Have AI detector scores been used as evidence? What the primary documents in the index state, and what 24 vendors say about acting on a score
The index holds no court filing, hearing transcript or disciplinary decision in which an AI detector score was entered as evidence, and this page cites none; what it holds, as of 9 November 2026, is 7 primary documents that state how a score should or should not be used as evidence, from Vanderbilt University's statement of 16 August 2023 disabling Turnitin's indicator because a 1% false positive rate on 75,000 papers a year implied "around 750 student papers" wrongly labelled (vanderbilt.edu, read 9 November 2026) to Turnitin's own FAQ, which says the company "does not make a determination of misconduct" and that the indicator "should not be used as the sole basis for action" (guides.turnitin.com, read 8 November 2026). Beside those documents, 24 of the 72 vendor cards carry a published caveat against acting on a score alone, and 18 carry none, by the vendor caveat field of every card read 8 and 9 November 2026 in the data file.
This page is a record, not advice. It answers the question from documents the index has read in the primary source: an institution's published decision, a vendor's help centre and blog, a lab's withdrawal notice, a study's discussion section, an examination board's policy pages and the vendors' own caveats. Press reports of accusations and court cases exist and are not cited here, because the index has not read a filing; a reader who needs one should read the filing, not a summary of it. The first table lists the seven documents with the sentence in each that bears on evidence; the second quotes every vendor caveat in the index. This page is one of the questions answered from the index; its siblings ask whether a 100% accurate AI detector exists, which detectors claim a false positive rate under 1% and what AI detector College Board uses, the false positive rate ranking holds every study-measured rate, and the AI detector register carries every card.
| Document | Who published it | What it states about a score as evidence | Source | Read date |
|---|---|---|---|---|
| Guidance on AI Detection and Why We're Disabling Turnitin's AI Detector, 16 August 2023 | Vanderbilt University, Brightspace and Instructional Technology Support | Turnitin's claimed 1% false positive rate on the 75,000 papers submitted in 2022 meant "around 750 student papers could have been incorrectly labeled"; the detector was disabled "for the foreseeable future"; "we do not believe that AI detection software is an effective tool that should be used" | vanderbilt.edu | 9 November 2026 |
| Turnitin's New AI Writing Detection (and what it means for your teaching), 4 April 2023 | Vanderbilt University | False positives were "not just a possibility, but an inevitability"; the tool should be "a guideline, not a grading metric" | vanderbilt.edu | 9 November 2026 |
| Turnitin's AI writing detection capabilities FAQs | Turnitin | "Turnitin does not make a determination of misconduct; rather, it provides data for the educators to make an informed decision"; the percentage "should not be used as the sole basis for action or a definitive grading measure" | guides.turnitin.com | 8 November 2026 |
| AI writing detection update from Turnitin's Chief Product Officer, 23 May 2023 | Turnitin | "real-world use is yielding different results from our lab"; documents scored under 20% show "a higher incidence of false positives"; "behind each false positive instance is a real student who may have put real effort into their original work" | turnitin.com | 8 November 2026 |
| New AI classifier for indicating AI-written text, 31 January 2023, withdrawal note 20 July 2023 | OpenAI | The classifier identified 26% of AI-written text and mislabelled human text 9% of the time; withdrawn "due to its low rate of accuracy" | openai.com | 8 November 2026 |
| Who wrote this? Evaluating the reliability of AI detection tools in higher education, 29 June 2026 | Van Vlasselaer, Van Droogenbroeck and Spruyt, International Journal for Educational Integrity | Turnitin scored all 40 fully AI-generated papers as human; the Discussion recommends that a flag from any tool trigger review rather than a conclusion | DOI 10.1007/s40979-026-00226-w | 9 November 2026 |
| AP Guidance for Artificial Intelligence Tools and Other Services, 2026-27 | College Board | Names no detector; College Board "reserves the right to investigate submissions where there is evidence of the inappropriate use of generative AI" and to "request from students copies of their interim work for review"; AP Seminar and Research rely on teacher checkpoints and affirmations | apcentral.collegeboard.org | 9 November 2026 |
Statements as quoted from each primary document on the date shown; the records behind rows 1 to 6 are the Vanderbilt 2023 record, the Turnitin claims record, the OpenAI classifier record and the Van Vlasselaer 2026 record.
