03 · What You Need to Know
Why Automated Suspicion and Research Misconduct Evidence Must Remain Separate
What Is the Difference Between Suspicion and Evidence?
Suspicion is a reason to consider whether a particular problem may exist.
Evidence is information relevant to evaluating whether a factual claim is supported.
Evidence may be incomplete, ambiguous, or insufficient. Its existence does not automatically establish the conclusion for which it is offered.
For example, an AI system may identify matching regions in two scientific images. The observed similarity may constitute evidence of duplication if independently verified.
However, the similarity does not independently establish that the duplication was improper or that a researcher intentionally falsified the images.
The central distinction is between observing a pattern and establishing the conduct that allegedly produced it.
What Are the Different Levels of an Integrity Concern?
Research-integrity assessment should distinguish several levels of evidentiary development.
| Level |
What is established? |
What remains unresolved? |
| Automated suspicion |
A system has identified a pattern meeting its screening criteria |
Whether the pattern is valid, relevant, or problematic |
| Verified anomaly |
The unusual pattern or discrepancy can be independently reproduced |
Whether it has a legitimate explanation |
| Verified research problem |
A specific error, misrepresentation, or other problem has been established |
Its cause, responsibility, and potential misconduct status |
| Credible misconduct allegation |
Available information may satisfy the threshold for further inquiry |
Whether misconduct can be substantiated |
| Formal misconduct finding |
The applicable authority has determined that the required criteria are met |
Any separate corrective or administrative consequences |
The exact procedural terminology varies across jurisdictions and institutions.
These distinctions help prevent an automated screening result from being treated as though an investigation had already established misconduct.
Research Misconduct Has Specific Definitions
Under the United States Public Health Service regulations at 42 CFR Part 93, research misconduct encompasses fabrication, falsification, and plagiarism as defined by that framework.
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 framework excludes honest error and differences of opinion from research misconduct.
Other institutions and jurisdictions may establish additional research-integrity requirements or different procedures.
An AI system cannot determine which formal standard applies merely by recognizing suspicious features in a manuscript.
Why Does Intent Matter?
Some formal misconduct frameworks require evidence concerning the researcher's culpability.
Under the applicable United States Public Health Service framework, a finding of research misconduct requires that the conduct be committed intentionally, knowingly, or recklessly, along with other specified elements.
An AI detector generally cannot establish a person's mental state from statistical patterns, image similarities, or writing style.
For example, two duplicated figure panels may result from accidental assembly, unauthorized manipulation, or another cause.
The visible duplication may be identical in each situation.
Determining responsibility and culpability requires evidence about the relevant actions and circumstances.
Researchers should not infer intent solely from how suspicious a result appears.
A Mathematical Impossibility Does Not Automatically Establish Fabrication
Some statistical discrepancies can be demonstrated with certainty under specified assumptions.
Suppose a study reports a mean of 3.47 for ten equally weighted integer-valued responses on a five-point scale.
Because the sum of ten integers must be an integer, the mean must be a multiple of 0.1. A value of 3.47 is incompatible with the stated calculation.
That conclusion is mathematically defensible if the assumptions are correct.
However, the discrepancy could reflect an incorrect sample size, an adjusted mean, a composite score, a reporting error, or fabricated observations.
The mathematical finding establishes a problem requiring explanation, not its cause.
The principles governing statistical anomalies associated with possible data fabrication illustrate why forensic results require contextual interpretation.
Image Similarity Is Evidence About Pixels, Not Automatically About Conduct
Computer vision systems can identify duplicated or closely matching image regions.
When independently verified, such findings may provide strong evidence that particular visual material has been reused.
However, reuse may be legitimate, accidentally introduced, or deliberately misleading.
Researchers should examine original acquisition files, figure legends, experimental records, and relevant processing procedures.
A finding of duplication should not be automatically transformed into an accusation of falsification.
The distinction between the observed image relationship and its cause remains essential.
Textual Similarity Is Not Proof of Plagiarism
A similarity detector identifies overlap between a manuscript and available source material.
That overlap may reflect copied wording, standard terminology, properly attributed quotation, or legitimate methodological description.
Conversely, plagiarism may involve appropriated ideas that produce little direct textual similarity.
Researchers should compare the actual source and manuscript passages and evaluate attribution under applicable standards.
