03 · What You Need to Know
What Collective Accountability Requires When Research Uses AI
AI Assistance Does Not Transfer Scholarly Responsibility
Researchers routinely use instruments, statistical software, databases, and other technologies to conduct scientific work.
Generative AI introduces additional forms of assistance by producing text, code, analytical suggestions, summaries, and interpretations.
However, the use of an automated system does not make that system responsible for the accuracy or integrity of the resulting publication.
The International Committee of Medical Journal Editors (ICMJE) states that AI-assisted technologies should not be listed as authors because they cannot assume the accountability required of scholarly authors.
The Committee on Publication Ethics (COPE) similarly distinguishes AI tools from accountable human authors.
These positions have an important practical implication.
If AI-generated material contains an error, the relevant human researchers must examine the problem and determine its consequences.
Explaining that a mistake originated from AI may describe how it occurred, but it does not resolve the authors' responsibilities.
Individual Responsibility and Collective Accountability Are Related but Different
Collaborative research involves specialized contributions.
One author may be responsible for conceptualization, another for statistical analysis, and another for qualitative interpretation.
It would be unreasonable to assume that every author possesses identical expertise across all aspects of a complex project.
Nevertheless, authorship involves responsibilities extending beyond an individual's immediate technical contribution.
Individual contribution responsibility
The responsibility to perform, verify, explain, and document the research activities assigned to a particular contributor.
Collective author accountability
The shared obligation, under the applicable authorship framework, to approve the work, understand how responsibilities are distributed, and help ensure that legitimate integrity concerns are appropriately investigated and resolved.
The ICMJE recommends that authors be accountable for all aspects of the work and ensure that questions about accuracy or integrity are appropriately investigated and resolved.
It also recognizes that authors should be able to identify which coauthors are responsible for specific parts of the work and have confidence in the integrity of their contributions.
These expectations provide a useful foundation for AI-assisted research, although authorship standards vary across disciplines and journals.
Why AI Can Create Accountability Gaps
AI-assisted workflows can make it difficult to identify who performed or verified particular research activities.
For example, one researcher may generate statistical code, another may execute it, and a third may incorporate the resulting table into the manuscript.
If the analysis contains an error, each collaborator may assume that someone else checked the relevant component.
The problem is not necessarily deliberate avoidance of responsibility.
It may arise because the team never explicitly assigned verification duties or distinguished between producing an output and approving its scientific validity.
This is why identifying human intellectual contributions in AI-assisted research matters for more than authorship credit.
Contribution transparency also helps determine who can explain consequential decisions and respond when problems emerge.
What Do Research Integrity Standards Require?
The European Code of Conduct for Research Integrity, revised by ALLEA in 2023, identifies reliability, honesty, respect, and accountability as fundamental principles.
These principles are relevant throughout the research process, including the preparation, verification, and publication of AI-assisted work.
Accountability includes taking responsibility for research activities and their consequences rather than treating the use of technology as a reason to avoid scrutiny.
The code also addresses responsible research practices, authorship, publication, and the handling of research integrity concerns.
For research teams, these principles support arrangements in which consequential AI-assisted activities remain attributable to identifiable human contributors.
However, the precise legal, contractual, or institutional responsibilities of particular collaborators depend on the applicable policies and agreements.
Why Collective Accountability Does Not Require Identical Expertise
Multidisciplinary research often depends on specialized knowledge that cannot reasonably be distributed equally among all authors.
For example, an educational researcher may contribute to conceptualization and interpretation, while a statistician develops and verifies the analytical model.
The educational researcher may not be able to reproduce every computational detail independently.
Nevertheless, the researcher should understand the analysis sufficiently to evaluate its relevance to the study's conclusions and know who is responsible for its technical verification.
Similarly, a statistician may not possess specialist knowledge of every theoretical framework used in the manuscript.
Collective accountability therefore requires appropriate division of responsibilities combined with communication and mutual confidence grounded in reasonable review.
It does not require pretending that every coauthor is an expert in every component.