Which primary documents in the index speak to scores as evidence?
Seven, from four kinds of author, and none of them is a court. The one institutional decision is Vanderbilt's: the university enabled Turnitin's indicator for faculty on 4 April 2023 with a warning to treat it as "a guideline, not a grading metric", then disabled it on 16 August 2023 on the arithmetic of the vendor's own claim, 1% of 75,000 papers, and on three further grounds, opacity, the withdrawal of other detectors and student-data privacy (vanderbilt.edu, read 9 November 2026; Vanderbilt 2023 record). No test was run, and the record notes that Vanderbilt rounded "less than 1%" to 1% and applied it to every paper rather than to the documents scored over 20% AI that Turnitin's rate describes.
The vendor whose tool Vanderbilt disabled says much the same about evidence in its own voice. Turnitin's FAQ states that the company "does not make a determination of misconduct" and that the indicator "should not be used as the sole basis for action", and its post of 23 May 2023 concedes that "real-world use is yielding different results from our lab" (guides.turnitin.com, read 8 November 2026; turnitin.com, read 8 November 2026). OpenAI's withdrawal note of 20 July 2023 is the one case of a lab removing its own detector for accuracy, after publishing a 26% true positive rate and a 9% false positive rate (openai.com, read 8 November 2026). Van Vlasselaer 2026 is the one study record whose Discussion addresses the use of a result: after Turnitin, GPTZero and Copyleaks scored all 40 fully AI papers as human, the authors recommend that a flag trigger review rather than a conclusion (DOI 10.1007/s40979-026-00226-w, read 9 November 2026). College Board's AP pages name no detector and define evidence as interim work and teacher checkpoints, as the College Board question sets out (apcentral.collegeboard.org, read 9 November 2026).
What do the vendors themselves say about acting on a score?
Twenty-four of the 72 cards carry a sentence on their own pages about acting on a result, and 18 carry none, by the vendor caveat field read 8 and 9 November 2026.
| Detector | Vendor caveat on acting on a score, as published | Source | Read date |
|---|---|---|---|
| QuillBot | "Never make an academic or employment decision on a detector score alone" | quillbot.com | 8 November 2026 |
| Grammarly | "you should never rely on the results of an AI detector alone"; "AI detection should never be used as a standalone verification method" | grammarly.com | 8 November 2026 |
| GPTZero | "Results should not be used to punish or as the final verdict"; "These results should not be used to punish students" | gptzero.me | 8 November 2026 |
| Phrasly | "These results should not be used to punish students or writers, instead consider them as part of a holistic assessment" | phrasly.ai | 8 November 2026 |
| Sapling | "No current AI content detector (including Sapling's) should be used as a standalone check"; "A detector score is one input to a review, not definitive proof" | sapling.ai | 8 November 2026 |
| JustDone | "Every AI detection result is a probability, not a verdict"; "An AI score is a probability, not proof" | justdone.com | 8 November 2026 |
| Originality.ai | "An AI score shouldn't be used alone for academic integrity discipline"; "Always keep a human reviewer in the loop" | originality.ai | 8 November 2026 |
| ZeroGPT | "You must not treat a ZeroGPT Output as the sole basis for a decision that could materially affect another person's rights, education, employment, admission" (terms of use) | zerogpt.com | 8 November 2026 |
| Proofademic | Results "should not be the sole basis for an academic-misconduct decision"; "A flag is not proof of misconduct" | proofademic.ai | 8 November 2026 |
| Walter Writes | "Treat AI and source-similarity scores as review signals, not final verdicts" | walterwrites.ai | 8 November 2026 |
| Turnitin | "Turnitin does not make a determination of misconduct"; the percentage "should not be used as the sole basis for action or a definitive grading measure by instructors" | guides.turnitin.com | 8 November 2026 |
