03 · What You Need to Know
Which AI Prompts Are Worth Preserving, and Why?
Why Might AI Prompts Count as Research Records?
Research records document how a study was planned, conducted, analyzed, and reported.
Prompts can become part of that documentation when they define instructions used to produce research outputs.
For example, a prompt specifying how an AI system should classify interview responses may function similarly to a coding protocol.
A prompt directing an AI system to generate experimental stimuli may help explain how those materials were created.
Likewise, an instruction specifying variables and analytical procedures may help reconstruct how AI-generated statistical code was developed.
In these situations, the prompt is not merely conversational text. It forms part of the documented process through which research activities were performed.
Whether it must be retained depends on applicable policies and its significance to the research.
Are Researchers Required to Save Every AI Prompt?
No universal rule requires the preservation of every prompt used during research.
Universities, journals, funders, and research agreements may establish specific documentation requirements.
However, general publisher expectations concerning AI disclosure should not automatically be interpreted as requiring a complete archive of every interaction.
A researcher who uses AI for minor punctuation corrections may not need to preserve every request unless an applicable policy requires it.
By contrast, prompts that determine how research data are classified or experimental materials are generated may be important for methodological transparency.
The appropriate approach is proportionate documentation based on the consequences of the AI-assisted activity.
What Do Major Publishing Authorities Say About AI Documentation?
The International Committee of Medical Journal Editors (ICMJE) recommends disclosure of AI-assisted technologies used in submitted research and emphasizes human responsibility for the resulting work.
The Committee on Publication Ethics (COPE) similarly emphasizes transparency concerning AI tools used in manuscript preparation and research activities.
These positions support accurate reporting of consequential AI assistance, but they should not be interpreted as establishing a universal requirement to archive every prompt.
Where AI contributes to research methodology, ordinary expectations concerning documentation and reproducibility may make prompt preservation relevant.
Researchers should distinguish explicit publisher requirements from recommended research-recordkeeping practices.
Which Types of Prompts Should Researchers Prioritize?
The importance of preserving a prompt depends on whether it influenced consequential research activities.
| AI Prompt Activity |
Why Preservation May Matter |
Suggested Documentation Priority |
| Routine grammar correction |
Usually concerns manuscript presentation. |
Lower, unless required by policy or substantive changes occurred. |
| Research question development |
May influence conceptual decisions. |
Context-dependent. |
| Qualitative coding instructions |
May determine classification categories and analytical outputs. |
High when used in the final analysis. |
| Literature screening instructions |
May influence which studies are selected. |
High for formal evidence syntheses. |
| Statistical code generation |
May influence computational implementation. |
High when code contributes to reported findings. |
| Experimental stimulus generation |
May determine materials presented to participants. |
High when relevant to the study design. |
| Rejected exploratory suggestions |
May explain consequential research decisions. |
Selective, according to significance. |
These priorities are practical recommendations rather than universal retention requirements.
Researchers should apply the standards governing their particular project.
Why Are Prompts Important for AI-Assisted Qualitative Analysis?
Prompts can influence how AI systems classify, summarize, or interpret qualitative research materials.
For example, asking an AI system to identify students' difficulties with online learning may produce different categories from asking it to identify evidence of poor student motivation.
The second prompt introduces a specific interpretive assumption.
If the resulting classifications contribute to the findings, preserving the instructions can help researchers examine how the analytical process was shaped.
Prompts may also define category descriptions, examples, and rules for ambiguous responses.
These details can be important when evaluating whether the coding procedure was applied consistently.
However, preserving prompts does not independently establish that the resulting qualitative interpretation is credible.
Should Prompts Used for Statistical Code Generation Be Saved?
Consequential prompts used to generate analytical code may be worth preserving because they can document how a computational procedure was initially specified.
For example, a researcher may ask ChatGPT to generate R code for a repeated-measures analysis with particular variables and assumptions.
The prompt may reveal the intended analytical design.
However, the final executed code may differ substantially from the generated version.
Researchers should therefore preserve the tested script, relevant software environment, and verification records alongside consequential prompts.
The prompt alone cannot establish which computations were actually performed.
The requirements concerning AI-generated research code are closely connected to this distinction.
Should Prompts Used for Literature Searching Be Preserved?
Prompts may be relevant when AI contributes to a formal literature-search or study-selection procedure.
For example, a systematic review may use AI to screen abstracts according to predefined eligibility criteria.
The instructions supplied to the model may influence which records are classified as relevant.
Preserving those instructions can support evaluation of the screening procedure.
However, an informal request for possible search terms may not warrant the same level of documentation.
The distinction between exploratory searching and formal evidence retrieval should guide the recordkeeping decision.
For systematic reviews, relevant reporting standards such as PRISMA 2020 may also apply.
Should Prompts Used to Generate Research Images Be Saved?
Prompts may be important when AI-generated images form part of a research method.
For example, a researcher may generate synthetic classroom scenes for an experiment examining students' perceptions of AI-created educational materials.
