03 · What You Need to Know
How to Respond to Undisclosed AI Use Within a Research Team
First, Establish the Facts Before Assigning Blame
The initial response should be a factual inquiry rather than an accusation.
Researchers should determine which AI system was used, what information was supplied, which outputs were generated, and how those outputs entered the research workflow.
Ask the coauthor to explain the activity and provide relevant records where available and appropriate.
For example, the coauthor may have used AI to correct grammar, generate a statistical script, summarize publications, or classify participant responses.
These activities require different forms of verification.
The team should also establish whether AI use had been discussed previously, whether a formal policy existed, and whether the coauthor understood the relevant expectations.
Using AI without informing collaborators is not automatically equivalent to intentionally deceiving them.
The distinction between misunderstanding and concealment should be evaluated from the available evidence rather than inferred from the mere absence of disclosure.
What Does Undisclosed AI Use Actually Mean?
The phrase may describe several different circumstances.
A coauthor may have used AI without telling the team, without informing the corresponding author, without documenting the contribution, or without including a disclosure required by the target journal.
These situations overlap but are not identical.
Undisclosed to collaborators
AI assistance was not communicated to other research team members, whether or not a formal team reporting requirement existed.
Undocumented in research records
The AI contribution was not recorded sufficiently to understand or reconstruct its role in the workflow.
Undisclosed to the journal
Required information about AI assistance was omitted from the manuscript or submission process.
Unauthorized AI use
The activity may have violated an applicable restriction concerning tools, data, procedures, or approvals.
A research team should identify which problem occurred before deciding what corrective action is necessary.
For example, a coauthor may have informed the journal about AI-assisted language editing but failed to tell collaborators. Conversely, the team may have known about the AI assistance but omitted a required publication disclosure.
The appropriate response depends on the actual circumstances.
Determine What the AI Contributed to the Research
The nature of the AI contribution is central to assessing its significance.
| Undisclosed activity |
Potential concern |
Recommended initial review |
| Grammar and readability assistance |
Possible disclosure omission or unintended meaning changes |
Compare original and revised passages; check applicable policy. |
| Substantive manuscript drafting |
Unsupported claims, inaccurate citations, or unclear attribution |
Verify claims and source material; establish the extent of generated content. |
| Literature searching and summarization |
Fabricated references or misrepresented findings |
Retrieve original publications and verify claim-source alignment. |
| Statistical programming |
Incorrect transformations, model specifications, or results |
Inspect code, validate operations, and reproduce consequential analyses. |
| Qualitative coding |
Misclassification or loss of contextual meaning |
Review coding criteria and compare outputs with original data. |
| Processing participant information |
Privacy, consent, confidentiality, or security concerns |
Establish what data were supplied and consult responsible institutional authorities where necessary. |
| Generating figures or analytical summaries |
Inaccurate representation of data or results |
Compare generated outputs with verified research records. |
These categories are practical distinctions rather than a universal hierarchy of misconduct.
The same activity may have different consequences depending on the study and applicable requirements.
Should the Team Immediately Remove All AI-Generated Material?
Not necessarily.
Removing every AI-assisted contribution before establishing its nature may be unnecessary and could disrupt valid research work.
However, consequential outputs should not continue to be used without adequate verification.
For example, AI-generated code that produced the primary results may need to be independently tested before the manuscript proceeds.
An AI-generated literature summary may require source verification before its claims are retained.
If the activity involved prohibited data sharing, immediate containment and institutional consultation may be necessary.
The appropriate response is therefore risk-based.
Preserve relevant original materials, suspend reliance on uncertain consequential outputs, and determine whether the work can be validated or must be replaced.
What if the Coauthor Believed AI Use Was Too Minor to Mention?
The team should examine whether that belief was reasonable under the applicable policy and research context.
Some researchers may regard routine language editing as comparable to conventional proofreading software.
Others may not recognize that AI-assisted programming or automated data classification can constitute a methodological contribution requiring documentation.
These misunderstandings are possible, particularly when institutional and journal policies differ.
However, personal beliefs do not automatically override explicit requirements.
If the team had established a rule requiring members to report all generative AI assistance, that expectation should be considered.
The response may involve clarification, training, documentation, or formal corrective action depending on the circumstances.
Researchers should distinguish a reasonable misunderstanding from knowingly disregarding an established requirement.
What if the Coauthor Deliberately Concealed AI Use?
Deliberate concealment raises different concerns from an inadvertent omission.
For example, a coauthor may have explicitly denied using AI after being asked to disclose relevant assistance.
Alternatively, the person may have concealed AI-generated analyses because they knew the team required independent validation.
Such conduct may involve misrepresentation or breach of collaborative responsibilities.
However, a formal finding of misconduct cannot be inferred solely from suspicion.
