03 · What You Need to Know
What AI Can and Cannot Establish About Research Misconduct
What Counts as Research Misconduct?
Research misconduct is defined differently across legal and institutional frameworks. Under the United States Public Health Service regulations at 42 CFR Part 93, research misconduct encompasses fabrication, falsification, and plagiarism in proposing, performing, reviewing, or reporting research results, subject to the regulation's specific definitions and exclusions.
Fabrication involves making up data or results and recording or reporting them. Falsification involves manipulating research materials, equipment, processes, data, or results so that the research is not accurately represented. Plagiarism involves appropriating another person's ideas, processes, results, or words without appropriate credit.
The applicable U.S. framework excludes honest error and differences of opinion from its definition of research misconduct. It also establishes requirements concerning the evidentiary and culpability elements of a misconduct finding.
Other jurisdictions and institutions may use broader concepts of research integrity violations, including additional questionable research practices.
Researchers should therefore distinguish a formal misconduct finding from an error, an editorial policy violation, or a concern requiring further examination.
Detection Is Not the Same as Determining Misconduct
AI detection systems generally identify patterns that may be associated with problematic research. They do not directly observe every aspect of how a study was conducted.
Detection signal
An observed pattern or anomaly that may justify closer examination, such as duplicated image regions or unusual statistical distributions.
Evidence of a problem
Verified information showing that a particular record, procedure, statement, or research output is incorrect, duplicated, manipulated, or otherwise problematic.
Finding of misconduct
A determination made through the applicable process that the relevant definition and evidentiary requirements for research misconduct have been satisfied.
These stages should not be collapsed into one another.
An image similarity system may correctly identify duplicated regions without establishing whether the duplication was accidental, permitted, or deliberately misleading.
Similarly, an AI system may detect an unusual numerical pattern without establishing that the data were fabricated.
The distinction between AI-generated suspicion and evidence of research misconduct is fundamental to fair and defensible integrity assessment.
What Types of Research Integrity Concerns Can AI Help Identify?
Different tools support different detection tasks. Their capabilities should be evaluated separately rather than grouped under a general claim that AI can detect scientific fraud.
| Potential concern |
Possible AI-assisted function |
Important limitation |
| Fabricated data |
Identify unusual numerical patterns or inconsistencies |
Statistical anomalies do not establish fabrication |
| Falsified images |
Identify duplicated regions, unexpected similarities, or possible alterations |
Image processing artifacts and legitimate reuse may resemble manipulation |
| Plagiarism |
Locate textual similarity and possible unattributed reuse |
Similarity does not independently establish plagiarism |
| Fabricated references |
Check citation metadata and identify unresolved bibliographic records |
Unmatched records may reflect incomplete metadata or indexing limitations |
| Paper-mill manuscripts |
Identify repeated templates, unusual submission patterns, or linked anomalies |
Individual indicators may have legitimate explanations |
| Manipulated peer review |
Flag unusual reviewer identities, contact information, or review patterns |
Behavioral anomalies do not establish that a review is fraudulent |
Some of these tasks can be supported by deterministic software, conventional statistical methods, machine learning, or generative AI. These technologies should not be treated as interchangeable.
Researchers should evaluate the actual detection method rather than assume that adding AI improves reliability.
Can AI Detect Fabrication in Research Data?
Fabricated datasets may contain patterns that differ from those expected under particular statistical models.
For example, investigators may examine repeated values, implausible distributions, unusual digit frequencies, or inconsistencies among reported summary statistics.
AI may help screen large datasets for such anomalies, but genuine data can also exhibit unusual patterns because of sampling, rounding, measurement procedures, or complex underlying distributions.
Conversely, fabricated data may be constructed to resemble expected statistical patterns.
Therefore, an anomaly detector's ability to identify unusual data does not establish its ability to identify fabrication reliably.
The specific challenges of detecting fabricated research data require attention to the data-generating process and independent source evidence.
Can AI Detect Falsified Scientific Images?
Image analysis tools may identify duplicated regions, unusual similarities, or patterns suggesting possible manipulation.
Such screening may be useful for microscopy images, Western blots, and other scientific figures where visual integrity matters.
However, image duplication can arise from legitimate reuse, figure preparation mistakes, or intentional manipulation. Image-processing procedures may also introduce artifacts that resemble suspicious alterations.
A flagged image should be examined alongside original image files, experimental records, figure legends, and the applicable image-handling requirements.
The question of AI detection of falsified scientific images should therefore be separated from the determination of researcher intent.
Text Similarity Detection Is Not a Plagiarism Verdict
Text-matching systems can identify passages that resemble previously available material.
They may help reviewers investigate unattributed copying, inappropriate text recycling, or unusual similarities among manuscripts.
However, similarity can arise from properly quoted material, standard methodological terminology, common technical expressions, or legitimate descriptions of established procedures.
A low similarity score also does not guarantee the absence of plagiarism. Appropriated ideas, paraphrased material, or copied content outside the comparison database may escape detection.
Researchers should examine the source, context, attribution, and applicable publication policies.
The limitations of AI-assisted plagiarism detection are especially important when screening results could affect a researcher's reputation or career.
AI-Generated Text Detection Is a Different Question
Some systems attempt to estimate whether a manuscript was written using generative AI.
Even if a detector identifies stylistic features associated with AI-generated text, that result does not establish fabrication, falsification, or plagiarism.
AI assistance may be permitted under applicable journal policies, particularly for certain language or editorial tasks, subject to disclosure and other requirements.
Conversely, research misconduct can occur in a manuscript written entirely by humans.
Researchers should avoid conflating AI-generated language with scientific fraud.
Detection performance can vary with language, genre, editing, and the systems used to generate or evaluate text. A detector should not be treated as definitive evidence of authorship or misconduct without adequate validation.