Assign Responsibility for Consequential AI-Assisted Activities
Research teams should identify who is responsible for evaluating AI-generated outputs before those outputs become part of the final research record.
| AI-assisted activity |
Principal verification concern |
Illustrative human responsibility |
| Literature summarization |
Fabricated or misrepresented sources |
Researcher responsible for reviewing the original publications and synthesis |
| Statistical code generation |
Incorrect implementation or model specification |
Qualified analyst responsible for testing and methodological verification |
| Data cleaning |
Unjustified transformations or exclusions |
Researcher responsible for data integrity and documented cleaning decisions |
| Qualitative analysis |
Unsupported codes, themes, or interpretations |
Qualified qualitative researcher reviewing the underlying material |
| Manuscript drafting |
Inaccurate claims or unsupported conclusions |
Authors responsible for the relevant content and substantive review |
| Figure generation |
Misleading or inaccurate representations |
Researcher responsible for validating the data and figure |
| Final manuscript approval |
Unresolved errors or accountability gaps |
All authors according to the applicable authorship requirements |
These assignments are illustrative rather than universally prescribed.
Teams should adapt responsibilities to their actual expertise, project organization, and institutional requirements.
Why Verification Must Be More Than a Formal Approval
A researcher may be designated as responsible for reviewing AI-assisted work without having the information or expertise needed to evaluate it.
For example, a coauthor might approve an AI-generated statistical table after checking its formatting but without examining whether the analytical method is appropriate.
Such approval does not establish meaningful verification.
Research on automation bias and complacency suggests that people may give automated recommendations inappropriate weight or reduce monitoring under certain conditions.
Parasuraman and Manzey (2010) examine these problems in human interaction with automation.
Research teams should therefore ensure that review involves substantive examination of relevant evidence.
This is consistent with designing genuine rather than ceremonial human oversight.
The reviewer should be able to identify problems, request corrections, and reject unsuitable outputs.
How Should Teams Verify AI-Generated Research Code?
AI-generated code should be evaluated according to its intended scientific function.
Successful execution does not establish that the implementation is correct.
A script may run without errors while performing an inappropriate transformation, omitting relevant observations, or fitting a model inconsistent with the study design.
Verification may involve inspecting code, testing known cases, comparing outputs with independent calculations, and documenting computational dependencies.
For consequential analyses, specialist review may be appropriate.
Wilson and colleagues (2017) discuss good practices in scientific computing, including testing, documentation, and version control.
These principles remain relevant regardless of whether code was written manually or generated with AI.
How Should Teams Verify AI-Assisted Literature and Writing?
AI-generated literature summaries and manuscript passages may contain factual errors, fabricated references, or interpretations that exceed the available evidence.
Researchers should verify substantive claims against original publications and research records.
For example, a generated discussion may describe a correlational finding as evidence of a causal effect.
Correcting the wording requires examining the study design and determining what conclusions are justified.
Coauthors responsible for the relevant sections should perform or coordinate this review.
Other authors should have sufficient information to understand how the claims were established and raise concerns when necessary.
Review should not be reduced to checking whether the prose sounds academically convincing.
Why Coauthors Need Appropriate Awareness of AI Use
Coauthors cannot meaningfully assess some risks if they do not know that AI contributed to consequential parts of the research.
For example, a collaborator may reasonably assume that a statistical analysis was developed and checked by a qualified team member.
If the analysis was instead generated by AI and implemented without verification, the collaborator's understanding of the work is incomplete.
Teams should therefore establish expectations about communicating material AI assistance.
This does not necessarily require reporting every routine prompt to every coauthor.
The appropriate level of awareness depends on the activity, its consequences, and applicable policies.
The issue is examined more specifically in coauthor awareness of AI-assisted research activities.
Can One Coauthor Verify AI-Assisted Work on Behalf of the Team?
Yes, in appropriate circumstances.
Specialized verification may reasonably be assigned to a qualified contributor.
For example, a statistician may validate the analytical code while other coauthors review whether the reported findings support the manuscript's substantive conclusions.
However, the team should know that verification occurred, who performed it, and whether important limitations remain.
Delegating specialized review does not mean other authors have no responsibilities concerning the integrity of the publication.
Under frameworks such as ICMJE, authors retain broader obligations to participate in resolving questions about the work's accuracy or integrity.
The distinction between individual verification duties and shared accountability is central to determining who verifies AI-assisted work produced by a coauthor.
Why Corresponding Authorship Does Not Eliminate Other Authors' Responsibilities
The corresponding author ordinarily manages communication with the journal during submission, peer review, and publication.
Specific responsibilities depend on the journal and applicable authorship framework.
However, designating a corresponding author does not automatically transfer every scientific responsibility away from the remaining authors.
Similarly, the principal investigator may have leadership and institutional responsibilities without personally performing every specialized analysis.
Research teams should distinguish administrative coordination from responsibility for the accuracy and integrity of scholarly contributions.
When an error is identified, the corresponding author may coordinate the response, but relevant contributors should participate in investigating the affected work.
How Can Contribution Statements Support Accountability?