The prompt may influence the appearance, content, and characteristics of the stimuli.
Preserving the instructions and selected outputs can help explain how the materials were developed.
However, the prompt alone may not reproduce the same image because outputs can depend on model versions, settings, and random variation.
Researchers should preserve the actual stimuli used in the experiment, subject to applicable rights and retention requirements.
They should also verify that the use of AI-generated imagery is permitted by the relevant ethics approval and publisher policy.
Should Researchers Save Prompts Used Only for Brainstorming?
Not necessarily. Exploratory brainstorming may generate numerous suggestions that are never adopted.
Preserving every such interaction may provide little additional methodological value.
However, a prompt may be worth recording if it materially influenced a research question, hypothesis, conceptual framework, or methodological decision.
For example, a researcher may adopt an AI-suggested analytical approach after evaluating alternatives.
Recording the consequential interaction may help explain the development of the study.
The distinction between AI-assisted brainstorming disclosure and internal documentation should remain clear.
Why Should Researchers Preserve Prompt Revisions?
Prompts may change during a research project as researchers refine instructions or correct problems.
For example, an initial coding prompt may produce inconsistent classifications, leading the researcher to revise category definitions.
If the revised instructions produce the final analytical outputs, the change may be methodologically significant.
Researchers should consider preserving consequential versions and identifying which version was used for each relevant stage.
This can help explain differences among outputs and prevent later confusion.
However, not every minor wording experiment needs a separate formal record.
The appropriate level of version control depends on whether the changes affected the research process or findings.
What Is the Difference Between a Prompt Template and an Actual Prompt?
A prompt template is a reusable structure containing instructions and placeholders for variable inputs.
An actual prompt is the specific instruction submitted during an interaction, including any data, examples, or contextual information supplied.
For example, a qualitative coding template may instruct the model to assign categories according to a predefined framework, while individual prompts insert different participant responses.
Preserving the template can document the general procedure.
However, where specific inputs or examples materially affect the outputs, additional records may be necessary.
Researchers should not present a sanitized template as an exact historical prompt if it differs from the original instructions.
Do Multi-Turn AI Conversations Need to Be Saved?
Sometimes. A consequential AI interaction may involve several messages rather than one isolated prompt.
The model's later responses may depend on earlier instructions, examples, corrections, or uploaded materials.
For example, a researcher may define coding categories in one message, supply examples in another, and request classification in a third.
Saving only the final message may omit information necessary to understand the resulting output.
Where conversation context materially affects the research, preserving the relevant sequence may be appropriate.
However, complete conversation histories may contain irrelevant exchanges or confidential materials that should not automatically be shared publicly.
Does Saving a Prompt Guarantee Reproducibility?
No. The same prompt may produce different outputs depending on the model, settings, input materials, conversation context, and provider-side changes.
Some generative AI services do not guarantee deterministic responses.
Even when the model identifier remains unchanged, technical updates may affect behavior.
Researchers should therefore distinguish preserving the instruction from reproducing the resulting output.
Prompt preservation
Retaining the instructions supplied to an AI system so the interaction can be understood or examined later.
Output reproducibility
Being able to reproduce or meaningfully evaluate the outputs using sufficiently documented procedures, systems, and materials.
Prompt preservation supports transparency but cannot guarantee identical AI behavior.
What Other Information Should Be Saved With Prompts?
Consequential prompts are more informative when connected to the AI system and outputs involved.
| Associated Record |
Purpose |
| AI tool and provider |
Identifies the application or service used. |
| Model identifier |
Documents the underlying system where known. |
| Date of use |
Establishes when the interaction occurred. |
| Relevant settings |
Records available parameters affecting generation. |
| Input materials |
Identifies the data or files processed. |
| Generated output |
Preserves what the AI system actually produced. |
| Human decision |
Records whether the output was accepted, revised, or rejected. |
| Verification evidence |
Explains how consequential outputs were checked. |
Not every item is necessary for every interaction. The relevant information depends on the research method and applicable recordkeeping requirements.
The broader structure can be maintained through a research AI-use log.
Should Researchers Save AI-Generated Responses Alongside Prompts?
Yes, where the outputs materially contribute to the research and their preservation is appropriate under applicable policies.
A prompt shows what the researcher requested, but the response shows what the AI system actually produced.
For example, a researcher may ask AI to classify survey responses according to predefined categories. The prompt documents the classification instructions, while the output records the categories assigned.
Both may be needed to understand the analytical process.
Preserving only the prompt may be insufficient because the same instruction may generate different responses later.
The question of archiving AI-generated research responses is therefore closely connected to prompt preservation.
Can Researchers Rely on ChatGPT Conversation History?
Conversation history may provide a convenient record, but it should not automatically be treated as a permanent research archive.
Access to historical interactions may depend on account settings, provider policies, product changes, or retention arrangements.
Some conversations may be deleted or become unavailable.
