Should Researchers Disclose the Prompts They Used With AI?

If AI Contributed to Your Research, Must You Reveal What You Asked It?

A researcher uses ChatGPT to assist with qualitative analysis. The researcher uploads de-identified interview excerpts and instructs the system to classify responses according to a predefined coding framework.

The manuscript explains that AI assisted with preliminary coding, but it does not include the prompts used to generate the classifications.

Would another researcher be able to evaluate the procedure without knowing the instructions supplied to the model?

Now consider a different situation. A researcher uses ChatGPT to correct grammar in an independently written manuscript. The interaction involves dozens of minor editing requests that have no direct role in the research methodology.

Should every prompt from that conversation also be published?

The distinction matters because prompts can function as methodological instructions in one study and ordinary writing assistance in another. Treating every AI conversation as equally important may create unnecessary documentation, while withholding consequential prompts can make an AI-assisted procedure difficult to evaluate.

Disclose Prompts When Required or Necessary to Understand the Research Method

The important question is not how many prompts were used, but whether their wording materially influenced the research. A single consequential prompt may be more important to methodological transparency than hundreds of minor editing instructions.

When AI Prompts Become Part of the Research Methodology

What Is an AI Prompt in Research?

An AI prompt is an instruction, question, or input supplied to an AI system to influence the output it produces.

Prompts may contain ordinary questions, detailed task descriptions, examples, classification rules, documents, datasets, or structured instructions.

For example, a researcher may ask a language model to explain a statistical concept, generate code, summarize a publication, or classify interview excerpts.

The methodological significance of the prompt depends on its role in the research.

A request to improve punctuation differs from a detailed instruction defining how an AI classifier should assign categories to participant responses.

Researchers should therefore consider prompts as potential research records rather than assuming that every interaction is merely informal conversation.

Why Can Prompt Wording Matter?

Generative AI outputs can be sensitive to the wording, structure, context, and examples included in prompts.

For example, asking an AI system to identify challenges experienced by students may produce a different analytical emphasis from asking it to identify evidence of low student motivation.

The second instruction introduces a more specific interpretive expectation.

Similarly, a prompt directing an AI system to classify responses using predefined categories may produce different results from one asking the model to generate categories inductively.

These differences matter because prompts can influence the analytical assumptions built into an AI-assisted procedure.

Where prompt wording materially shapes the findings, documenting the instructions may be necessary for methodological transparency.

What Do Major Publishing Authorities Require?

The International Committee of Medical Journal Editors (ICMJE) recommends disclosure of AI-assisted technologies and an explanation of how they were used in submitted research.

The Committee on Publication Ethics (COPE) similarly emphasizes transparency about AI tools used in manuscript preparation and research activities.

However, these general positions should not be interpreted as establishing a universal requirement to publish every prompt from every AI interaction.

Springer Nature, Elsevier, and individual journals may establish additional requirements concerning AI-use declarations and methodological documentation.

Where AI is integral to the research procedure, ordinary expectations of methodological transparency may also make prompt information relevant even without a dedicated prompt-disclosure rule.

The journal's current instructions and the research design should therefore guide the decision.

What Is the Difference Between Prompt Disclosure and Prompt Documentation?

Prompt disclosure concerns what researchers communicate to editors, reviewers, or readers.

Prompt documentation concerns what researchers preserve as part of the research record.

These activities overlap but are not identical.

A researcher may retain consequential prompts without publishing the complete conversation history.

Conversely, a manuscript may need to report a prompt template or classification instruction even when the full interaction archive cannot be shared because of confidentiality restrictions.

Which Types of AI Prompts Are Most Likely to Matter?

These are methodological considerations rather than universal disclosure rules.

The appropriate level of reporting depends on the research question, the AI system's role, and the applicable policy.

Do Researchers Need to Disclose Prompts Used for Grammar Correction?

Not universally.

Some publishers exempt defined routine AI-assisted copyediting from declaration, while others require broader reporting of generative AI writing assistance.

