What Should Researchers Archive So Future Researchers Can Understand Exactly How AI Contributed?

What Should You Preserve So Someone Else Can Understand Your AI-Assisted Research?

A researcher publishes a study in which generative AI helped classify interview responses, generate statistical code, and identify alternative explanations for the findings.

The manuscript identifies the AI provider and states that all outputs were reviewed by the authors. Two years later, another research team attempts to understand how the classifications were produced.

The original AI model is no longer available. The researcher preserved the final dataset but not the classification instructions, original AI outputs, or records of human corrections. The statistical code remains available, although nobody can determine which parts were generated by AI or subsequently modified.

The published disclosure is accurate, but it provides too little information to reconstruct the analytical process.

What should researchers preserve while the project is still active so that future researchers can understand what AI contributed, how its outputs were evaluated, and which scientific decisions remained under human control?

Archive the Evidence Needed to Trace AI Contributions From Input to Final Research Output

What a Reconstructable AI-Assisted Research Archive Should Contain

Why Is an AI Disclosure Statement Not Enough?

A disclosure statement identifies AI involvement but may not explain how the system affected the research.

For example, stating that a language model assisted with qualitative coding does not identify the coding framework, instructions, data supplied, generated categories, or subsequent human revisions.

Future researchers may need these details to evaluate the analytical procedure.

Similarly, a statement that AI generated statistical code does not establish which script was executed or how its calculations were verified.

Disclosure and archiving serve different purposes. Disclosure communicates the contribution, while archiving preserves evidence that can support later inspection.

What Does It Mean to Reconstruct an AI-Assisted Research Process?

Reconstruction involves establishing how a reported research output was produced from the available inputs, procedures, and decisions.

For an AI-generated classification variable, reconstruction may require the original data, coding instructions, model information, generated labels, human corrections, and validation results.

For AI-generated statistical code, it may require the final script, software environment, input dataset, and verification records.

Reconstruction does not always mean reproducing the exact original AI response.

Some generative systems are stochastic, and their underlying models or configurations may change or become unavailable.

The archive should therefore preserve what actually occurred rather than depend entirely on the possibility of rerunning the original system.

Which AI Contributions Deserve Detailed Archiving?

Detailed archiving is particularly important when AI materially influences research methods, measurements, analytical results, or conclusions.

Examples include AI-generated classification variables, data-cleaning scripts, automated evidence extraction, qualitative coding suggestions used in final interpretation, and consequential methodological recommendations.

By contrast, a discarded suggestion for a manuscript heading may have little relevance to future scientific scrutiny.

Researchers should distinguish the importance of an interaction from its apparent technical complexity.

A short prompt that defines how thousands of responses are classified may be more consequential than a lengthy conversation about manuscript wording.

What Information Should Be Preserved for Each Consequential AI Task?

Not every task requires every element. Researchers should preserve the information necessary to understand the consequential procedure and comply with applicable research obligations.

Why Exact Prompts May Matter

Prompts can function as procedural instructions in AI-assisted research.

For example, a prompt may define classification categories, specify how ambiguous responses should be handled, or establish the format of generated analytical outputs.

Small wording changes can sometimes affect model behavior.

Where prompt content materially determines research measurements, preserving the exact instructions can support methodological evaluation.

However, a prompt alone may be insufficient because model versions, system instructions, generation settings, and external tools can also affect the output.

Researchers should preserve relevant context rather than assume that the prompt is a complete specification of the procedure.

Should Researchers Archive Original AI Outputs?

Original outputs may be important when future regeneration cannot be guaranteed.

For example, a model may assign 2,000 interview excerpts to predefined categories. Preserving only the final corrected labels would make it difficult to determine which decisions came from the model and which were changed by researchers.

Archiving the original classifications alongside the final dataset can support evaluation of human oversight and correction procedures.

For generated code, preserving the original script may help explain consequential revisions.

However, original outputs should be retained according to the project's information protection and retention requirements.

They should not be made public automatically when they contain confidential or restricted information.

Why Human Decisions Must Be Preserved

AI outputs do not explain why researchers accepted, modified, or rejected them.

A model may recommend excluding unusual observations, but the researcher may decide to retain them after examining the original data.