| AIDetectPlus | "Can I use a result to accuse a student? No, and we would ask you not to. A detection score is a reason to open a conversation, not evidence to close one" | aidetectplus.com | 9 November 2026 |
| GPT-2 Output Detector | "strongly suggested not to use it as a ChatGPT detector for the purposes of making grave allegations of academic misconduct against undergraduates and others" | huggingface.co | 9 November 2026 |
| Codequiry | "Codequiry does not determine whether plagiarism, academic misconduct or unauthorised use has occurred"; results "should be independently reviewed and interpreted by qualified personnel in light of all relevant circumstances before any action is taken" | codequiry.com | 9 November 2026 |
| Writer | "We don't recommend using Writer (or any detector) without a human watching over it" (disclaimer, January 2024 revision) | web.archive.org | 9 November 2026 |
| Smodin | "An indicator in a report is not definitive proof of inappropriate AI use"; "Smodin AI Detector is a Tool, Not a Judge" | smodin.io | 9 November 2026 |
| BrandWell | Results "do not verify authorship and should not be used as proof in academic, employment, legal, or disciplinary decisions" (Moxby listing) | moxby.com | 9 November 2026 |
| Surfer | "AI content analyzers are good for initial screening, but human oversight is necessary to ensure that their results are correct" | surfer.ai | 9 November 2026 |
| Compilatio | "This score is not proof, but rather an indicator that must be interpreted" | compilatio.net | 9 November 2026 |
| TruthScan | "You set the thresholds, so most teams route borderline scores to review rather than auto-deny" (pricing FAQ) | truthscan.com | 9 November 2026 |
| Sightengine | The detector "should never be used as a substitute for human judgment" | sightengine.com | 9 November 2026 |
| Pindrop Pulse | The product provides "a probabilistic determination as to whether a Meeting Participant may be human or a machine" (meeting terms); a result "can help inform routing, verification, agent handling, or fraud review rather than automatically determining the outcome of every call" | pindrop.com | 9 November 2026 |
| SynthID | "not a standalone guarantee; we recommend collecting multiple points of data before making a decision" | synthid.com | 9 November 2026 |
| Reality Defender | "Uncertain material needs further review; a media assessment does not establish whether a recorded claim is true" | realitydefender.com | 9 November 2026 |
Caveats as quoted on each card from the page read on the date shown. The 18 cards without a sentence on acting on a result are Scribbr, Winston AI, Pangram, Copyleaks, Hive, GLTR, WasItAI, Undetectable AI, AI or Not, Grammarly Authorship, Turnitin Clarity, Quetext, DetectGPT, Deepware, Sensity, Codeleaks, Illuminarty and Resemble Detect; three of those cards record the absence in terms, Winston AI ("No sole-basis style caveat published"), Pangram ("No explicit not-the-sole-basis statement by Pangram itself on pages read") and Copyleaks ("No explicit sole-basis statement on pages read"), and Quetext's page says the point "is not about punishing the use of AI", a statement of purpose rather than of use.
The pattern across the 24 is one sentence in different words: a score is a probability, a signal or an indicator, and a decision needs something else. Four vendors name the decisions they mean. ZeroGPT's terms name "rights, education, employment, admission"; BrandWell's listing names "academic, employment, legal, or disciplinary decisions"; QuillBot names "an academic or employment decision"; Proofademic names "an academic-misconduct decision" (zerogpt.com, quillbot.com, read 8 November 2026; moxby.com, read 9 November 2026). Two of the sentences come from the two detectors the studies have tested most: Turnitin, with 5 records, and GPTZero, with 16 (ranking by independent testing).
What would a score need before it could stand as evidence, on the record?
