03 · What You Need to Know
What Makes Undisclosed AI Use More or Less Serious?
The Task Matters More Than the AI Tool's Name
A tool such as ChatGPT can correct grammar, propose research questions, summarize literature, generate code, classify text, or interpret findings. These functions do not carry identical scientific implications.
For example, correcting punctuation generally does not change how research evidence is generated. By contrast, using AI to assign qualitative codes may influence the analytical categories from which conclusions are developed.
The appropriate assessment therefore begins with the function performed rather than the brand or apparent sophistication of the AI system.
Researchers should identify the input, the requested operation, the output, and the way that output entered the research workflow. The same system can perform a low-consequence task in one study and a highly consequential task in another.
How Directly Did AI Influence the Scientific Evidence?
One useful consideration is the proximity of the AI contribution to the evidence supporting the article's claims.
Assistance with spelling is generally distant from the production of scientific evidence. AI-generated data transformations, coding decisions, analytical models, and interpretations may be much closer to the study's substantive conclusions.
This distinction does not mean that every AI-assisted analytical task is inappropriate. Researchers routinely use computational tools, including machine-learning methods, in legitimate scientific work. The concern arises when a consequential contribution is not adequately documented, validated, or disclosed.
| AI contribution |
Potential scientific consequence |
Primary assessment question |
| Spelling and punctuation correction |
Usually limited when meaning remains unchanged |
Was the activity exempt from disclosure? |
| Substantive manuscript rewriting |
May change claims, emphasis, or interpretation |
Were meanings preserved and disclosure requirements met? |
| Literature synthesis |
May introduce unsupported claims or misrepresent sources |
Were all sources and attributed claims independently checked? |
| Qualitative coding |
May influence categories, themes, and interpretations |
Can coding decisions be traced to the original data? |
| Statistical code generation |
May introduce computational or methodological errors |
Were the code and results independently validated? |
| Research image generation or alteration |
May misrepresent evidence or violate journal restrictions |
Does the image accurately represent the underlying research? |
| Processing confidential participant data |
May create privacy, consent, or information-security concerns |
Was the transfer authorized and adequately protected? |
These categories illustrate possible risks rather than assigning universal severity levels. A seemingly minor writing change can become consequential if it reverses a finding, while thoroughly validated analytical assistance may not compromise scientific reliability.
Did AI Generate, Transform, or Merely Present Information?
Another important distinction concerns the relationship between AI output and the underlying research material.
Presentation assistance
AI helps express information already established through independently supported research, without materially changing its meaning.
Evidence-related assistance
AI generates, selects, transforms, classifies, or interprets information used to support scientific claims.
Presentation assistance may still require disclosure under a particular publisher's policy. Nevertheless, evidence-related assistance often warrants additional methodological scrutiny because it can affect the substance of the study.
Suppose AI rewrites a paragraph explaining an independently verified regression result. The principal concern may be whether the explanation remains accurate. If AI instead selects variables, generates the regression code, and interprets the model without adequate verification, the concern extends to the validity of the analytical process.
Was the AI Output Independently Verified?
The existence of independent verification can materially change the assessment.
An AI-generated script that was inspected, tested against known results, and incorporated into a reproducible analytical workflow presents a different evidentiary situation from an unexamined script accepted solely because it produced plausible numbers.
Likewise, AI-assisted qualitative coding that was reviewed against original transcripts and documented through a transparent analytical procedure may be more defensible than classifications accepted without checking their relationship to participants' statements.
Verification does not eliminate disclosure obligations. It does, however, help establish whether the AI contribution introduced errors or uncertainty into the research.
When verification records are unavailable, researchers should not assume the work was reliable simply because the published conclusions appear reasonable.
Did the Omission Conceal a Material Methodological Detail?
Scientific reporting should provide enough information for readers to understand how evidence was produced and evaluated. Reporting standards such as CONSORT and STROBE illustrate the broader principle that relevant methodological decisions should be described transparently, although their requirements are specific to the study designs they address.
When AI participates in a consequential research procedure, omitting that involvement may leave readers with an inaccurate account of the method.
For example, a manuscript describing exclusively human thematic coding may be misleading if AI-generated categories substantially influenced the analysis. Similarly, a study may require clarification if AI was used to construct or transform variables but the methods section describes a different procedure.
The seriousness depends partly on whether the missing information would reasonably affect methodological evaluation, interpretation, or attempts to reproduce the work.
Did AI Introduce an Error or Make the Results Unreliable?
Undisclosed AI use becomes particularly consequential when the contribution introduces an error affecting the reported evidence.
Possible problems include fabricated citations, incorrect statistical calculations, misclassified observations, altered participant quotations, unsupported interpretations, or misleading research images.
These problems should be assessed according to their effects. An incorrect reference that can be replaced without changing the argument may be correctable. An analytical error that invalidates the study's central findings may require substantially more serious editorial action.
COPE's retraction guidance focuses on circumstances in which findings are unreliable or other recognized grounds for retraction exist. The fact that AI contributed to an error does not, by itself, determine whether a correction or retraction is appropriate.
Could the AI Activity Have Harmed Participants or Compromised Confidentiality?
Scientific reliability is not the only relevant dimension. AI use may also create risks for research participants, particularly when identifiable or sensitive information is transferred to a system without appropriate authorization.
A model may produce an accurate summary of confidential interview material while the underlying data transfer still violates an applicable privacy, consent, contractual, or ethics requirement.
In such circumstances, the concern cannot be resolved solely by verifying the AI-generated text or adding a disclosure statement.
The research team may need to consult its ethics committee, data protection officer, information-security personnel, or other responsible authority. Whether external notification is legally required depends on the jurisdiction, the information involved, and the applicable obligations.
Does Deliberate Concealment Make the Situation More Serious?
Intent is relevant but should not be inferred from nondisclosure alone.
An author may misunderstand an evolving policy, overlook a required declaration, or fail to learn that a coauthor used AI. Another author may knowingly conceal prohibited AI involvement after receiving explicit instructions.
These situations may warrant different institutional responses even if the underlying AI activity was similar.
Nevertheless, an innocent omission can still require correction, while deliberate nondisclosure does not automatically establish that the scientific findings are false.
It is therefore necessary to distinguish the reliability of the published research from the circumstances surrounding the author's conduct.
Can the Seriousness of Undisclosed AI Use Be Expressed as a Score?
A structured assessment can improve consistency, but there is no universally accepted numerical severity score for undisclosed generative AI use across scholarly disciplines and publishers.
Researchers may examine several dimensions: the significance of the AI task, the strength of independent verification, the consequences of any errors, applicable disclosure obligations, privacy risks, and evidence concerning intent.
These dimensions should not simply be added into an arbitrary total. Some concerns, such as unauthorized disclosure of sensitive participant information or unreliable central findings, may independently warrant urgent action regardless of how other dimensions are rated.
A defensible assessment explains the reasons for its conclusion rather than presenting a numerical score as though it were a validated research-integrity instrument.
