the institution's approved disposal process rather than assume that deleting a visible file completes every required action.
What If the AI Provider Deletes Its Copy Before the Research Retention Period Ends?
Researchers should not rely exclusively on the provider to preserve consequential research interactions.
Where authorized, relevant prompts, outputs, and metadata should be preserved independently while accessible.
If the provider's copy becomes unavailable, the institutionally retained record may still support examination of the original procedure.
However, independent preservation does not guarantee that the original model can later be accessed or that identical outputs can be regenerated.
The archive preserves historical evidence rather than the AI provider's complete technical environment.
How Should Researchers Document Archival Decisions?
A research archive may benefit from a concise record explaining which AI materials were retained, where they are stored, and which retention requirements apply.
For example, an archive inventory may identify the AI-use log, consequential conversation exports, generated analytical code, and verification records.
It may also identify restricted-access materials and the responsible custodian.
However, the inventory should not unnecessarily reproduce confidential content.
The objective is to ensure that authorized personnel can locate and interpret the records during the required retention period.
How Can Researchers Determine Whether an AI Record Is Safe to Dispose Of?
A practical assessment should consider whether the record is required by policy, materially supports the findings, or remains subject to a preservation obligation.
For example, a discarded AI suggestion with no influence on the study may have limited archival value.
By contrast, the original AI-generated coding matrix used to develop published findings may be necessary to reconstruct the analysis.
Researchers should also consider whether the record contains sensitive information that should not be retained unnecessarily.
The decision should follow the applicable retention schedule rather than rely solely on personal preference.
What Makes an AI Research Archive Adequate?
An adequate archive should allow authorized reviewers to identify consequential AI assistance and examine its relationship to the final research.
For example, the records should make it possible to determine which AI tool was used, which input materials were processed, what outputs were generated, and how researchers evaluated them.
The archive should also preserve the relevant methodological context and comply with applicable access and retention requirements.
However, adequacy does not mean retaining every conversation indefinitely.
The broader question of how much AI-use documentation is enough should be answered in relation to the significance of the assistance and the purposes the records must serve.
