01 · The Question
Should You Decide How AI Will Be Used Before the Research Starts?
Many research teams discuss authorship, data collection, responsibilities, ethics, and publication before beginning a project. AI use is often handled differently: someone starts using a tool, another collaborator follows, and the team works out the rules later.
That approach may be harmless when AI is limited to incidental, low-consequence tasks. It becomes much less comfortable when AI touches participant data, analytical code, qualitative coding, manuscript drafting, confidential material, or methodological decisions.
An AI-use plan allows the team to make those decisions before different practices become embedded in the project.
03 · What You Need to Know
An AI-Use Plan Is Most Valuable Before Practices Diverge
Collaborative Research Already Benefits From Early Agreements
Planning important responsibilities at the beginning of collaborative research is established responsible-conduct practice.
The U.S. Office of Research Integrity advises collaborators to develop a common understanding from the beginning about project goals, roles, data collection and sharing, changes to research design, publication responsibilities, reporting, and intellectual property. ORI's guidance on written agreements similarly notes that complex collaborations benefit from documenting rights and responsibilities rather than relying entirely on informal understandings.
An AI-use plan extends this logic to a newer set of research decisions.
It answers a practical question: if several people will use AI during the project, what should everyone know before the first consequential prompt is sent?
The Plan Should Start With External Requirements
Do not begin by asking the team what AI practices it prefers. Begin by identifying what it is already required to do.
Applicable requirements may come from university policy, research ethics approval, information-security rules, funder conditions, contracts, data-use agreements, applicable law, and anticipated publication policies.
The plan then translates those requirements into project procedures.
External requirement
The university prohibits confidential participant information from being entered into public generative AI services.
Project interpretation
Raw interview transcripts and identifiable survey responses are classified as restricted for public AI services.
Operational rule
Team members may use public AI tools only with public or approved non-confidential project material.
Escalation rule
Any proposed AI processing of participant-level data requires confirmation that the system and procedure are institutionally and ethically permitted.
This is more useful than copying several pages of institutional policy into the project folder and hoping everyone interprets them identically.
Decide Which AI Uses Are Permitted, Restricted, and Approval-Required
A practical plan can classify activities according to the project rather than attempting to produce a universal taxonomy of AI.
| Category |
Possible Project Examples |
| Permitted |
Brainstorming public search terms, explaining generic code errors, improving non-confidential prose, or working with synthetic examples |
| Permitted with verification |
Generating analytical code, suggesting manuscript restructuring, summarising researcher-provided public material, or proposing preliminary classifications |
| Approval required |
Introducing a new AI system into data analysis, processing sensitive research information, or materially changing an approved research method |
| Prohibited |
Entering information into an AI service where applicable institutional, ethics, contractual, or legal requirements forbid the transfer |
The examples should be adapted to the project. A task permitted in one study may be inappropriate in another because the data, participants, contractual obligations, or methods differ.
Specify the Tool, Not Merely “AI”
AI services are not interchangeable. A university-managed enterprise system may have different contractual and data-handling arrangements from a public consumer service. Different versions of the same branded product may also have different features or settings.
The plan should therefore identify approved systems or establish a process for checking them. "ChatGPT allowed" is less useful than specifying which account or service configuration is approved for which type of information, assuming the institution itself has made that distinction.
This becomes especially important when different units within the same institution impose different AI requirements.
Decide What Information AI Systems May Receive
The plan should address inputs before outputs.
Researchers may encounter participant data, confidential collaborator material, unpublished manuscripts, grant applications, proprietary datasets, passwords, API keys, licensed content, reviewer material, and commercially sensitive information during a project. The fact that a team member has access to information does not automatically mean they may provide it to an AI service.
A useful plan identifies categories of information that require additional approval or must not be entered into specified systems.
Watch Out
Do not invent a project-level data classification that contradicts your institution's official categories. Where the institution already classifies information and approved systems, use that framework in the AI-use plan.
Define Verification Before AI Outputs Enter the Research Record
AI output can be fluent while being wrong, incomplete, fabricated, biased, methodologically inappropriate, or subtly inconsistent with the research context.
The plan should therefore define who checks consequential outputs and what checking entails.
Generated references should be checked against the original publications. Generated code should be inspected and tested. Statistical explanations should be evaluated by someone competent in the method. AI-assisted qualitative interpretations should be compared with the underlying data and analytical approach. Manuscript claims should be verified against evidence rather than accepted because the wording sounds plausible.
The European Commission's updated 2026 Living Guidelines continue to frame responsible generative AI use around research-integrity principles including accountability, transparency, and responsibility. An AI-use plan can translate those principles into named human responsibilities.
Agree on Documentation Before You Need the Records
Teams frequently discover the need for documentation at the worst possible time: when a journal asks for disclosure, a collaborator questions an analysis, or an ethics committee asks what changed.
The beginning of the project is a better time to decide what will be recorded.
The plan might require documentation when AI materially contributes to research design, code, data processing, analysis, interpretation, figures, or substantive manuscript content. Records might include the system, purpose, relevant inputs or input categories, outputs retained in the workflow, verification performed, and researcher responsible.
The level of detail should be proportionate. A project should not become a prompt archive merely because someone used autocomplete to finish a sentence.
Decide How AI Use Will Be Disclosed
The team may not know the eventual journal when the project begins, so it cannot predict every publication requirement. It can nevertheless establish a principle: consequential AI use will be documented well enough to permit accurate methodological reporting and disclosure later.
That avoids the awkward end-of-project question, "Does anyone remember exactly what we used this tool for eighteen months ago?"
Publication requirements should then be checked again when a journal is selected, including what to do if the journal has no clear AI policy.
Assign Decision Authority
An AI-use plan should identify who can approve changes rather than merely listing rules.
Can any team member adopt a new AI tool? Does the principal investigator decide? Must the data manager review services that process project data? When must the research office, information-security team, or ethics committee become involved?
The answer depends on project structure and external governance. What matters is that the decision path is explicit.
The separate question of who should decide which AI uses are permitted becomes particularly important when methodological expertise and formal project authority are held by different people.
Plan for Mistakes and Unapproved Use
A useful governance plan should anticipate that someone may accidentally or deliberately depart from it.
The response should depend on the consequence. Using an unapproved AI tool to rewrite a public sentence is different from uploading identifiable participant data to an unauthorised service.
The plan can specify that researchers should stop the affected activity, preserve relevant records, notify the appropriate project lead, assess what information was involved, and determine whether institutional, ethics, privacy, funder, or other reporting is required.
This gives the team a starting process if a collaborator uses AI in a way the team did not approve.
The Plan Must Be Able to Change
An AI-use plan written at project inception should not become a historical artefact that everyone respectfully ignores after six months.
The European Commission describes its own responsible generative AI guidance as "living" guidance, reflecting the rapid development of AI technologies. Research projects need a similar capacity for revision, although at a much smaller scale.
The plan should specify review points or triggers: a new AI tool, a new dataset, a new collaborator, a change in ethics approval, a new funder condition, selection of a target journal, or a relevant institutional policy update.
Long projects should also establish a process for monitoring changes in external AI policies.