The documents above name the missing pieces without supplying them. A base: Turnitin's rate is stated for documents scored over 20% AI on 700,000 to 800,000 pre-ChatGPT papers of unpublished composition, and the study bases in the index run from 10 human samples in Perkins 2024 to 1,992 in Jabarian and Imas 2025 (Turnitin claims record; false positive rate ranking). A version and a date: every Turnitin figure in the index predates the first of seven dated model changes on 6 December 2023, and no study has measured whether the same text scores the same twice, as the question on repeated results records. A threshold: Jabarian and Imas 2025 reports Pangram's false positive rate as 0.0000 at a fixed 0.5 threshold and up to 0.0075 per genre at Youden thresholds on the same texts (NBER w34223, read 9 November 2026). And a second kind of record: the two cards built as records rather than scores, Grammarly Authorship and Turnitin Clarity, log how a text was written and return no probability (Grammarly Authorship card; Turnitin Clarity card), which is the design College Board's checkpoints describe without naming a product.
A detector score is one vendor's probability about one text on one date, and no document in the index treats it as proof of anything else.
The law firms view, added 9 November 2026, reads this page's 7 documents beside the 9 cards that state both a retention undertaking and an enterprise deployment or report export, the fields a firm would check before any score exists.
Questions about AI detector results as evidence
Have AI detector results been used as evidence in hearings or court?
The index holds 0 court filings, hearing transcripts or disciplinary decisions and cites none; its 7 primary documents on the question are an institution's decision (Vanderbilt, 16 August 2023), a vendor's FAQ and blog (Turnitin), a lab's withdrawal note (OpenAI, 20 July 2023), a study's recommendation (Van Vlasselaer 2026) and an examination board's policy (College Board), all read 8 and 9 November 2026.
Does Turnitin say its AI score is evidence of misconduct?
No: Turnitin's FAQ states that the company "does not make a determination of misconduct" and that the percentage "should not be used as the sole basis for action or a definitive grading measure by instructors" (guides.turnitin.com, read 8 November 2026).
Why did Vanderbilt stop using Turnitin's AI detector?
Because Turnitin's claimed 1% false positive rate applied to 75,000 papers a year implied about 750 wrongly labelled papers, and because of "no insight into how it works", the withdrawal of other detectors and student-data privacy, per its post of 16 August 2023 (vanderbilt.edu, read 9 November 2026).
How many AI detector vendors tell users not to act on a score alone?
24 of the 72 cards carry such a sentence on the vendor's own pages, read 8 and 9 November 2026, from QuillBot's "Never make an academic or employment decision on a detector score alone" to Turnitin's "should not be used as the sole basis for action"; 18 cards carry none.
Has any detector vendor withdrawn its tool over accuracy?
One lab has: OpenAI removed its AI text classifier on 20 July 2023 "due to its low rate of accuracy", having published a 26% true positive rate and a 9% false positive rate at launch (openai.com, read 8 November 2026). The question on shut-down detectors lists every withdrawal on the record.
Sources
All read 9 November 2026 unless an earlier date is shown; the vendor caveat source of every card is linked in the second table.
