Skip to main content
📝 Annotated Essay Tutorial

AI Detectors Are Not Evidence: Why Schools Must Stop Treating Them as Proof

Published accuracy studies show that AI detectors misfire unevenly, punishing some students more than others. A score should start a conversation, never end one.

2,150 words APA 7 Undergraduate 8 notes ~9 min read Written November 2026 Facts as of November 18, 2026

PaperDue offers an AI-writing checker, TextChecker. This essay is an example of argumentative writing and does not evaluate any specific product.

Illustration of a school desk with handwritten essay drafts and a closed laptop.

I. A Classroom Episode That Exposed the Problem

In spring 2026, an economics course at Brown University became the center of a highly public dispute over generative AI and academic integrity (Topaz & Bahl, 2026).A1 For the 59 students who finished the course, the average score dropped from 95.7 out of 100 on a take-home midterm to 48.8 on a proctored final, a fall the instructor blamed on unauthorized generative AI use during the midterm (Topaz & Bahl, 2026). Yet a statistical preprint — not yet peer-reviewed — analyzing the publicly released scores found that the correlation between each student's midterm and final was only 0.06, and that permutation tests detected no statistically significant association between the two sets of scores, a pattern consistent with multiple explanations including a harder final, test anxiety, and regression to the mean, not only AI use (Topaz & Bahl, 2026). The episode crystallizes a question that every school now faces: when a number — whether from a detector algorithm or an anomalous grade — appears to suggest AI misuse, how much weight should that number carry on its own?

Schools should not treat an AI-detector score as evidence of misconduct on its own: published accuracy studies show false positives that fall unevenly on some students, so a score should at most prompt a conversation backed by process evidence such as drafts and version history (Karr et al., 2026; Liang et al., 2023; MIT Sloan Teaching & Learning Technologies, n.d.; Vanderbilt University, 2023; Weber-Wulff et al., 2023).

II. What the Accuracy Studies Actually Show

A widely cited independent test of AI-writing detectors tested two commercial systems widely used in academic settings along with 12 publicly available tools, and its authors concluded that the tools are “neither accurate nor reliable,” with a primary bias toward classifying output as human-written rather than flagging AI-generated text (Weber-Wulff et al., 2023).A2 That peer-reviewed study, published in the International Journal for Educational Integrity, also found that content obfuscation techniques — simple steps a student could take to disguise AI-generated prose — significantly worsened the performance of every tool tested (Weber-Wulff et al., 2023). In other words, a student who deliberately disguises AI text is the one most likely to escape a flag. Both peer-reviewed studies cited in this essay were published in 2023 and tested the detectors available at that time (Liang et al., 2023; Weber-Wulff et al., 2023).

A separate preprint — not yet peer-reviewed — studying published English abstracts from 2013 to 2015 versus 2023 to 2025 found that light editing of abstracts, described as a proxy for guideline-compliant AI assistance, was flagged by detectors at rates of 38 to 80%, while after humanization software was applied, fewer than 4% of AI-labeled rewrites remained flagged, giving a false-negative rate above 96% (Karr et al., 2026). That same preprint concluded that editing honestly with AI carries a greater risk of sanction than using a humanizer to evade detection, meaning the students most likely to be caught are precisely those who used AI transparently and within guidelines (Karr et al., 2026). This inversion — where rule-following raises risk and rule-breaking lowers it — is not a minor calibration problem; it is a structural failure that makes detector scores an unreliable basis for any disciplinary action.

Turnitin, one of the most widely deployed systems, says its tool achieves a less than 1% false positive rate, but that figure is Turnitin's own claim about its product (Chechitelli, n.d.). Vanderbilt University noted that even accepting Turnitin's stated rate at face value, applying it to the roughly 75,000 papers Vanderbilt submitted in 2022 would have produced around 750 incorrectly labeled student papers (Vanderbilt University, 2023). A rate that sounds reassuringly small in percentage terms can translate into hundreds of real students facing unwarranted scrutiny, which is why institutional decisions about deployment cannot rest on vendor-supplied accuracy claims alone (Chechitelli, n.d.; Vanderbilt University, 2023).