Where language assistance is reportable, the journal may require only the tool and purpose rather than every editing prompt.

For example, a researcher who used ChatGPT to correct grammar in an independently written manuscript may not need to publish dozens of individual requests.

However, substantive rewriting should not be concealed under a description of routine grammar correction.

The requirements concerning AI-assisted grammar correction disclosure should be evaluated before deciding whether the prompts themselves need to be shared.

Should Prompts Used for Research Brainstorming Be Disclosed?

Not necessarily under every policy.

Researchers may use AI to explore preliminary topics, research questions, hypotheses, or methodological possibilities.

Some interactions may have little influence on the final study, while others may materially shape its conceptual direction.

Where a journal requires disclosure of consequential conceptual assistance, the researcher should explain the AI contribution accurately.

However, that does not automatically mean the complete brainstorming transcript must be published.

The distinction between disclosing AI-assisted brainstorming and publishing every brainstorming prompt should remain clear.

Why Are Prompts Important in AI-Assisted Qualitative Coding?

Prompts may determine how an AI system interprets and classifies qualitative data.

For example, a researcher may instruct a language model to assign one of five predefined codes to each open-ended response.

The prompt may define the categories, provide examples, establish rules for ambiguous cases, and specify the required output format.

These instructions can influence the resulting classifications.

Without knowing the coding instructions, readers may be unable to evaluate how the AI system applied the framework.

Where AI-assisted coding materially affects the findings, prompt documentation may therefore be an important part of the methodology.

However, preserving prompts does not establish that the coding framework is valid or that the classifications are accurate.

Can Prompt Wording Introduce Bias Into Qualitative Analysis?

Potentially. Prompts may direct the AI system toward particular interpretations or categories.

For example, asking a model to identify evidence of academic dishonesty in student interviews may encourage attention to suspicious behavior while overlooking legitimate educational uses of AI.

A more neutral instruction may produce a different distribution of suggested categories.

Researchers should examine whether their prompts introduce assumptions inconsistent with the chosen analytical approach.

In reflexive qualitative research, prompts may also become part of the researcher's analytical positioning and decision-making.

Documenting consequential instructions can help make these influences visible, although documentation alone does not eliminate bias.

Should Prompts Used for Literature Screening Be Disclosed?

Prompts may be important when AI contributes to study selection in a systematic review or other evidence synthesis.

For example, an AI system may be instructed to classify abstracts according to prespecified inclusion and exclusion criteria.

The wording of those criteria and the instructions for handling uncertain cases can influence screening decisions.

PRISMA 2020 requires transparent reporting of study-selection methods, including relevant automation tools.

Although PRISMA does not impose a universal requirement to publish every generative AI prompt, the actual screening instructions may be necessary to explain how the automation was used.

Researchers should document the procedure and any human verification or adjudication of AI-assisted decisions.

Should Prompts Used to Generate Statistical Code Be Shared?

Prompts may help explain how analytical code was developed, particularly when they contain important methodological instructions.

For example, a prompt might specify variable transformations, model structure, missing-data treatment, or statistical assumptions.

However, the prompt alone does not establish what the final executed code actually did.

The researcher may have modified the generated script substantially before producing the published results.

For computational reproducibility, the final tested code, data-processing procedures, software environment, and execution instructions are often more directly relevant than the original prompt alone.

Consequential prompts may still be useful supporting records, particularly when they document analytical decisions or explain the development of the software.

The relationship between AI-generated research code and disclosure therefore involves more than publishing the instructions used to generate it.

Are Prompts Research Data When AI Is the Subject of the Study?

They may be. When researchers systematically investigate AI behavior, prompts can form part of the experimental inputs or research materials.

For example, a study may compare how different language models respond to the same educational assessment questions.

The prompts used to elicit responses are central to interpreting the results.

Changes in wording, examples, or instructions may affect the observed outputs.

In such studies, preserving and reporting the prompts may be essential for methodological evaluation.

The researcher should also document relevant model information, settings, dates, and response collection procedures.

