03 · What You Need to Know
Building Shared AI Expectations Across a Research Team
Why Individual AI Preferences Are Not Enough
Research collaborators may differ in their familiarity with AI, methodological expertise, institutional permissions, and willingness to use automated tools.
These differences are not necessarily problematic.
A statistician may find AI-assisted programming useful, while a qualitative researcher may be more cautious about delegating interpretation. Another collaborator may use AI only for language editing.
However, individual preferences become a governance problem when AI assistance affects shared research material or decisions for which multiple researchers are responsible.
For example, one collaborator's decision to upload confidential data can create consequences for the entire project.
Similarly, unverified AI-generated findings may be incorporated into a manuscript approved by several coauthors.
A shared agreement helps distinguish individual discretion from decisions requiring collective authorization.
What Is a Research Team AI-Use Agreement?
A research team AI-use agreement is a documented understanding of how AI systems may contribute to collaborative research and how resulting responsibilities will be allocated.
It may be incorporated into an existing collaboration agreement, laboratory policy, research management plan, or project-specific operating procedure.
It need not always be a separate legal document.
Its purpose is to translate applicable requirements into practical expectations for the team.
The agreement should identify what collaborators may do independently, when consultation is necessary, and which activities are not permitted.
It should also establish how the team will evaluate consequential AI contributions before accepting them into shared research outputs.
Begin With the Rules That Already Apply
Before establishing internal preferences, collaborators should identify the requirements governing their project.
Relevant sources may include institutional AI policies, research ethics approvals, data-sharing agreements, funder conditions, confidentiality commitments, and journal instructions.
The European Code of Conduct for Research Integrity, revised by ALLEA in 2023, identifies reliability, honesty, respect, and accountability as fundamental principles.
These principles provide a foundation for responsible collaboration, but they do not replace specific institutional rules.
A research team cannot authorize processing prohibited by a binding agreement merely because all members support it.
Where requirements conflict or remain unclear, the relevant institutional authority should be consulted.
Distinguish Individual Permissions From Collective Decisions
Not every AI interaction needs approval from the entire research team.
Routine activities may be covered by standing permission, provided they comply with applicable requirements.
For example, a collaborator might use AI to explain a publicly available statistical concept or improve the wording of a nonconfidential email.
By contrast, using AI to change participant eligibility criteria or modify the primary analysis may require collective methodological discussion and possibly formal protocol authorization.
The distinction should depend on the consequences of the activity rather than the number of prompts involved.
| AI-assisted activity |
Suggested agreement category |
Primary condition |
| General concept explanations |
Standing permission |
Verify consequential information before use |
| Preliminary literature-search terminology |
Standing permission |
Conduct appropriate independent searching |
| Nonconfidential language editing |
Standing permission where allowed |
Preserve scientific meaning and satisfy disclosure rules |
| AI-generated analytical code |
Conditional permission |
Require testing and appropriate review |
| Substantive methodological changes |
Prior team discussion |
Document justification and obtain required authorization |
| Processing restricted research data |
Specific authorization required |
Use an approved processing arrangement |
| AI-generated interpretation of principal findings |
Collective review where material |
Verify against actual evidence |
| Fabricating research evidence |
Prohibited |
Research outputs must represent actual procedures and observations |
These categories are recommendations, not universally binding rules. A team's agreement must reflect its actual research context.
Specify Which AI Tools and Environments Are Authorized
Different AI services may have different data-processing arrangements, technical capabilities, and contractual protections.
Even products offered by the same provider may differ according to account configuration or institutional agreement.
The team should establish which systems may be used for particular research tasks and information categories.
For example, a system approved for public literature exploration may not be authorized to process identifiable participant records.
Approval to access a tool does not automatically authorize every processing purpose.
The agreement should identify who can approve additional tools and what evidence is needed before introducing them.
Establish Confidentiality Rules Before Sharing Data
Collaborative research may involve participant records, unpublished findings, restricted institutional documents, or confidential intellectual property.
One collaborator's decision to submit such information to an AI service may affect obligations shared by the entire research team.
Consequently, data-processing boundaries should be established before AI is used with restricted material.
The team should determine which data may be processed, through which systems, and under what safeguards.
De-identification may reduce privacy risks but does not automatically eliminate them.
Where authorization is uncertain, the relevant processing should not proceed until clarification is obtained.
Why AI-Generated References Need an Assigned Verifier
Generative AI can produce plausible bibliographic details and summaries that are inaccurate or unsupported.
For example, a collaborator may include a genuine article while attributing a finding that the publication does not report.
The team should identify who is responsible for checking references and substantive claims before they enter the shared manuscript.
Verification should include examining original sources rather than merely confirming that a DOI resolves.
Assigning responsibility helps prevent each collaborator from assuming that someone else has already checked the evidence.
The broader issue of responsibility for verifying AI-assisted contributions requires both individual task ownership and collective scholarly accountability.
Establish Verification Procedures for AI-Generated Code
AI-generated code may be useful for data cleaning, statistical analysis, and visualization.
However, code can execute successfully while implementing incorrect transformations or unsuitable methods.
The team should determine who will inspect the implementation, test consequential operations, and evaluate the resulting outputs.
For example, a research assistant may generate preliminary code, while a qualified analyst verifies its behavior and the responsible researchers approve the analytical approach.
Wilson and colleagues (2017) describe good practices in scientific computing, including testing and documentation.
These principles remain applicable when the initial code is generated by AI.
Who Decides Whether an AI Recommendation Is Accepted?