- Vanderbilt University, Guidance on AI Detection and Why We're Disabling Turnitin's AI Detector (16 August 2023): https://www.vanderbilt.edu/brightspace/2023/08/16/guidance-on-ai-detection-and-why-were-disabling-turnitins-ai-detector/
- Vanderbilt University, Turnitin's New AI Writing Detection (4 April 2023): https://www.vanderbilt.edu/brightspace/2023/04/04/turnitins-new-ai-writing-detection-and-what-it-means-for-your-teaching/
- Turnitin AI writing detection FAQs: https://guides.turnitin.com/hc/en-us/articles/28477544839821-Turnitin-s-AI-writing-detection-capabilities-FAQs (read 8 November 2026)
- Turnitin, AI writing detection update from the Chief Product Officer (23 May 2023): https://www.turnitin.com/blog/ai-writing-detection-update-from-turnitins-chief-product-officer (read 8 November 2026)
- OpenAI, New AI classifier for indicating AI-written text (31 January 2023, withdrawal note 20 July 2023): https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/ (read 8 November 2026)
- Van Vlasselaer et al. 2026: https://doi.org/10.1007/s40979-026-00226-w
- College Board, AP Guidance for Artificial Intelligence Tools and Other Services: https://apcentral.collegeboard.org/exam-administration-ordering-scores/administering-exams/exam-policies/artificial-intelligence-tools
- Jabarian and Imas 2025, NBER w34223: https://www.nber.org/papers/w34223
- QuillBot AI detector: https://quillbot.com/ai-content-detector (read 8 November 2026)
- Grammarly AI detector: https://www.grammarly.com/ai-detector (read 8 November 2026)
- GPTZero FAQ: https://gptzero.me/faq (read 8 November 2026)
- Phrasly AI detector: https://phrasly.ai/ai-detector (read 8 November 2026)
- Sapling AI detector: https://sapling.ai/ai-content-detector (read 8 November 2026)
- JustDone AI detector: https://justdone.com/ai-detector (read 8 November 2026)
- Originality.ai, AI accuracy: https://originality.ai/blog/ai-accuracy (read 8 November 2026)
- ZeroGPT terms of use: https://www.zerogpt.com/terms-of-use (read 8 November 2026)
- Proofademic AI text detector: https://proofademic.ai/ai-text-detector/ (read 8 November 2026)
- Walter Writes AI detector API: https://walterwrites.ai/ai-detector-api/ (read 8 November 2026)
- AIDetectPlus AI detector: https://aidetectplus.com/ai-detector
- GPT-2 Output Detector model card: https://huggingface.co/openai-community/roberta-base-openai-detector
- Codequiry AI code detection: https://codequiry.com/detect-ai-written-code
- Writer AI content detector disclaimer (archived): http://web.archive.org/web/20251209115654/https://writer.com/disclaimer-aicd/
- Smodin AI detector: https://smodin.io/ai-content-detector
- Moxby AI Content Detector Mod (BrandWell): https://moxby.com/marketplace/ai-content-detector/
- Surfer AI detector: https://surfer.ai/ai-content-detector/
- Compilatio AI detector information: https://www.compilatio.net/en/ai-detector-info/
- TruthScan pricing: https://truthscan.com/pricing
- Sightengine AI image detector: https://sightengine.com/detect-ai-generated-images
- Pindrop Pulse meeting terms: https://www.pindrop.com/legal/pulse-meeting-terms
- SynthID: https://synthid.com/
- Reality Defender deepfake video detection: https://www.realitydefender.com/deepfake-video-detection
Revision log: 2026-11-09: page created from the primary documents in four study records, the College Board question and the vendor caveat field of 42 cards at index version 2026.11.7; no court record is held or cited.; 2026-11-09: eighth-wave pages linked
Related pages: Which AI detectors detect both AI text and AI images? The 15 cards that list both, with the study record behind each modality, Which AI detectors have a free API tier? The ten vendors whose free allowance covers API calls, and the twenty that sell an API with no free tier, AI detectors ranked by independent testing Two records added in the fifth wave on 9 November 2026, Pegoraro 2023 and Fleckenstein 2024, changed nine counts and none: Pegoraro 2023 tested ZeroGPT, GPTZero, Copyleaks, Originality.ai, Writer, Content at Scale (BrandWell), DetectGPT, the GPT-2 Output Detector and GLTR, lifting GPTZero to 12, the GPT-2 Output Detector to 10, Copyleaks to 7 and Originality.ai and ZeroGPT to 6; Fleckenstein 2024 tested 289 teachers and no software, so it counts for no card ([Fleckenstein 2024 record](/studies/fleckenstein-2024-teachers/)). Illuminarty, one of four cards added in the same wave, joined with its one image record from Ha and Passananti 2024, so 18 of 72 cards now qualify. and Questions about AI detectors, answered from the index.