On August 14, the federal Office of Research Integrity announced guidance on generative AI in research-misconduct proceedings. It recommended that researchers disclose how they used AI and verify their results. It also warned institutions against relying solely on automated tools to establish misconduct.
The guidance addresses responsibility on both sides of a research finding. The researcher must account for the work being reported. The institution must account for the evidence supporting any finding that the work involved misconduct.
AI can assist research. Automated tools can also assist scrutiny of that research. Neither contribution removes the need for human judgment.
That makes the guidance useful beyond the question of whether a researcher used a chatbot. It shows what people must examine before they rely on an output, and what they must establish before they accuse someone of misconduct.
The Regulation Supplies the Obligations
ORI oversees the Public Health Service research-misconduct framework under 42 CFR Part 93. The applicability provisions cover institutions applying for or receiving PHS support for biomedical or behavioral research, research training, and related activities.
The framework reaches covered funding proposals, research activities, and research records. A proposal can fall within its scope even if it never receives an award. This is broader than preparing a submission to a federal public health agency, and more specific than a general rule for every academic assignment or publication.
The August guidance is nonbinding. It does not bind the public, ORI, or HHS, and following it does not guarantee compliance. The regulation prevails if the two conflict.
The distinction matters. ORI is explaining how institutions can apply existing obligations when generative AI is involved. It has not created a comprehensive AI research-governance mandate.
Disclosure Makes the Method Examinable
The guidance recommends identifying the AI tools used in research, manuscripts, and grant preparation and explaining how they were used. It also recommends verifying all results, including those produced through processes involving generative AI.
Those are connected tasks. Saying that AI was used does not tell a reader whether it helped draft prose, processed an image, generated data, or contributed to an analysis. The role of the tool affects what needs to be checked.
Disclosure can help another researcher understand the method and reproduce the work. Verification addresses whether the result being reported is supported. A statement acknowledging AI use cannot perform that verification.
ORI also explains that disclosure does not prevent a misconduct finding. A disclosed practice can still depart significantly from accepted research practices. Undisclosed or inaccurately described AI processing can be evidence of fabrication or falsification.
The relevant question is what happened to the research and how accurately the researcher represented it.
An AI Output Is Not a Misconduct Finding
The regulation’s misconduct test requires a significant departure from accepted research-community practices, intentional, knowing, or reckless conduct, and proof by a preponderance of the evidence. The conduct must also fall within the regulatory definition of research misconduct: fabrication, falsification, or plagiarism.
Preponderance means the evidence establishes that the fact at issue is more likely true than not. It is a standard for evaluating the evidence, not a confidence score supplied by a detection product.
ORI says automated tools that detect data manipulation or plagiarism can provide evidence. It recommends against using their outputs as the sole evidence in a misconduct proceeding. It also warns that plagiarism detectors can have limited ability to identify copied material in AI-generated scientific content.
These are tools for identifying and examining concerns. The institution must still determine what conduct occurred, whether it falls within the definition, and whether the required state of mind and evidentiary standard are established.
A flag can justify further examination. It does not supply all those conclusions.
The Research Record Can Show What Actually Happened
ORI points to evidence beyond the manuscript itself. Purchasing records, intermediate image-processing files, and access logs for animal facilities can help establish whether experiments occurred as reported.
Such records can support or weaken an allegation. They can distinguish an experiment from a synthetic result presented as an experiment, or help explain how an image changed during processing.
The guidance also identifies a less obvious possibility: a commonly used tool can alter data without the researcher understanding its processing. It gives the example of automatic image alteration in smartphone cameras.
That possibility is relevant to honest error. It is not an automatic defense. ORI states that a respondent bears the burden of proving honest error, and the regulation requires the finder of fact to give due consideration to credible evidence of it.
Human review therefore involves reconstructing the work, examining the records, and assessing the explanation. Recognizing that a tool altered an image begins that analysis. It does not finish it.
Expertise Determines Which Questions to Ask
Accepted practices can differ across research communities. An investigator needs to understand the relevant scientific work to assess whether a particular use of AI significantly departed from those practices.
The regulation requires appropriate scientific expertise on investigation committees and reasonable steps toward an impartial, unbiased investigation. ORI recommends considering additional AI expertise where it would help explain the tools involved.
Technical knowledge and scientific knowledge serve different purposes. Understanding how an image-processing tool works helps explain the transformation. Understanding the research helps determine whether that transformation changed what the evidence legitimately showed.
Institutions also have obligations to obtain relevant records, pursue significant leads, document the investigation, and provide the required procedural protections. Those requirements give substance to the human assessment.
A Fabricated Reference Can Raise Different Questions
ORI’s treatment of invented citations illustrates why classification matters. Fabrication under the regulation concerns data or results. A reference listed in a manuscript is generally not itself a result of that manuscript’s scientific inquiry.
In a literature review, however, references can constitute research data. Inventing them can therefore support an allegation of fabrication in that context.
An invented reference remains a problem requiring attention. But its existence does not automatically establish the particular federal misconduct category. Institutions must assess the role the reference played and the other elements of the standard. Other institutional or publication rules can also apply.
Copied text raises a separate question. ORI notes that generative AI tools can reproduce text or ideas from existing works and result in plagiarism, and it leaves institutions to determine whether a particular use qualifies. The regulatory definition reaches the appropriation of another person’s work. The question is therefore not whether the researcher credited the tool but whether the output reproduced an identifiable person’s work, which is what the detection limits described above make hard to locate. The institution still carries the burden of proof on each element, so an unidentifiable source weakens the allegation rather than the respondent’s position.