The team should preserve relevant evidence, establish what representations were made, and follow applicable institutional procedures when serious concerns arise.
The distinction between misunderstanding an AI policy and deliberately concealing AI use requires attention to both the rules and the available evidence concerning conduct.
Who Is Responsible for Verifying the AI-Assisted Work?
Verification responsibilities should reflect the research team's roles and expertise.
The coauthor who used AI should explain the contribution and provide relevant materials needed for review.
However, independent verification may require another qualified team member or specialist.
For example, a statistician may need to examine AI-generated analytical code, while researchers responsible for qualitative analysis may need to review AI-generated themes.
The principal investigator or corresponding author may coordinate the process according to the project's governance arrangements.
The International Committee of Medical Journal Editors emphasizes that authors should be accountable for their contributions and cooperate in resolving questions concerning the accuracy or integrity of the work.
That does not mean every coauthor must personally perform every technical calculation. It means the team must have a defensible process for establishing the reliability of its shared research.
What if the AI Output Is Scientifically Correct?
Accurate AI-assisted work may still create a reporting or governance issue if its use was not disclosed as required.
Suppose a coauthor used AI to generate a statistical script, then independently validated every calculation.
The analysis may be scientifically defensible, but the team may still need to document the AI contribution or disclose it to the journal.
Conversely, disclosure alone does not establish that the analysis is valid.
The research team should distinguish the accuracy of the output from compliance with reporting requirements.
Neither should be assumed to resolve the other.
What if AI Introduced Errors Into the Shared Research?
If verification reveals inaccuracies, the team should correct the affected work before proceeding.
The necessary response depends on the error.
Incorrect code may require reanalysis. Fabricated citations may require revising the literature review. Distorted participant summaries may require returning to the original transcripts and reconsidering affected themes.
The team should determine which downstream results or conclusions were influenced.
The broader process for responding when AI introduces a research error includes containment, independent verification, correction, and appropriate communication.
Blaming the coauthor or AI system does not replace the obligation to restore the reliability of the research.
What if Confidential Participant Data Were Entered Into AI?
This possibility requires particular attention.
A coauthor may have uploaded interview transcripts, survey responses, institutional records, or other restricted material to an external AI service.
The research team should establish what information was shared, which service received it, what contractual or technical protections applied, and whether the activity was permitted.
Do not assume that data were protected merely because the service was widely used or because the coauthor believed the information was anonymous.
De-identified data may still contain contextual details that permit reidentification.
Where a possible privacy or ethics breach occurred, the team should follow institutional incident-response and reporting procedures.
Depending on the circumstances, consultation with an ethics committee, data-protection officer, information-security team, or other responsible authority may be necessary.
A publication disclosure does not retrospectively authorize prohibited processing of participant information.
Should the Team Tell the Journal?
The answer depends on the manuscript's stage, the applicable policy, and the nature of the AI contribution.
If the manuscript has not been submitted, the team should resolve the issue internally and prepare the required disclosure before submission.
If the manuscript is under review and a required disclosure was omitted, the corresponding author should consult the journal about updating the submission.
If the undisclosed AI use materially affected methods, results, or research integrity, more extensive editorial communication may be necessary.
If the article has already been published, the journal may need to consider a correction or other documented action.
Researchers should not assume that every undisclosed AI interaction requires the same editorial response.
The appropriate process for correcting missing AI disclosures begins with establishing what the applicable policy required.
What if the Coauthor Refuses to Explain the AI Use?
A refusal to provide information necessary for verifying consequential research work can prevent the team from establishing its reliability.
Researchers should first clarify what information is needed and why.
For example, the team may require the generated code and analytical specification rather than an unrestricted copy of every private AI conversation.
Requests should be proportionate and consistent with confidentiality, institutional policy, and legitimate research oversight.
If the coauthor continues to withhold information essential to verification, the responsible research leaders may need to pause approval of affected work and seek institutional guidance.
The team should avoid making unsupported allegations while ensuring that unverified consequential outputs are not represented as reliable findings.
What if Coauthors Disagree About Whether AI Use Was Acceptable?
Disagreement may concern scientific validity, publication policy, research ethics, or the team's expectations.
These questions should be considered separately.
One coauthor may object to AI assistance as a matter of personal scholarly preference. Another may regard it as acceptable when adequately verified and disclosed.
Personal disagreement does not necessarily establish a policy violation.
However, a team member's preference cannot override an applicable prohibition or a requirement for ethics approval.
Researchers should identify the governing standards, examine the actual contribution, and determine whether the resulting work is scientifically defensible.
Where a serious disagreement remains unresolved, institutional research-integrity or supervisory procedures may provide an appropriate route for resolution.
Does Undisclosed AI Use Affect Authorship?