These requirements arise from the research design as well as any publisher-specific AI disclosure policy.

Can Researchers Share Only a Representative Prompt?

Sometimes, but the adequacy of that approach depends on the study.

A representative prompt may illustrate a general workflow, but it may not reveal important differences among instructions used across the dataset.

For example, if a researcher repeatedly revised the prompt after observing classification errors, reporting only the final version could conceal changes that influenced earlier outputs.

Researchers should distinguish a stable prompt template from a procedure involving substantial iterative modification.

Where prompt variation materially affects the findings, the relevant versions and their use should be documented.

The reporting should not imply that one prompt was used consistently when the actual procedure involved multiple consequential revisions.

What About Multi-Turn AI Conversations?

Some AI-assisted procedures involve several exchanges rather than a single prompt.

A researcher may provide initial instructions, review the output, request revisions, and supply additional examples.

Later responses may depend on the preceding conversation context.

Publishing only the final user message may therefore omit information necessary to understand how the output was produced.

Where the interaction is methodologically consequential, researchers should consider preserving the relevant conversational sequence or documenting the procedure sufficiently to explain the dependencies.

However, complete conversation histories may contain confidential information or irrelevant exchanges that should not automatically be made public.

Do System Prompts and Hidden Instructions Matter?

Potentially. AI outputs may be influenced by instructions or configurations that are not visible to the researcher.

Commercial applications may incorporate provider-defined system instructions, safety mechanisms, retrieval components, or other processing steps.

Researchers may not have access to these details.

Consequently, publishing the visible user prompt does not necessarily reveal every factor influencing the generated response.

Where this limitation matters to the methodology, researchers should acknowledge it rather than imply that the visible prompt provides a complete specification of the system.

Prompt disclosure can improve transparency without guaranteeing full technical reproducibility.

Does Publishing the Exact Prompt Guarantee the Same AI Output?

No. Identical prompts do not necessarily produce identical responses.

Outputs may depend on the model, settings, conversation history, input materials, random variation, and provider-side changes.

Some services may update the underlying system while retaining the same application name.

Researchers should therefore distinguish prompt reproducibility from output reproducibility.

A published prompt allows others to understand or attempt to repeat the instruction, but it does not guarantee that the resulting response will match the original.

Where outputs are central to the findings, preserving the actual responses may be as important as documenting the prompts.

Should Researchers Report the Model Alongside the Prompt?

Model identification may be important when prompts form part of a research procedure.

The same prompt may produce different outputs across models or model versions.

Researchers should therefore record the AI tool, underlying model where known, access date, and relevant settings when these details are methodologically significant.

However, the exact information required depends on the journal and research design.

Researchers should not invent model identifiers or settings that were unavailable at the time of use.

The question of identifying AI tools and model versions should be considered alongside prompt documentation.

Can AI Prompts Contain Confidential Research Information?

Yes. Prompts may include unpublished manuscripts, participant quotations, institutional documents, proprietary data, or other restricted materials.

Publishing the complete prompt could therefore disclose information that researchers are obligated to protect.

For example, a qualitative coding prompt may contain a detailed interview excerpt that indirectly identifies a participant even after names have been removed.

Researchers should assess confidentiality and privacy obligations before sharing prompt histories.

Where necessary, they may provide redacted prompts, sanitized templates, or controlled-access documentation.

However, redaction should be described accurately when it removes information relevant to understanding the original procedure.

Does Removing Names Make Prompts Safe to Publish?

Not necessarily. Detailed narratives can contain combinations of contextual information that permit indirect identification.

For example, a participant's occupation, institution, location, and unusual experience may make them recognizable even without a name.

Researchers should therefore consider re-identification risks rather than relying exclusively on removal of direct identifiers.

The appropriate approach depends on the research ethics approval, consent arrangements, institutional requirements, and applicable data-protection law.

Transparency does not override obligations to protect participants.

Can Researchers Share Prompt Templates Instead of Actual Data-Containing Prompts?

Potentially. A prompt template may communicate the analytical instructions without reproducing confidential research data.