Why False Positives Are Particularly Consequential
A false positive occurs when a detection system flags material as suspicious even though the targeted problem is absent under the applicable reference standard.
In research integrity, false positives can have serious consequences if screening results are interpreted as accusations.
A researcher may face reputational damage, delayed publication, unnecessary investigation, or professional uncertainty.
The possibility of false accusations makes specificity and positive predictive value important, particularly when the underlying prevalence of confirmed misconduct is low.
Researchers should distinguish the rate of incorrect flags from the probability that a flagged case actually involves misconduct.
The risks associated with AI incorrectly labeling honest research as suspicious justify cautious communication and proportionate follow-up.
Why Positive Predictive Value Matters
A tool can have apparently strong sensitivity and specificity while producing many false alarms when the condition being detected is uncommon.
Consider a hypothetical screening system evaluated in a population where 1% of manuscripts contain the specific problem being targeted.
In this hypothetical setting, only about 15% of positive flags would correspond to the targeted condition, despite the apparently favorable sensitivity and specificity.
The calculation assumes that the performance estimates apply to the screened population and that the target condition has a valid reference standard. It does not represent the measured performance of any real research misconduct detector.
Formal misconduct is particularly difficult to use as a reference label because establishing it may require investigation of evidence and culpability that an automated system cannot observe.
AI Detectors May Perform Differently Across Disciplines
Research practices vary across fields. Data distributions, manuscript structures, image conventions, citation practices, and statistical methods may differ substantially.
A system trained on biomedical articles may not perform similarly when applied to mathematics, qualitative education research, or humanities scholarship.
For example, repeated technical phrases may be ordinary in standardized laboratory methods but unusual in another genre.
Likewise, statistical anomaly thresholds developed for one measurement process may be inappropriate for another.
Researchers should require evidence that a detection tool performs adequately in the intended disciplinary and publication context.
General benchmark accuracy should not be interpreted as proof of reliable detection in every field.
AI Systems May Miss Sophisticated or Previously Unseen Misconduct
Detection methods depend on the patterns they are designed or trained to recognize.
Fabricated data may be statistically plausible. Manipulated images may evade particular similarity checks. Plagiarized ideas may be expressed in substantially different language.
Organized publication fraud may also adapt to known screening procedures.
Consequently, a manuscript that receives no automated flags is not necessarily free from integrity concerns.
Researchers should avoid treating a negative screening result as a certification of research integrity.
Effective integrity assessment may require source records, methodological evaluation, editorial scrutiny, and other evidence unavailable to an AI detector.
Research Misconduct Findings Require Appropriate Procedures
Formal research misconduct processes differ among institutions and jurisdictions.
Under the applicable U.S. Public Health Service framework, allegations and findings are governed by defined procedures and evidentiary requirements. Other systems may establish different investigative authorities, standards, and protections.
COPE provides guidance to editors addressing concerns about submitted or published work, but journal editors do not necessarily possess the authority or evidence needed to make formal institutional misconduct findings.
Where an AI system identifies a concern, the response should follow the relevant procedure rather than bypass it.
The question of whether an AI misconduct flag should trigger an investigation depends on the credibility and specificity of the evidence, applicable policies, and the appropriate threshold for escalation.
Human Review Must Examine Primary Evidence
Human oversight is meaningful only when reviewers can examine evidence relevant to the suspected problem.
For a possible image manipulation, this may involve original image files and experimental records. For suspicious numerical data, it may require raw observations, data-processing logs, and measurement documentation.
For possible plagiarism, reviewers should examine the original source and determine whether attribution is appropriate.
The reviewer should also consider plausible innocent explanations and whether the detection method is suitable for the material being assessed.
The requirements for human review following an AI research-integrity flag should be proportionate to the seriousness and evidentiary strength of the concern.
A human reviewer who merely accepts the AI system's classification has not independently established misconduct.
Confidentiality and Procedural Fairness Matter
Research integrity concerns can affect individuals and institutions before any allegation has been substantiated.
Screening results should therefore be communicated carefully and shared only with individuals who have a legitimate role in evaluating them.
Researchers and editors should avoid public accusations based solely on automated flags.
Where formal procedures apply, individuals should receive the protections established by those procedures, which may include confidentiality, impartial assessment, opportunities to respond, and appropriate documentation.
AI systems should not be used to bypass institutional safeguards or transform uncertain signals into public judgments about individual researchers.
How Should Institutions Evaluate an AI Misconduct Detection Tool?
Before deployment, institutions should define the specific problem the tool is intended to identify.
They should then examine validation evidence, relevant false-positive and false-negative rates, performance across research contexts, and the consequences of acting on its outputs.
Evaluation should consider whether the tool produces interpretable evidence that reviewers can independently examine.
Institutions should also establish how flags are recorded, who may access them, when they are escalated, and how incorrect classifications are corrected.
Model updates or changes in screening populations may require reassessment.
The National Institute of Standards and Technology's AI Risk Management Framework provides a general structure for identifying, measuring, and managing AI-related risks, although it is not a specialized research misconduct adjudication standard.
AI Should Support Evidence Gathering Rather Than Replace Judgment
AI may be particularly useful for directing attention toward records or inconsistencies that warrant closer examination.
For example, a system may identify a reference whose bibliographic details cannot be verified or highlight duplicated regions in a scientific figure.
These observations can guide targeted checks.
The appropriate conclusion should follow from independently verified evidence, not the detector's confidence score or an automatically generated allegation.
Watch Out
Do not label a researcher, manuscript, or dataset fraudulent solely because an AI detector assigns a high suspicion score. A screening signal is not a formal finding, and unsupported accusations can cause substantial harm.