A similarity percentage does not directly quantify plagiarism.
Nor does a low similarity score establish originality.
The evidentiary question concerns the relationship between the source material and the author's representation of it.
AI Confidence Scores Can Be Misleading
Automated systems may express suspicion through numerical scores or percentages.
However, a score is not necessarily a calibrated probability that misconduct occurred.
For a probabilistic interpretation to be defensible, the detector would need appropriate validation against a suitable reference standard and relevant deployment conditions.
Research misconduct presents a difficult validation problem because confirmed cases may be heterogeneous and may require evidence unavailable to automated systems.
A model's apparent precision should not be confused with evidentiary certainty.
Why Prevalence Changes the Meaning of a Positive Flag
The probability that a positive screening result identifies the target condition depends partly on how common that condition is in the screened population.
This relationship is expressed through positive predictive value.
The calculation is illustrative and does not estimate actual misconduct prevalence or the performance of a particular detection system.
Its significance is conceptual: a positive classification should not be interpreted without considering the detector's error rates and the population in which it operates.
Multiple AI Systems Can Share the Same Error
Agreement among several AI systems may appear to strengthen a misconduct allegation.
However, models may rely on similar training material, statistical assumptions, or detection features.
Two systems may therefore reproduce the same mistaken interpretation.
For example, several language-based detectors may flag formulaic academic writing without establishing that the manuscript was produced fraudulently.
Independent corroboration should involve evidence capable of testing the allegation, not merely another automated classification.
Model consensus is not a substitute for primary evidence.
Can AI-Generated Suspicion Become Relevant Evidence?
Yes, but the distinction must be precise.
An AI output may document that a screening system identified a particular feature at a particular time.
That record may be relevant to understanding why an assessment began.
If the flagged feature is independently verified, the underlying observation may become evidence relevant to a specific integrity concern.
However, the AI system's broader accusation remains an interpretation that must be justified separately.
For example, an automated image match may lead reviewers to original files establishing that the same image was presented as two different experiments.
The verified source evidence may be consequential. The original suspicion score does not independently establish misconduct.
Primary Evidence Is Often More Informative Than Detection Scores
Primary evidence may include original datasets, laboratory records, image acquisition files, survey administration documentation, manuscript histories, or other records relevant to the alleged conduct.
These materials can help reconstruct what occurred during the research process.
However, their evidentiary value depends on authenticity, completeness, provenance, and context.
Researchers should not assume that every original file is automatically conclusive.
Records may contain errors, be incomplete, or require specialized interpretation.
The objective is to establish a defensible factual account using relevant evidence rather than rely on a detector's classification.
Absence of Evidence Is Not Automatically Evidence of Fabrication
A research team may be unable to produce particular source records.
This can limit verification and may raise questions about compliance with data-management or retention requirements.
However, unavailable records do not automatically establish that the reported data were invented.
Records may be missing because of technical failures, institutional transitions, privacy restrictions, or other circumstances.
Researchers should examine whether the records were required to be retained and what alternative evidence remains available.
Any conclusion should reflect the actual evidentiary limitations rather than presume misconduct from missing information alone.
Scientific Unreliability and Research Misconduct Are Different Findings
A research article may contain serious errors that undermine its conclusions without a formal finding of misconduct.
For example, an incorrect statistical analysis may make the reported results unreliable even when the error was unintentional.
Conversely, misconduct may be established in relation to a particular research procedure without automatically resolving every question about the validity of other findings.
Editors may need to correct or retract unreliable publications independently of whether misconduct has been formally established.
Researchers should distinguish the reliability of the scholarly record from the culpability of individuals.
Both are important, but they involve different questions and sometimes different decision-making authorities.
Why Does Procedural Fairness Matter?
Research misconduct allegations can affect professional reputation, employment, funding, and scholarly standing.
Consequential decisions should therefore follow applicable procedures that support impartial evaluation and protect relevant rights.
Depending on the jurisdiction and institution, these may include confidentiality requirements, evidence preservation, opportunities to respond, and designated investigative responsibilities.
AI-generated suspicion should not bypass these safeguards.
Procedural fairness also supports the quality of the investigation by allowing alternative explanations and relevant evidence to be examined.
Who Has Authority to Determine Research Misconduct?
