03 · What You Need to Know
How to Recognize Research-Integrity Problems Involving AI
Research integrity concerns the reliability of the research process, not simply the tools used
Research integrity requires researchers to represent their methods, evidence, findings, and contributions accurately. It also involves maintaining records that allow others to understand how conclusions were reached.
Generative AI introduces additional opportunities for error because it can produce convincing text, references, images, code, and numerical information without independently establishing their truth.
Nevertheless, using AI is not inherently a research-integrity violation.
A researcher who uses AI to improve grammar, generate programming suggestions, or organize preliminary ideas may be acting appropriately, provided that the activity complies with applicable requirements and the resulting work is properly verified.
The integrity concern emerges when AI involvement results in unreliable evidence, misleading reporting, prohibited conduct, or failures of accountability.
The International Committee of Medical Journal Editors (ICMJE) emphasizes that human researchers remain responsible for the accuracy and integrity of material produced with AI assistance.
What is the difference between an AI error, a questionable practice, and research misconduct?
These categories should not be treated as equivalent.
| Category |
What it means |
Illustrative situation |
| AI-related error |
An incorrect output or mistake that may occur without misconduct. |
An AI coding suggestion produces an incorrect calculation that is identified and corrected during validation. |
| Questionable research practice |
Conduct that may undermine research quality or transparency without necessarily meeting a formal misconduct definition. |
A researcher relies heavily on AI-generated interpretations without documenting the analytical decisions. |
| Policy violation |
Conduct that breaches an applicable institutional, journal, funder, or contractual requirement. |
A researcher fails to disclose AI assistance where the journal expressly requires disclosure. |
| Research misconduct |
Conduct satisfying a governing formal definition and its evidentiary requirements. |
A researcher knowingly submits AI-fabricated experimental observations as genuine results. |
The United States Office of Research Integrity (ORI) defines research misconduct in the relevant federal framework around fabrication, falsification, and plagiarism in proposing, performing, reviewing, or reporting research.
Its framework also distinguishes misconduct from honest error and differences of opinion. Formal findings require the applicable evidentiary and culpability standards to be satisfied.
Other institutions, funders, and countries may adopt broader definitions of unacceptable research practices. A violation under one policy does not necessarily receive the same classification elsewhere.
When does AI-generated data become fabrication?
Fabrication occurs when data or results are made up and recorded or reported as though they were genuine research evidence.
Generative AI can produce realistic-looking survey responses, interview transcripts, laboratory measurements, and datasets. The ability to produce plausible information does not make that information an actual observation.
For example, a researcher who asks AI to generate responses from 100 hypothetical participants and then reports those responses as if 100 people completed a survey misrepresents the origin of the evidence.
The central issue is not that the data were generated by AI. It is that generated observations were falsely represented as empirical data.
By contrast, appropriately disclosed synthetic data may be legitimate for software testing, simulation, or methodological development.
The distinction between using AI to fabricate research data and legitimate synthetic data generation depends on the purpose, provenance, reporting, and claims made from the material.
When does AI-assisted analysis become falsification?
Falsification concerns manipulating research materials, processes, data, or results, or changing or omitting information so that the research record no longer accurately represents what occurred.
AI can facilitate such conduct when it is used to alter results, remove inconvenient observations without methodological justification, or create misleading representations of experimental evidence.
Consider a researcher who instructs an AI system to modify a dataset until the statistical analysis produces a desired significant result, then reports the modified values as the original observations.
The resulting problem is not ordinary statistical assistance. The researcher has misrepresented the evidence.
However, not every AI-assisted data transformation constitutes falsification. Data cleaning, outlier handling, imputation, and other analytical procedures may be legitimate when methodologically justified, documented, and accurately reported.
The critical question is whether the procedure preserves an honest account of the research or conceals changes that materially distort its findings.
Can AI-generated references create a research-integrity problem?
Yes. Generative AI can produce citations that appear credible but refer to nonexistent publications, contain incorrect bibliographic details, or misrepresent what a real source establishes.
A fabricated reference can compromise the evidentiary support of a manuscript, especially when it is used to justify a research gap, theoretical claim, or methodological decision.
Researchers should distinguish an accidental citation error from knowingly presenting invented sources as genuine evidence.
Both may require correction, but a formal misconduct determination depends on the applicable definition and circumstances.
Verification should involve locating the original source, confirming its bibliographic identity, and checking whether it actually supports the claim attributed to it.