Without a decision record, future readers may be unable to determine whether the exclusion rule was considered or why it was rejected.

Relevant records may include analytical memos, code revisions, validation decisions, and explanations of methodological changes.

The aim is not to preserve every passing thought. It is to document decisions that materially affect the scientific process.

What Should Be Archived for AI-Generated Statistical Code?

When AI assists with statistical programming, the final validated code is a central research artifact.

Researchers should preserve the script actually executed, relevant software dependencies, input-data references, and instructions necessary to reproduce the downstream calculations where feasible.

Original AI-generated versions may also be relevant when substantial corrections occurred.

For example, a script may initially reverse-code a questionnaire item incorrectly. Preserving the corrected script and verification evidence is essential, while the original erroneous version may help explain the audit trail.

Researchers should distinguish the AI's contribution to writing code from the independently executed computational procedure.

Attempting to regenerate the script through a future AI model is generally not a substitute for preserving the code used in the original analysis.

What Should Be Archived for AI-Assisted Qualitative Analysis?

Qualitative research may require records of coding frameworks, AI-generated suggestions, original excerpts, researcher revisions, and analytical memos.

The appropriate archive depends on the qualitative methodology.

For structured content analysis, coding rules and reference judgments may be relevant.

For reflexive thematic analysis, documentation may emphasize interpretive development, researcher engagement, and the rationale for final themes.

Researchers should not assume that preserving a table of AI-generated codes is sufficient to explain the analytical process.

At the same time, confidentiality obligations may restrict access to original interview material.

Archives should support appropriate scrutiny without exposing participants unnecessarily.

How Should Researchers Preserve Model Information?

Researchers should record the model information available at the time of consequential processing.

This may include the provider, product, model identifier, available snapshot or version information, execution date, and relevant settings.

Some services expose detailed model identifiers, while others provide only a general product name.

Researchers should not invent version information when it is unavailable.

Instead, they should preserve the available evidence and acknowledge limitations in reconstructing the original system.

For long projects, a model change log can identify which versions were associated with particular analytical outputs.

Why Input Dataset Versions Matter

An AI output cannot be fully understood without identifying the information processed.

For example, a classification procedure applied to 500 responses may produce different results after duplicate removal, transcript correction, or exclusion of incomplete records.

Researchers should preserve dataset identifiers, relevant preprocessing rules, and version relationships.

Where appropriate, checksums or cryptographic hashes may help establish whether archived files have changed.

However, a hash identifies file integrity relative to a recorded value. It does not establish that the underlying data are scientifically valid or ethically obtained.

Researchers should connect input versions to the outputs and analyses derived from them.

How Can Researchers Preserve Relationships Between Research Artifacts?

An archive becomes more useful when its components can be traced to one another.

For example, an AI-generated classification file should be associated with the prompt and model used to produce it, the reference dataset used for validation, and the final analytical dataset.

Researchers may use a structured manifest or other documentation system to describe these relationships.

File names alone may be insufficient when several collaborators generate multiple versions of similar outputs.

A clear record should explain which artifacts are preliminary, superseded, validated, or included in the final study.

Traceability is especially important when errors are discovered after publication.

What Is the Difference Between Archiving and Public Sharing?

Researchers may be required to retain records without being permitted to publish them openly.

For example, identifiable interview transcripts may need secure institutional retention while remaining unavailable for public download.

Open-science principles do not eliminate confidentiality, consent, intellectual property, or legal obligations.

Researchers should distinguish what can be made publicly accessible from what must remain available only through authorized procedures.

How Do FAIR Principles Apply to AI Research Archives?

The FAIR Guiding Principles address the findability, accessibility, interoperability, and reusability of research data and related resources.

They provide a useful conceptual foundation for organizing AI-assisted research artifacts.

For example, a dataset or analytical script may benefit from descriptive metadata, persistent identifiers where appropriate, clear provenance, and documented reuse conditions.

FAIR does not mean that every dataset must be openly accessible.

Access can be restricted while metadata and governance arrangements support appropriate discovery and authorized reuse.

Researchers should apply FAIR principles consistently with confidentiality and institutional obligations.

What Should Researchers Do When Exact AI Reproduction Is Impossible?

Some AI systems may no longer be available when future researchers attempt to repeat the procedure.