The appropriate authority depends on the applicable institutional, regulatory, and publication framework.
Journal editors may assess manuscript reliability, request clarification, or refer serious concerns to institutions.
Research-integrity offices or other designated bodies may conduct formal investigations under their governing procedures.
An AI vendor or automated detector does not acquire investigative authority simply because its system produces a misconduct classification.
The distinction addressed in whether AI flags should trigger investigations is particularly important when deciding what action is justified.
What Should Happen Before an Allegation Is Communicated?
Reviewers should identify the specific concern and determine whether its underlying evidence has been verified.
They should distinguish preliminary observations from formal allegations.
Communications should accurately describe what is known and what remains uncertain.
For example, stating that two figures appear to contain duplicated regions is more precise than asserting that the authors falsified their results.
Confidentiality should be maintained according to applicable procedures.
Premature public accusations may cause harm even when subsequent assessment resolves the concern.
How Can Researchers Avoid Evidence Laundering?
Evidence laundering can occur when an unsupported claim gains apparent credibility through repetition or presentation as an established finding.
In research-integrity screening, an AI-generated suspicion may be copied into an editorial report, summarized in an institutional memorandum, and later cited as though the underlying allegation had already been verified.
The repetition does not create new evidence.
Each subsequent report should preserve the original evidentiary status of the concern.
Reviewers should distinguish what the AI claimed, what was independently checked, and what the available records actually establish.
Otherwise, an initial classification error may acquire unwarranted institutional authority.
Why Documentation Must Separate Observations From Interpretations
A defensible research-integrity record should identify observed facts, analytical interpretations, unresolved questions, and procedural decisions separately.
For example, a record might state that a statistical inconsistency was independently reproduced but that its cause remains undetermined.
That statement is more informative than a general conclusion that the dataset appears fraudulent.
Where AI contributed, the system's output should be recorded as part of the assessment history rather than presented as an independently established fact.
The requirements for human review after an AI research-integrity flag include documenting the reasoning that connects evidence to a decision.
Can a Strong AI Flag Justify Immediate Action?
A serious, specific concern may justify prompt protective or editorial action under applicable procedures.
For example, an editor may temporarily pause manuscript processing while verifying a credible discrepancy.
However, precautionary action is not equivalent to a final misconduct finding.
The scope and duration of the response should be proportionate to the concern and consistent with the responsible authority's procedures.
Researchers should distinguish temporary measures intended to preserve evidence or prevent harm from sanctions based on established misconduct.
Why the Burden of Proof Cannot Be Replaced by Algorithmic Confidence
Formal misconduct frameworks may specify who bears the burden of proof and what standard must be satisfied.
Under the United States Public Health Service framework, the relevant procedures establish evidentiary requirements for findings of research misconduct.
A model's confidence score does not alter those requirements.
Nor can an institution assume that a researcher must prove innocence simply because an AI detector produced a suspicious classification.
Investigative responsibilities should follow the applicable framework, not an automated presumption.
Where different jurisdictions or institutional policies apply, their specific standards should be verified rather than generalized.
How Should Unresolved Concerns Be Reported?
Not every assessment produces a definitive conclusion.
Sometimes the available evidence establishes a discrepancy but does not identify its cause.
In other cases, original records may be unavailable or competing explanations may remain plausible.
Researchers should report uncertainty accurately and avoid treating inconclusive evidence as proof.
An unresolved concern may still justify editorial caution or additional verification, depending on the circumstances.
However, the language used should distinguish uncertainty about research reliability from a substantiated allegation against an individual.
What Does Responsible AI Research-Integrity Governance Require?
Institutions should define the permitted role of AI screening and the evidentiary status of its outputs.
They should evaluate detection performance, establish human verification procedures, and assign authority for escalation and formal determinations.
Screening systems should be assessed for false positives, false negatives, and potential differences in performance across research contexts.
Records should preserve the distinction between automated suspicion and verified findings.
The NIST AI Risk Management Framework provides general guidance for evaluating and managing AI-related risks, while specialized research-integrity procedures determine how allegations should be handled.
Responsible governance requires both technical scrutiny and procedural accountability.
Watch Out
Do not convert an AI-generated suspicion score into an allegation of research misconduct by repeating it in an editorial report or institutional record. A claim does not become evidence merely because it has been restated by several people or systems.