Researchers should therefore consider preserving consequential interactions in an approved research storage system rather than relying exclusively on the AI provider's interface.
The preserved record should be linked to the relevant research activity and protected according to applicable confidentiality requirements.
What Is the Best Format for Saving AI Prompts?
No single format is universally required.
Researchers may preserve prompts in text files, structured spreadsheets, electronic laboratory notebooks, version-controlled repositories, or approved research management systems.
The appropriate format depends on the research workflow, collaboration requirements, and sensitivity of the information.
For example, prompt templates used in computational experiments may be well suited to version control.
Qualitative research prompts containing participant excerpts may require restricted storage with appropriate access controls.
The important qualities are accuracy, retrievability, contextual documentation, and appropriate protection.
Should Researchers Use Version Control for Prompts?
Version control may be useful when prompt changes affect a research procedure.
For example, a project using AI-assisted document classification may develop several prompt versions during pilot testing.
Recording the changes can help identify which instructions produced the final classifications.
Version control can also distinguish exploratory prompts from the finalized procedure.
However, researchers should avoid placing confidential participant data into publicly accessible repositories.
Where prompts contain restricted information, private or institutionally approved storage may be necessary.
The objective is to preserve a meaningful history of consequential methodological changes.
Can Saved Prompts Contain Confidential Research Data?
Yes. Prompts may contain participant quotations, unpublished manuscripts, institutional records, or other sensitive materials.
For example, a researcher may paste an interview excerpt into an AI system while requesting preliminary coding.
The saved prompt then becomes a record containing part of the research dataset.
Researchers should apply appropriate confidentiality, access, and retention safeguards.
Removing direct identifiers may not eliminate all re-identification risks, particularly for detailed qualitative narratives.
The fact that a prompt is useful for transparency does not mean it can be shared without restriction.
Should Researchers Save Prompts Containing Personal Information?
Only under arrangements consistent with applicable ethics, consent, institutional, and data-protection requirements.
Researchers should avoid unnecessary duplication of personal information in AI-use logs.
Where possible, records may reference securely stored inputs rather than reproduce sensitive material across multiple files.
However, any redaction or modification should be documented when it changes information necessary to understand the original procedure.
Researchers should also determine whether the original transfer of information to the AI service was authorized.
Preserving the prompt does not retrospectively resolve an unauthorized data-processing activity.
Should Researchers Save Rejected AI Prompts?
Not every rejected prompt needs preservation.
However, rejected prompts may be relevant when they explain consequential methodological changes or research decisions.
For example, a researcher may abandon an initial classification prompt after discovering systematic errors affecting a particular category.
Preserving that version and the reason for its rejection may help explain the development of the final procedure.
By contrast, an unsuccessful request for a more concise sentence may have little methodological significance.
The value of preserving rejected prompts depends on whether they contribute to a defensible account of the research process.
Should Researchers Disclose Every Saved Prompt?
No. Internal preservation and publication disclosure are different activities.
A researcher may save prompts as part of the research archive without publishing them in the manuscript.
Some prompts may contain confidential information or technical details unnecessary for readers.
Others may be essential to understanding the method and therefore appropriate for supplementary materials or controlled-access sharing.
The question of whether AI prompts should be disclosed depends on the journal's requirements and the prompts' methodological significance.
Researchers should not assume that retaining a prompt automatically makes it suitable for public release.
How Long Should Saved AI Prompts Be Retained?
Retention should follow the requirements governing the research records.
Where prompts materially contribute to published findings, they may need to be preserved with the broader research archive.
However, no universal retention period applies to every AI prompt across institutions and disciplines.
Researchers should consider funder requirements, institutional policies, research agreements, and confidentiality obligations.
They should also distinguish records that must be retained from incidental interactions that do not contribute meaningfully to the research.
Preservation should be planned rather than treated as indefinite storage by default.
What If Researchers Forgot to Save Their Prompts?
The appropriate response depends on whether the prompts can be recovered and how important they are to the research.
Researchers may examine available conversation histories, research notes, code repositories, or other records.
However, reconstructed prompts should not be presented as exact historical instructions unless supported by evidence.
For example, a researcher may remember asking AI to classify responses using five categories but not recall the precise wording.
The researcher should distinguish that recollection from a preserved prompt.
Where missing instructions materially limit the reproducibility or evaluation of the procedure, the limitation may need to be acknowledged.
Can Prompt Preservation Help During Research Integrity Investigations?
Potentially. Preserved prompts may help establish what instructions were supplied to AI and how generated outputs entered the research workflow.
For example, a prompt history may show whether an AI-generated interpretation was requested as exploratory assistance or incorporated into the final analytical procedure.
However, prompts alone do not establish whether the researcher verified the output or acted responsibly.
Investigators may need to examine additional evidence, including outputs, drafts, analytical records, and researcher decisions.
Authenticity and chronology also matter. A prompt reconstructed after an allegation should not be represented as a contemporaneous record.
Prompt preservation can support an audit trail, but it does not replace the broader research record.