III. The Unequal Burden on Non-Native English Writers

Beyond overall accuracy, the research reveals a distributional problem: the errors made by AI detectors do not fall evenly across the student population (Liang et al., 2023). Researchers at Stanford University tested seven popular GPT detectors on 91 essays that non-native English speakers wrote for the TOEFL and found an average false-positive rate of 61.3%, while the same detectors classified essays by US eighth-grade students accurately (Liang et al., 2023).A3 All seven detectors agreed that 19.8% of the TOEFL essays, though written by humans, were AI-authored, and 97.8% of them were flagged by at least one detector (Liang et al., 2023).

The mechanism behind this bias is the statistical measure most detectors rely on: text perplexity, which gauges how easily a language model can predict the next word in a sentence (Liang et al., 2023). Non-native writers tend to use a more limited range of vocabulary and grammatical structures, which produces lower text perplexity — the same signal that detectors associate with AI-generated prose — even when every word is the student's own (Liang et al., 2023). When the Stanford researchers used ChatGPT to enrich the vocabulary of the TOEFL essays, the average false-positive rate dropped by 49.7 percentage points, from 61.3% to 11.6%, confirming that the detectors were responding to linguistic simplicity, not to AI authorship (Liang et al., 2023).

The Stanford authors also found that the same vocabulary-enrichment intervention that helped non-native writers evade false flags also allowed genuinely AI-generated college essays to evade detection, with detection rates dropping to near zero after a simple self-edit prompt (Liang et al., 2023). The practical consequence is a system that is simultaneously over-inclusive for students who write plainly in a second language and under-inclusive for students who know how to manipulate an AI's output. Deploying such a system as a disciplinary instrument does not neutrally enforce academic integrity; it systematically disadvantages students who are already navigating higher education with fewer linguistic resources (Liang et al., 2023; Vanderbilt University, 2023).

IV. What Institutions Have Concluded and What They Advise Instead

Facing these findings, some universities have moved away from automated detection entirely (MIT Sloan Teaching & Learning Technologies, n.d.; Vanderbilt University, 2023). Vanderbilt University disabled Turnitin's AI detection tool in August 2023, citing the false-positive problem, the disproportionate flagging of non-native English writers, and the absence of any transparent explanation from the vendor of how its detection algorithm works (Vanderbilt University, 2023).A4 Vanderbilt’s guidance called AI detection “a very difficult task for technology to solve” and warned that the problem will grow as AI tools spread and improve (Vanderbilt University, 2023).

MIT Sloan Teaching and Learning Technologies tells instructors that AI detection software makes frequent errors that can lead to false accusations of misconduct, and it recommends against relying on detectors as a primary enforcement mechanism (MIT Sloan Teaching & Learning Technologies, n.d.). Instead, MIT Sloan recommends a combination of clear syllabus policies, open dialogue with students, assignment redesign, and process-based evidence such as asking students to submit a process statement explaining how they completed their work (MIT Sloan Teaching & Learning Technologies, n.d.). These institutional positions do not argue that AI misuse is acceptable; they argue that detection software is the wrong tool for identifying it (MIT Sloan Teaching & Learning Technologies, n.d.; Vanderbilt University, 2023).

The MIT Sloan resource also cites survey evidence from a 2026 survey that covered more than 3,000 students, instructors, and administrators at over 750 institutions, which found that respondents at institutions that integrated AI were more likely to consider their policy effective than those at institutions that banned it (MIT Sloan Teaching & Learning Technologies, n.d.). That contrast suggests that the energy institutions spend on detection and prohibition may be better invested in redesigning assessments so that AI assistance does not undermine the learning objectives the assignment is meant to serve.

Concrete alternatives to automated detection include requiring students to submit drafts, outlines, and version histories alongside final papers, comparing submitted work against a student's established writing style from earlier in the semester, and holding brief oral follow-up conversations about the ideas in a submitted essay (MIT Sloan Teaching & Learning Technologies, n.d.; Vanderbilt University, 2023). Each of these approaches generates process evidence that is harder to fabricate than a polished final document and does not carry the systematic bias that detector algorithms embed.

Continue reading the full tutorial

Read the full annotated essay.

4 of 6Sections read
4 of 8Notes shown
~3 minRemaining

Read the remaining sections, full references, and all 8 editor annotations — plus the full library of annotated tutorials.

