03 · What You Need to Know
Why AI-Detection Scores Are Problematic as Misconduct Evidence
What Does an AI-Detector Score Actually Establish?
An AI-text detector analyzes characteristics of submitted writing and estimates whether they resemble patterns associated with machine-generated text.
Depending on the system, the classification may involve statistical predictability, machine-learning models, or other linguistic features. Most general-purpose detectors do not directly observe the researcher's writing process.
Consequently, a detector score establishes that a particular system produced a particular classification under its operating conditions. It does not independently establish the historical origin of the manuscript.
AI-detection result
A computational classification based on the submitted text and the detector's methods.
Misconduct finding
An institutional determination that a person violated a defined academic or research-integrity requirement, reached through the applicable evidentiary and procedural standards.
These are different kinds of conclusions. The first concerns a technical assessment. The second concerns conduct, responsibility, and compliance with institutional rules.
Understanding why an AI-detection score is not proof is essential before assigning it evidentiary weight.
Why Must Universities Distinguish AI Use From Misconduct?
The use of generative AI does not automatically constitute academic or research misconduct.
Universities may permit some forms of AI assistance while restricting others. A policy might allow grammar correction but require disclosure of AI-generated drafting, prohibit undisclosed AI use in a particular assessment, or regulate the processing of confidential research data.
The requirements may differ across undergraduate coursework, postgraduate dissertations, funded research, and journal submissions.
Even if AI assistance were independently confirmed, an institution would still need to determine whether the activity violated the rules applicable at the relevant time.
Research misconduct also has a more specific meaning in some regulatory frameworks. For example, the United States Office of Research Integrity defines research misconduct in Public Health Service-supported research as fabrication, falsification, or plagiarism, subject to its regulatory requirements. Undisclosed AI use is not automatically equivalent to one of these categories.
Institutions may separately regulate unauthorized AI assistance as an academic integrity violation. Such distinctions should be explicit rather than collapsed into a single allegation of misconduct.
What Does Research Say About Detector Reliability?
Published evaluations have identified substantial limitations in AI-text detection, including false positives, false negatives, and sensitivity to changes in writing style.
Weber-Wulff and colleagues evaluated 14 detection tools in a 2023 study published in the International Journal for Educational Integrity. Their findings raised concerns about the accuracy and reliability of the tested systems for identifying AI-generated writing.
Liang and colleagues, writing in Patterns, found that seven evaluated GPT detectors disproportionately misclassified essays written by non-native English speakers. Across the tested detectors, the average false-positive rate for their 91 TOEFL essays was 61.3%.
These findings do not establish the error rate of every contemporary commercial detector. They demonstrate why institutional use requires independent validation rather than reliance on general accuracy claims.
Because AI detectors can incorrectly flag human-written material, a positive classification cannot safely be treated as a verified account of how a student or researcher produced a document.
What Do AI-Detection Providers Say About Disciplinary Use?
Turnitin's official guidance acknowledges that its AI writing detection model may misidentify human-written and AI-generated text. The provider states that its results should not be used as the sole basis for adverse action against a student.
This limitation is significant because a detector's numerical presentation can appear more definitive than the underlying evidence supports.
A university should consult the current documentation for the specific product and version it uses, including score definitions, supported languages, minimum text lengths, and known limitations.
Provider guidance is not a substitute for independent validation, however. A commercial performance claim does not automatically establish that the tool is suitable for high-stakes institutional decisions.
Can a Detection Score Be Considered Evidence at All?
The word evidence requires careful interpretation.
In a broad sense, a detector report is evidence that a particular software system classified the submitted text in a particular way. It may provide information relevant to deciding whether additional examination is warranted.
That does not mean it provides strong evidence that AI was actually used.
Its evidentiary value depends on whether the system has been validated for the relevant task, whether its error rates are known, and whether the report can be interpreted correctly.
A detector that has not been evaluated on comparable academic writing may provide little meaningful support for an authorship allegation.