AI assistance does not automatically determine whether a human contributor qualifies for authorship.
Authorship criteria concern substantive contributions and associated responsibilities under the applicable disciplinary or journal standards.
The ICMJE recommends that AI tools should not be listed as authors because they cannot fulfill authorship responsibilities.
A coauthor who used AI may still qualify for authorship based on their genuine intellectual contributions and accountability.
Conversely, generating text through AI does not automatically establish that a person satisfies every authorship criterion.
If undisclosed AI use reveals that a contributor's actual role was materially different from what the team understood, authorship and contribution statements may need review.
Such decisions should follow the relevant standards rather than treating AI use itself as an automatic basis for adding or removing an author.
Should the Team Remove the Coauthor?
Removing a coauthor should not be an automatic response to undisclosed AI assistance.
Authorship changes can have serious professional consequences and may be governed by journal and institutional procedures.
The team should establish whether the coauthor satisfies the applicable authorship criteria, whether the AI use affected the integrity of their contribution, and whether any separate conduct concerns require investigation.
If the manuscript has already been submitted, changes to the author list may require journal approval and agreement among the relevant authors.
Disputed authorship should be handled through appropriate institutional processes rather than unilateral retaliation.
The immediate scientific priority remains verification and correction of the affected research.
Does Undisclosed AI Use Automatically Constitute Research Misconduct?
No.
Formal research-misconduct definitions vary by jurisdiction and institution.
Under the United States Public Health Service framework, research misconduct concerns fabrication, falsification, or plagiarism within the scope of the applicable rules and required standards.
Undisclosed AI use does not automatically satisfy those definitions.
However, the circumstances may raise concerns about knowingly false declarations, fabricated research material, unauthorized data processing, or other violations.
The team should distinguish the absence of communication from evidence of deliberate misrepresentation.
Where serious concerns arise, formal assessment should follow institutional procedures rather than informal assumptions about intent.
What if the Manuscript Has Already Been Approved by All Authors?
Author approval should be based on an accurate understanding of the work being submitted.
If a consequential AI contribution was unknown when coauthors approved the manuscript, the team may need to reconsider the affected sections or analyses.
For example, an author may have approved statistical results without knowing that the primary script was generated through AI and had not been independently validated.
The new information may justify additional review before submission or continued editorial consideration.
It does not automatically invalidate the earlier approval, but it may reveal that the information available to the authors was incomplete.
The team should document any material re-verification and obtain renewed agreement where necessary.
How Should the Research Team Document the Incident?
A proportionate record should establish what happened and how the issue was resolved.
Relevant information may include the AI tool, tasks performed, approximate dates, materials processed, outputs incorporated, verification undertaken, applicable policies, and final corrective decisions.
When available and necessary, preserve relevant prompts, generated code, manuscript versions, and analytical records.
Do not collect or circulate confidential participant information unnecessarily.
The record should distinguish established facts from unresolved questions.
If institutional investigation becomes necessary, preserve relevant evidence and follow the responsible authority's procedures.
A clear record can help prevent later disagreement about which AI contributions were known, verified, and approved.
How Can Teams Prevent Undisclosed AI Use?
Research teams should establish expectations before consequential AI-assisted activities begin.
A project-level agreement can identify permitted tools, prohibited data inputs, tasks requiring prior approval, verification responsibilities, and disclosure procedures.
Researchers should also distinguish routine assistance from activities that materially affect research evidence or conclusions.
For example, AI-generated code used in the primary analysis might require documented review by a qualified analyst, while routine language assistance may follow a simpler reporting procedure.
Periodic discussion of AI use can help identify contributions that would otherwise remain undocumented.
Before submission, the corresponding author should confirm that each contributor has reported relevant AI assistance and reviewed the final disclosure statement.
Such procedures are most effective when expectations are clear, proportionate, and consistent with applicable policies.
Build a Collaborative Culture That Encourages Accurate Reporting
Research teams may inadvertently discourage transparency if every disclosure of AI use is treated as evidence of incompetence or wrongdoing.
Conversely, an environment in which AI assistance is considered entirely inconsequential may encourage researchers to bypass verification and reporting.
A more defensible approach distinguishes permitted assistance from prohibited activity and makes the relevant responsibilities explicit.
Team members should be able to report AI use, ask questions about policy, and raise concerns about output reliability without assuming that disclosure automatically triggers punishment.
This does not remove accountability for deliberate concealment or serious breaches.
It makes accurate reporting more likely and enables the team to identify problems before they affect the scholarly record.
Watch Out
Do not assume that undisclosed AI use is harmless merely because the resulting text appears accurate. Equally, do not assume that the coauthor committed misconduct solely because AI assistance was not communicated. Establish the activity, verify consequential outputs, assess applicable requirements, and use fair procedures when evaluating responsibility.