An AI-generated citation is not evidence merely because it has the familiar appearance of a scholarly reference.
When does AI use become plagiarism?
Plagiarism involves appropriating another person's ideas, processes, results, or words without appropriate acknowledgment under the relevant standards.
AI-generated content may reproduce or closely resemble existing material. Researchers who incorporate such content without checking its provenance may create attribution problems.
However, AI-generated text is not automatically plagiarized simply because it was produced by a model trained on existing works.
The relevant assessment concerns whether protected or attributable material has been improperly appropriated and how the work is represented.
COPE and ICMJE emphasize that researchers remain responsible for appropriate attribution and the originality of submitted material, including content produced with AI assistance.
Researchers should not assume that AI-generated paraphrasing removes the obligation to credit an underlying source.
Is undisclosed AI assistance automatically research misconduct?
No. The consequences depend on what assistance occurred, whether disclosure was required, and how the omission affected the representation of the research.
Some journals require disclosure of AI assistance used in manuscript preparation, data analysis, or figure generation. Other institutions may distinguish routine language correction from substantive AI involvement.
The ICMJE recommends that authors disclose AI-assisted technologies used in producing submitted work and explain their use. Its guidance recognizes that nondisclosure may require corrective action and may be construed as misconduct in some circumstances.
That does not establish that every omitted AI disclosure satisfies a formal misconduct definition.
The more specific question of whether undisclosed AI use constitutes research misconduct requires examining the applicable policy, the nature of the assistance, and the relevant evidence.
Can AI-assisted scientific images create integrity violations?
Yes, particularly when generated or altered images are represented as authentic scientific observations.
For example, generating a microscopy image and presenting it as an actual experimental result would misrepresent the evidence. AI-based alteration of image features may also constitute falsification if it changes the scientific information conveyed.
However, legitimate image processing and AI-assisted visualization are not inherently prohibited.
The distinction depends on the image's purpose, the processing performed, the underlying evidence, and applicable reporting standards.
Researchers should preserve original images and appropriate processing records when scientific images form part of the evidence.
The issue of AI-assisted scientific image falsification is especially consequential because visual evidence can appear persuasive even when its provenance is uncertain.
Can AI create an integrity problem without fabricating or falsifying anything?
Yes. Research integrity extends beyond the most serious categories of misconduct.
AI use may create problems when researchers cannot explain analytical decisions, fail to verify important outputs, breach confidentiality obligations, or present speculative interpretations with unjustified certainty.
For example, an AI-generated thematic analysis may produce plausible categories that do not adequately reflect the original interview transcripts.
If the researcher reports those categories without examining the underlying data, the study may contain serious methodological weaknesses even when there is no evidence of intentional fabrication.
Similarly, uploading confidential participant records or unpublished peer-review manuscripts to an unauthorized service may violate privacy or confidentiality requirements independently of the accuracy of the AI output.
The ICMJE specifically cautions reviewers and editors against uploading confidential manuscripts to AI systems where confidentiality cannot be assured without appropriate permission.
Does the researcher's intention determine whether misconduct occurred?
Intention and culpability matter to formal misconduct determinations, but their legal significance depends on the applicable framework.
Under the United States federal research-misconduct framework, a finding requires more than identifying an incorrect result. The relevant standards address whether conduct was intentional, knowing, or reckless, among other requirements.
A researcher who discovers an AI-generated error and corrects it transparently is in a different position from someone who knowingly submits invented results.
Nevertheless, claiming that AI produced the information does not automatically remove human responsibility.
Researchers remain accountable for how AI-generated material is incorporated into research and what claims are made from it.
The relationship between human accountability and prohibited AI-assisted research actions becomes particularly important when automated systems perform substantial parts of a research workflow.
How should researchers respond when an AI-related integrity problem is discovered?
The appropriate response depends on the seriousness of the issue and the stage of the research.
An error identified during drafting may be corrected through ordinary verification and revision. A problem affecting submitted or published results may require notification of collaborators, institutional officials, or journal editors.
Researchers should preserve relevant records, including original data, analytical code, material AI interactions, and documentation of corrective actions.
They should avoid silently replacing evidence in a way that obscures what occurred.
Where fabrication, falsification, plagiarism, or another serious violation is suspected, the institution's established research-integrity procedures should be followed.
Researchers should distinguish investigating a concern from concluding that misconduct has occurred. Formal findings require appropriate processes and evidence.