Universities should therefore distinguish between an initial screening signal, corroborating information, and evidence sufficient to support a formal finding.
| Possible Use |
Appropriate Interpretation |
Important Limitation |
| Preliminary screening |
A possible reason for proportionate further examination |
Only justified where tool performance and screening procedures support such use |
| Request for clarification |
One contextual consideration alongside independent concerns |
A score alone should not create a presumption of wrongdoing |
| Formal misconduct investigation |
Limited technical information requiring critical evaluation |
Cannot independently establish prohibited conduct |
| Disciplinary finding |
Insufficient as standalone proof |
Requires applicable standards, corroboration, and procedural fairness |
Some institutions may reasonably decide not to use AI detectors at all in misconduct procedures because their limitations outweigh their practical value.
Why Does the Burden of Proof Matter?
When a university alleges misconduct, it should follow the burden and standard of proof established by its applicable regulations and procedures.
These standards vary by jurisdiction, institution, and type of proceeding. A university disciplinary process may use different rules from a formal research misconduct investigation governed by national regulations.
A detector score should not silently reverse the evidentiary burden by requiring a researcher to prove innocence simply because software generated an unfavorable classification.
For example, a researcher may have no complete keystroke history or may have drafted sections using software that does not preserve earlier versions. The absence of such records does not automatically establish AI use.
Supporting records can help clarify authorship, but they should be evaluated proportionately and according to the requirements that actually applied.
Why Are False Positives Particularly Serious in Disciplinary Proceedings?
A false positive occurs when human-written text is incorrectly classified as AI-generated.
In routine experimentation, such an error may simply reduce a detector's measured accuracy. In a disciplinary setting, it can initiate questioning, create anxiety, delay academic progress, or contribute to an unjustified allegation.
The consequences may be especially significant for doctoral candidates, early-career researchers, and scholars whose professional standing depends on institutional trust.
The severity of these potential consequences means that the acceptable evidentiary standard should be substantially more demanding than what might be sufficient for an exploratory technical demonstration.
A high overall accuracy figure is not enough. Institutions should examine false-positive rates, the proportion of flagged cases that are genuinely positive, and whether errors differ across relevant writing populations.
Could AI-Detection Policies Disadvantage Researchers Writing in an Additional Language?
Yes, this is a documented concern for some detectors.
Liang and colleagues found substantial disparities in the classification of essays written by native and non-native English speakers. Their findings suggest that linguistic predictability may contribute to erroneous AI classifications.
Although their results cannot be generalized to every detector or research manuscript, they demonstrate why institutions should examine subgroup error rates before deploying these systems.
A university that applies identical software to all researchers is not necessarily ensuring equal treatment if the software performs differently across populations.
The possibility of language-related AI-detection bias is particularly relevant in internationally diverse universities.
What Evidence Should Universities Examine Instead?
When a specific concern about undisclosed or prohibited AI use warrants investigation, evidence connected to the actual writing process may be more informative than statistical text classification.
Depending on the allegation and available records, relevant evidence may include manuscript drafts, revision histories, research notes, correspondence, documented AI interactions, and explanations from the researcher.
Such materials should be evaluated for authenticity, completeness, and relevance. A version history does not automatically establish that every sentence was written without AI assistance, and missing records do not automatically establish wrongdoing.
Investigators should also identify the specific conduct alleged and explain how the evidence relates to the applicable rule.
The evaluation of evidence beyond an AI-detector score should form part of a proportionate and fair inquiry.
What Procedural Safeguards Should Universities Establish?
Universities considering AI detection for integrity-related purposes should establish safeguards before deploying the technology.
These may include documented validation requirements, restrictions on the use of automated scores, clear thresholds for initiating an inquiry, and procedures for evaluating independent evidence.
Researchers should be informed of the specific concern, the applicable policy, and the evidence being considered. They should have a meaningful opportunity to respond and access to the review or appeal procedures provided by institutional rules.
Institutions should also address confidentiality, data protection, conflicts of interest, and the appropriate retention of detection reports.
These safeguards are particularly important because human reviewers can be influenced by automated classifications. Merely adding a human decision-maker does not guarantee that an unsupported score will receive appropriate scrutiny.