The agreement should distinguish the generation of suggestions from the authority to make final research decisions.
AI may recommend alternative models, exclusion criteria, or interpretations.
However, consequential choices should be made by appropriately qualified and authorized researchers.
For example, a change to a prespecified analysis may require collective methodological justification and documentation.
It may also require additional approval when it affects an authorized research protocol.
The principle of meaningful human oversight requires the ability to question, reject, and correct automated recommendations.
Merely approving an AI output without examining its basis is insufficient.
How Should AI Use Be Documented?
Documentation should make consequential research decisions understandable and, where appropriate, reproducible.
A useful record may identify the AI system, the task performed, relevant output, verification undertaken, and final human decision.
For computational work, preserving tested scripts and analytical environments may be more useful than retaining every conversational exchange.
Complete prompt histories are not universally necessary.
However, studies investigating AI interactions directly may require systematic logging as part of their approved methodology.
The agreement should specify which records are necessary and how confidential information within them will be protected.
Agree on Internal Disclosure Before Journal Submission
Coauthors may need to understand which parts of the research involved consequential AI assistance before approving a manuscript.
For example, an AI-generated analytical procedure may influence findings that all authors are expected to evaluate.
The team should establish when collaborators must report AI use internally and what information the disclosure should contain.
The question of whether all coauthors should know which research activities involved AI concerns collective accountability rather than mere administrative reporting.
Internal disclosure should support verification and informed approval.
Distinguish Internal Disclosure From Journal Disclosure
Research teams should not assume that internal communication automatically satisfies publication requirements.
Journals and publishers may establish specific rules concerning AI-assisted technologies.
The International Committee of Medical Journal Editors provides recommendations on AI-assisted technologies and human author responsibility.
Other publication venues may apply different requirements.
The corresponding author should confirm the current instructions of the intended journal.
However, publication disclosure should reflect accurate information supplied by the relevant collaborators rather than assumptions about their individual practices.
Can One Collaborator Use AI Without Informing Others?
The answer depends on the activity and the team's established expectations.
Routine permitted assistance may not require individual notification.
However, material AI contributions to methods, analysis, evidence, or interpretation generally warrant communication because they affect shared scholarly responsibilities.
For example, a collaborator who uses AI to improve the grammar of a nonconfidential email may not need to inform every coauthor.
A collaborator who uses AI to determine which observations to exclude from the final analysis should not treat that decision as merely private tool use.
The distinction is explored further through undisclosed AI use among research coauthors.
How Should the Team Handle Different Institutional AI Policies?
Collaborators may belong to institutions with different AI permissions, data-handling requirements, or approved tools.
One institution may authorize a processing arrangement that another prohibits for the same shared dataset.
The team should identify applicable obligations before deciding how AI will be used.
Where requirements conflict, collaborators should seek institutional clarification or adopt an arrangement satisfying all applicable restrictions.
The agreement should not assume that permission from one institution automatically overrides another institution's requirements.
For international research collaborations involving AI, additional legal and contractual considerations may arise.
What If a Collaborator Does Not Want to Use AI?
Researchers may reasonably differ in their willingness to use generative AI.
Some may prefer established analytical tools or direct engagement with source material.
Others may consider AI useful for selected tasks.
The team should distinguish personal preference from binding requirements and methodological necessity.
Where AI is optional, collaborators may be able to use different authorized workflows while meeting shared standards for accuracy and accountability.
Where a proposed AI activity affects shared data or consequential decisions, the team should resolve the disagreement before proceeding.
How Should Disagreements Be Resolved?
The agreement should identify who has authority to decide particular matters and when institutional consultation is necessary.
For example, a disagreement about statistical modeling may require methodological review.
A dispute concerning confidential data processing may require guidance from an information-security or research ethics authority.
Questions about authorship or manuscript approval may involve the team's established publication agreement and relevant journal policies.
The principal investigator or corresponding author may coordinate these discussions, but their role does not automatically override binding institutional obligations.
Decisions should be documented where they materially affect the research.
Should AI Use Affect Authorship Decisions?
AI assistance does not automatically determine who qualifies as an author.
Authorship should be assessed according to the relevant disciplinary and publication criteria for human scholarly contributions.
The International Committee of Medical Journal Editors states that AI-assisted technologies should not be listed as authors because they cannot meet human authorship responsibilities.
However, the significance of a human collaborator's AI-assisted work may still need evaluation.
The team should distinguish tool use from the intellectual contribution and accountability of the researcher.
The separate question of how AI-assisted contributions affect academic authorship deserves careful attention when establishing publication arrangements.
How Should Research Teams Address Errors Discovered Later?
AI-assisted research may contain errors that become apparent after internal review or manuscript submission.
The agreement should identify how concerns will be reported, investigated, corrected, and communicated.
For example, if an AI-generated data transformation incorrectly excludes observations, the responsible researchers should evaluate the affected analyses and revise the findings where necessary.
Scientific correction should not be delayed merely because the original procedure was approved by a team member.
Potential misconduct concerns should be handled through applicable institutional procedures rather than assumed from the presence of an error.
Why the Agreement Should Be Reviewed During the Project
AI tools, research tasks, and institutional requirements may change.
An agreement established during literature exploration may not adequately address later processing of confidential participant records.
Similarly, manuscript submission may introduce publisher-specific disclosure requirements.
Research teams should revisit the agreement at meaningful milestones, particularly before introducing new AI systems or consequential processing activities.
Changes should be communicated and documented appropriately.
The agreement should remain a practical governance instrument rather than an inflexible list of assumptions established at the project's beginning.