Sign Up Now

V. Why a Score Alone Cannot Carry the Weight of a Misconduct Finding

The Brown University episode illustrates a broader principle: even striking quantitative patterns — a dramatic score drop, a high detector percentage — do not by themselves establish what caused them (Topaz & Bahl, 2026). The statistical preprint analyzing that episode reported that individual changes ranged from a gain of 4 points to a loss of 100 points across the 59 students, and that a model assuming no link between a student's midterm and final fit the data about as well as models allowing one (Topaz & Bahl, 2026).A5 The preprint’s authors were explicit that the scores cannot separate AI use from a final exam that tested substantially different material, so the data alone cannot show which students, if any, had unauthorized help (Topaz & Bahl, 2026).

The same logical structure applies to detector scores: a high percentage output is consistent with AI authorship, but it is also consistent with a non-native writer's limited vocabulary, an unusually formal writing style, or a subject domain that happens to use predictable academic phrasing (Karr et al., 2026; Liang et al., 2023). The preprint on academic integrity detection found that unmodified human-written abstracts from 2023 to 2025 were flagged at rates of 9 to 15%, with much higher rates in non-STEM fields than in STEM, a gap the authors attributed to long-token and Academic Word List density rather than authorship intent (Karr et al., 2026). If a human-written abstract in a humanities field is flagged simply because it uses formal academic vocabulary, then the detector is measuring a stylistic property of the discipline, not evidence of wrongdoing.

Academic misconduct findings carry serious consequences — grade penalties, transcript notations, suspension, and lasting reputational harm — and the evidentiary standard for imposing those consequences should be commensurate with their severity (Liang et al., 2023; Vanderbilt University, 2023). The Stanford researchers warned that a false accusation harms a student’s reputation even if it is later withdrawn, and that relying on these tools risks treating students as guilty until they can prove otherwise (Liang et al., 2023).A6 A detector score that cannot reliably distinguish AI text from non-native human writing, and that can be defeated by a simple prompt revision, does not meet the evidentiary standard that the gravity of a misconduct charge demands (Karr et al., 2026; Liang et al., 2023; Weber-Wulff et al., 2023).

VI. Conclusion: Scores as Conversation Starters, Not Verdicts

The research record is consistent across multiple independent studies: AI detectors are neither accurate enough nor equitable enough to serve as standalone evidence of academic misconduct (Karr et al., 2026; Liang et al., 2023; Weber-Wulff et al., 2023). A peer-reviewed study covering 14 tools found them unreliable and biased toward false negatives, a peer-reviewed study of 7 detectors found a 61.3% false-positive rate for non-native English writers, and a preprint found that honest AI editing is penalized more harshly than deliberate evasion (Karr et al., 2026; Liang et al., 2023; Weber-Wulff et al., 2023). Taken together, these findings mean that a school relying on a detector score to initiate formal misconduct proceedings is building a case on a foundation that the scientific literature has repeatedly found to be unstable.

None of this means that AI misuse in academic work is a fiction or that institutions should abandon efforts to uphold integrity (MIT Sloan Teaching & Learning Technologies, n.d.; Vanderbilt University, 2023). It means that the right response to a suspicious detector score is a structured conversation with the student, supported by process evidence — drafts, version histories, outlines, and the student's own explanation of how the work was produced. MIT Sloan’s guidance recommends open dialogue with students, and Vanderbilt went further, switching its detector off altogether (MIT Sloan Teaching & Learning Technologies, n.d.; Vanderbilt University, 2023).

The Brown University episode, still unresolved in the peer-reviewed literature, is a preview of the disputes that will multiply as AI tools become more capable and more widely used (Topaz & Bahl, 2026).A7 Schools that build their integrity frameworks around detector scores will find themselves in the position that the statistical preprint described: holding data that are consistent with multiple explanations and unable to distinguish among them. The more defensible path — and the one the evidence supports — is to treat a detector score as the beginning of an inquiry, never its conclusion, and to reserve formal sanctions for cases where process evidence, not an algorithm's output, establishes what actually happened (Karr et al., 2026; Liang et al., 2023; MIT Sloan Teaching & Learning Technologies, n.d.; Topaz & Bahl, 2026; Vanderbilt University, 2023; Weber-Wulff et al., 2023).A8

References APA 7 · 7 sources

The toolkit behind the tutorials

Read the example. Then write your own.

Every annotation maps to a tool — outline, thesis, citations, references. Full library access · cancel anytime.

Sign Up Now