For example, a researcher might publish the classification rules and replace the original participant response with a clearly marked placeholder.

This can help readers understand the procedure while reducing unnecessary disclosure of sensitive material.

However, the template should accurately represent the instructions used.

If the original prompts contained examples, contextual information, or additional directions that materially influenced the outputs, those differences may need to be documented.

A sanitized template should not be presented as an exact historical prompt when it has been modified.

Where Should AI Prompts Be Reported?

The appropriate location depends on their role and the journal's instructions.

A methods section may describe the prompt structure and explain how instructions were developed or applied.

Supplementary materials may contain complete prompt templates or additional examples where permitted.

A research repository may provide access to consequential prompts and outputs, subject to ethical, legal, and institutional restrictions.

A dedicated AI-use declaration may identify the tool and summarize its role without reproducing the full prompt history.

The distinction between AI disclosure locations in a manuscript and supporting methodological documentation should remain clear.

What Should a Prompt Disclosure Statement Say?

A useful statement should explain the role of prompts in the procedure and identify where relevant instructions can be examined.

This example assumes that the described procedures occurred and that the prompt template is actually available.

Researchers should identify the specific model and provide additional information where required.

The statement should not claim that prompts were stable or classifications were comprehensively reviewed unless those conditions were satisfied.

Should Researchers Save Every Prompt They Use?

Not universally. The appropriate documentation depends on the significance of the interaction and applicable retention requirements.

For example, preserving every punctuation-correction request may provide limited methodological value.

By contrast, prompts used to generate experimental stimuli, classify research data, or determine analytical outputs may be important records.

Researchers should prioritize interactions that materially influence the research procedures or findings.

The separate question of saving AI prompts used in research concerns recordkeeping even when publication disclosure is not required.

Should Researchers Preserve Rejected or Revised Prompts?

That depends on whether the changes influenced consequential research decisions.

For example, a researcher may revise a classification prompt after discovering that the model systematically misinterprets certain responses.

The revision may affect the final analytical procedure and should be documented where necessary to explain how the classifications were produced.

However, not every unsuccessful exploratory prompt needs to appear in the published manuscript.

Researchers should distinguish minor experimentation from methodological changes that affect the evidence or findings.

The value of documenting rejected AI suggestions depends on their relevance to reconstructing the research process.

What If the Journal Does Not Mention Prompt Disclosure?

Researchers should review the journal's broader AI-use policy and methodological reporting requirements.

Where prompts materially influence the research, documenting them may be appropriate even without an explicit prompt-disclosure rule.

If uncertainty remains, the researcher may ask the editorial office whether prompt templates or interaction records should be provided.

The inquiry should describe the actual procedure rather than ask whether all ChatGPT prompts must be published.

Researchers should also consider confidentiality restrictions before offering to share complete interaction histories.

Policy silence does not eliminate the need for sufficient methodological reporting, but it should not be interpreted as a universal requirement to disclose every conversation.

When Two Different Prompts Produce Different Qualitative Classifications

The example illustrates why prompt wording can be part of the analytical method rather than a trivial technical detail.

If the manuscript reported only that AI was used for coding, readers might not understand how the instructions shaped the resulting classifications.

At the same time, publishing the prompts would not independently establish that the final coding framework was valid.

Methodological transparency and substantive verification remain separate responsibilities.

Misconceptions About AI Prompt Disclosure

How Should You Decide Which AI Prompts to Disclose?

Researchers should evaluate whether the prompts were incidental to manuscript preparation or consequential to the research procedures and findings.

They should also distinguish what must be preserved internally from what can or must be shared publicly.

Researchers should preserve the connection between prompts, model information, outputs, and subsequent human decisions when that relationship matters to the findings.

The appropriate standard is sufficient methodological transparency without unnecessary or unauthorized disclosure of research information.

What to Verify Before Disclosing AI Prompts

Frequently Asked Questions About Disclosing AI Prompts

Disclose the Prompts That Matter to the Research, Not Every Conversation

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.