03 · What You Need to Know
How to Distinguish Disclosure Problems From Research-Integrity Violations
Research Integrity Is Broader Than Compliance With an AI Disclosure Policy
Research integrity concerns the responsible conduct and communication of research. The Singapore Statement on Research Integrity identifies honesty, accountability, professional courtesy and fairness, and good stewardship of research as fundamental principles.
These principles apply regardless of whether a researcher uses generative AI, conventional statistical software, laboratory equipment, or entirely manual procedures.
An AI disclosure requirement serves a particular purpose within this broader framework. It helps readers, editors, and reviewers understand how AI contributed to the work. However, compliance with a disclosure policy does not automatically establish that the underlying research is sound.
A researcher could accurately disclose AI-assisted analysis while still relying on invalid calculations. Conversely, an author might omit a required declaration despite having independently verified every scientific claim.
Both situations warrant attention, but they are not the same problem.
A Disclosure Violation Is Not Automatically a Formal Finding of Misconduct
Research-integrity concerns, publication-policy violations, and formal research misconduct are related but distinct categories.
Publication-policy violation
Failure to comply with an applicable journal requirement, such as a mandatory declaration of substantive AI-assisted writing.
Research-integrity concern
Conduct or circumstances that raise credible questions about honesty, accountability, research reliability, ethical obligations, or the accuracy of the scholarly record.
Formal research misconduct
Conduct established under the specific definitions, evidentiary standards, and procedures of the responsible institution, funder, or jurisdiction.
For example, the United States Public Health Service definition of research misconduct addresses fabrication, falsification, and plagiarism in proposing, performing, reviewing, or reporting research. It does not classify every editorial-policy violation as research misconduct.
Other institutions and jurisdictions may recognize broader categories of unacceptable research practice. Researchers should therefore avoid treating one legal or institutional definition as universally applicable.
The fact that an AI disclosure was omitted may justify examination. It does not, without further evidence, establish a particular misconduct offense.
When Undisclosed AI Use Conceals How Evidence Was Produced
A significant concern arises when the published methodology does not accurately describe the procedures used to produce research findings.
Suppose a qualitative study states that two researchers independently coded interview transcripts, but the initial coding was actually performed by a generative AI system and the researchers accepted most of its categories without documented review.
The omission is not merely about acknowledging software assistance. It may misrepresent the analytical procedure and the basis for the study's claims.
Similarly, an article describing a manually constructed analytical model may be materially incomplete if an AI system selected variables, generated code, and recommended analytical decisions that shaped the final results.
Not every methodological omission invalidates a study. Its significance depends on whether the missing information affects readers' ability to evaluate the research, understand its limitations, or reconstruct the procedure.
When AI Produces Fabricated or Falsified Research Material
Generative AI systems can produce plausible but nonexistent citations, inaccurate descriptions of sources, invented numerical values, or synthetic text resembling empirical evidence.
If such material enters a manuscript as genuine evidence, the problem extends beyond disclosure.
For example, an author who presents AI-generated participant quotations as statements from actual interviewees misrepresents the underlying research record. Likewise, invented observations presented as collected data raise concerns about fabrication, regardless of whether AI produced them.
It is important to distinguish an AI-generated falsehood that an author negligently failed to identify from intentional fabrication or falsification. Both can undermine a publication, but formal classifications of misconduct depend on the applicable standards and evidence.
A disclosure stating that AI was used does not make fabricated empirical evidence acceptable.
When Undisclosed AI Use Compromises Research Reliability
Research reliability may be affected when AI-generated outputs are accepted without adequate validation.
Consider an AI-generated statistical script that incorrectly recodes missing values as zero. If the error changes the reported treatment effect, the findings may no longer support the article's conclusions.
The omission of AI involvement creates a transparency problem. The analytical error creates a separate reliability problem.
COPE's retraction guidance emphasizes the integrity of the scholarly record, including situations in which findings are unreliable. Editorial action should reflect the actual defect rather than treating the use of AI itself as grounds for retraction.
Where a limited error can be corrected without undermining the overall findings, a correction may be appropriate. Where the central results cannot be relied upon, more substantial action may be necessary.