Contributor statements help identify the activities performed by individual researchers.
The CRediT taxonomy provides a standardized vocabulary for describing roles such as Conceptualization, Methodology, Software, Formal analysis, Writing – original draft, and Writing – review & editing.
These descriptions can help readers and collaborators understand how work was distributed.
However, CRediT is not itself an authorship eligibility standard.
Nor does assigning a role establish that the work was adequately verified.
For example, identifying a researcher as responsible for Formal analysis does not prove that the analytical code was correct.
Contribution statements should therefore be supported by actual research practices and appropriate verification records.
What Should Be Documented When AI Contributes to Research?
Documentation should be proportionate to the significance of the AI-assisted activity.
For consequential uses, relevant records may include the system used, the task performed, the human decision-maker, the verification procedure, and the resulting analytical or interpretive decision.
Research teams may also preserve code versions, analytical scripts, source-verification records, and substantive manuscript revisions.
These records can help reconstruct how a decision was made if concerns arise later.
However, indiscriminate retention of prompts or uploaded materials may create confidentiality and data-protection risks.
Documentation practices should comply with applicable institutional policies, research protocols, and legal requirements.
The objective is to make consequential work accountable without collecting unnecessary sensitive information.
How Should Teams Respond to an AI-Related Error?
Discovering an error does not automatically establish research misconduct.
Errors can arise from misunderstanding, inadequate verification, software problems, or other causes.
Research teams should investigate the nature and consequences of the problem before drawing conclusions about responsibility or intent.
A proportionate response may involve:
- Identifying the affected AI-assisted output and the relevant research records.
- Determining which analyses, claims, or conclusions depend on the output.
- Assigning qualified researchers to investigate and verify the problem.
- Correcting affected work and assessing whether the findings remain valid.
- Following applicable institutional and journal procedures for reporting or correcting the scholarly record.
The Committee on Publication Ethics provides guidance concerning corrections, retractions, and publication integrity.
The appropriate response depends on the seriousness of the error and its consequences for the reliability of the publication.
Teams should avoid both minimizing consequential errors and treating every AI-related mistake as evidence of deliberate misconduct.
Why Accountability Must Continue After Publication
Scholarly responsibility does not end when a manuscript is accepted.
Readers, reviewers, or other researchers may later identify problems in data processing, citations, statistical interpretation, or AI-assisted content.
Authors should be prepared to examine legitimate concerns and cooperate with appropriate corrective processes.
For example, an AI-generated analytical script may contain an error discovered after publication.
The team should determine whether the error affects reported estimates, conclusions, or the reliability of the work.
If correction of the published record is necessary, the authors should follow the journal's procedures.
Collective accountability therefore includes maintaining sufficient knowledge and records to investigate the work after publication.
How Can Research Teams Prevent Responsibility From Becoming Diffused?
Responsibility diffusion can occur when several collaborators assume that someone else has checked an AI-assisted output.
For example, an analyst may believe that the principal investigator will review the interpretation, while the principal investigator assumes that the analyst has already verified everything.
Clear assignments can reduce this ambiguity.
Teams should identify who performs technical verification, who evaluates the substantive implications, and who approves consequential changes.
These responsibilities should be communicated rather than inferred from job titles alone.
The broader objective is to prevent AI recommendations from acquiring authority without meaningful human evaluation.
Why Accountability Is Not the Same as Blame
Accountability involves being answerable for decisions and participating in appropriate investigation and correction.
It should not automatically be equated with assigning blame whenever an error occurs.
A researcher may have followed a reasonable verification procedure and still encounter an unexpected technical problem.
Another may have knowingly bypassed required checks.
These situations warrant different assessments.
Research integrity processes should examine the evidence, applicable standards, and circumstances rather than assume that every mistake reflects misconduct.
A constructive accountability system encourages researchers to identify and correct problems without obscuring legitimate responsibilities.
Can Collective Accountability Be Preserved in International Research Teams?
International collaborations may involve different institutional policies, disciplinary conventions, data-protection requirements, and publishing expectations.
These differences can complicate AI-assisted workflows.
For example, one institution may permit an AI service that another restricts for confidential research data.
Research teams should establish shared procedures that respect the obligations applicable to participating institutions.
Agreements may identify authorized tools, data-processing restrictions, verification responsibilities, and procedures for resolving conflicts.
These arrangements can form part of research collaboration agreements addressing AI use.
Collective accountability does not require identical institutional policies, but it does require sufficient coordination to ensure that shared research obligations are fulfilled.