Others may produce variable outputs even when the original prompt is reused.

Researchers should preserve the original outputs, available model information, relevant configurations, validation evidence, and downstream analytical artifacts.

These records may support methodological reconstruction and evaluation even when exact model-output repetition is impossible.

However, researchers should distinguish reconstructability from exact computational reproducibility.

An archive should not claim a level of reproducibility that the preserved evidence cannot support.

How Long Should AI Research Records Be Retained?

Retention requirements vary according to institutional policies, funder agreements, journal expectations, applicable law, and the type of research information.

There is no universal retention period that applies to every AI prompt, generated output, or analytical artifact.

Researchers should determine which records form part of the official research record and which retention schedules apply.

Where records contain personal information, retaining them indefinitely may conflict with applicable data protection principles.

The archive should balance scientific accountability with legitimate retention and deletion requirements.

Researchers should consult the responsible institutional offices when obligations differ across record categories.

Should Researchers Archive Every AI Conversation?

Not necessarily.

An indiscriminate archive of every interaction may create substantial storage and privacy burdens while making consequential decisions difficult to locate.

For example, hundreds of discarded wording suggestions may contribute little to understanding a validated statistical analysis.

By contrast, one prompt defining how primary outcome categories were generated may be essential.

Researchers should prioritize interactions that explain consequential methods, outputs, revisions, and scientific decisions.

The archive should be complete enough for its intended accountability purpose without collecting unnecessary information.

How Should Researchers Protect Archived Prompts and Outputs?

Prompts and generated outputs may contain participant information, unpublished findings, or other restricted material.

Researchers should apply appropriate access controls, secure storage, retention procedures, and institutional governance.

Where possible, descriptive metadata can be separated from protected source information.

For example, an archive manifest may identify a confidential interview dataset through an internal reference without reproducing participant statements in a public document.

However, the underlying authorized archive must still preserve sufficient evidence for legitimate inspection.

Transparency should not create new confidentiality violations.

What if the AI Provider Does Not Allow Complete Export?

Some AI services may not expose full configuration information or provide convenient access to complete interaction histories.

Researchers should establish these limitations before relying on the system for consequential analytical procedures.

Where permitted, they may preserve relevant prompts, outputs, and model information through an authorized research documentation process.

If essential information cannot be retained, the system may be unsuitable for a task requiring detailed reconstruction.

The limitation should be acknowledged rather than concealed through an incomplete account of the workflow.

What Should a Future Researcher Be Able to Determine?

A useful archive should allow an appropriately qualified person to identify the scientific role of AI and understand how consequential outputs were produced.

The person should be able to distinguish original AI-generated material from subsequent human corrections and determine what verification supported the final results.

They should also be able to identify important limitations, including unavailable model versions or restricted source data.

Not every study can provide public access to every artifact, and not every AI process can be repeated exactly.

The defensible standard is sufficient, accurate, and appropriately protected evidence of how AI contributed to the research.

Preparing an AI Research Archive for Future Methodological Inspection

Two years later, the original AI model may no longer be available. Nevertheless, the archive can show which classifications the system produced, which were corrected, and how the final variables entered the analysis.

The research team cannot necessarily guarantee exact regeneration of the original model outputs. It can, however, preserve the evidence needed to examine the procedure and independently reproduce downstream calculations where the necessary data and code remain available.

Misconceptions About Archiving AI-Assisted Research

Preserve Enough Evidence to Explain What AI Did and What Researchers Decided

Before completing a research project, identify which AI contributions materially influenced the evidence, analysis, or conclusions and determine what records are necessary to reconstruct those contributions.

Researchers should distinguish the interactions that require logging during the project from the records that must remain available in the long-term archive.

The archive should support reconstruction of the AI-assisted research process and preserve the evidence necessary for reproducing consequential analytical results where feasible.

Researchers should also connect archived records to the findings of the prepublication workflow audit and assess whether model changes affect the interpretation of preserved outputs.

A defensible archive should allow future researchers to distinguish the AI system's contribution from the scientific judgments and corrections made by accountable researchers.

Before Finalizing an AI-Assisted Research Archive

Frequently Asked Questions About Archiving AI Research Contributions

Archive the Evidence Needed to Understand AI's Actual Contribution

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.