When AI Use Violates Participant Confidentiality or Research Ethics Requirements
Undisclosed AI use may also raise integrity concerns even when the AI-generated output is scientifically accurate.
A researcher might upload identifiable interview transcripts, confidential clinical records, or unpublished proprietary data to an external AI service without confirming that the transfer is authorized.
The resulting privacy or confidentiality concern exists independently of whether the system produces a useful analysis.
Applicable obligations may arise from informed consent, institutional ethics approval, data-sharing agreements, information-security policies, contractual restrictions, or data protection law.
These obligations vary across jurisdictions and research contexts. Researchers should not assume that removing participant names is sufficient to anonymize a dataset or authorize its transfer.
Where a potentially unauthorized disclosure has occurred, appropriate institutional or privacy procedures may be required alongside any journal communication.
When an Author Knowingly Conceals AI Involvement
Deliberate concealment can raise a different concern from an inadvertent omission.
For example, a researcher might knowingly provide a false answer to a mandatory submission question asking whether generative AI was used, despite understanding that the journal requires disclosure.
Such conduct may implicate honesty and compliance with publication obligations. Its formal classification depends on the applicable institutional or editorial framework.
However, intent should not be inferred solely from the absence of a disclosure. The author may have misunderstood the policy, relied on an outdated version, or been unaware of a coauthor's AI activity.
Establishing what the researcher knew, what was required, and what was communicated is essential before drawing conclusions about deliberate concealment.
Does the Extent of AI Use Determine Whether Research Integrity Is at Risk?
The amount of AI-generated content is an inadequate standalone measure.
A manuscript containing several pages of AI-assisted language editing may retain accurate scientific content. By contrast, a single fabricated citation or altered participant quotation could materially affect a central claim.
The relevant issue is the relationship between the AI contribution and the scientific or ethical obligation at stake.
| Undisclosed activity |
Potential concern |
What requires examination |
| Basic language correction |
Possible disclosure compliance issue |
Applicable journal policy and whether meaning changed |
| Substantive manuscript drafting |
Transparency and accuracy of scholarly claims |
Disclosure requirements and independent verification |
| AI-assisted data analysis |
Methodological reporting and reliability |
Code, analytical decisions, inputs, and results |
| AI-generated empirical evidence presented as real |
Possible fabrication or falsification |
Authenticity of the underlying research records and applicable misconduct standards |
| Unauthorized processing of participant information |
Confidentiality, privacy, or ethics breach |
Consent, approvals, data protections, and reporting obligations |
| Knowingly false AI declaration |
Possible deliberate misrepresentation |
Policy requirements, submission records, intent, and consequences |
This classification identifies concerns requiring assessment. It is not a universal misconduct taxonomy or a substitute for institutional investigation.
Who Determines Whether Undisclosed AI Use Constitutes Misconduct?
Different authorities have different responsibilities.
Journal editors evaluate the reliability and accuracy of the publication and determine appropriate editorial action. Research institutions and designated research-integrity bodies may investigate allegations of misconduct under their governing procedures.
Ethics committees, privacy officers, funders, or regulatory authorities may have responsibilities when participant protection or data handling is implicated.
These processes can overlap without producing identical determinations.
For example, a journal may issue a correction because a methods description is incomplete, while an institution separately determines that the omission does not meet its formal definition of misconduct.
Similarly, an institution may investigate deliberate misrepresentation even when the published findings remain scientifically reliable.
What Evidence Is Needed Before Escalating a Concern?
A credible assessment should rely on identifiable facts rather than assumptions about AI use.
Relevant records may include saved AI interactions, draft histories, source data, analysis scripts, research protocols, submission declarations, correspondence, and applicable policies.
Researchers should preserve original records and document what is known, what remains uncertain, and which parts of the research may be affected.
AI-text detection scores alone are not sufficient to establish that a particular author used generative AI or deliberately concealed its use. Such tools may produce false positives and false negatives, and their outputs require careful interpretation.
Where a serious concern exists, the appropriate response is a fair, evidence-based assessment through established procedures, not a presumption of guilt.
