Examples
Worked examples
- Is an instance
An institutionally-licensed detector (e.g. Turnitin's AI writing detection module) flagging a submitted student paper as likely containing AI-generated text, routed to a human for review rather than an automatic penalty
- Is an instance
A publisher-side integrity screening tool (e.g. Springer Nature's Geppetto, now part of the STM Integrity Hub) flagging a manuscript section for further human investigation
Counter-examples
Looks similar, but isn't
- Not an instance
A tool that verifies an embedded, generator-created watermark is watermarking verification, not the forensic pattern-based inference this entry covers
- Not an instance
A human reviewer's subjective impression that a document 'reads like AI' with no software tool involved is not a detection tool in this sense
Editorial commentary
These tools infer likely AI origin from statistical patterns — unusually uniform sentence structure, low ‘perplexity’ or predictability, characteristic phrasing — rather than from any cooperative signal deliberately embedded by a generator. That inferential, probabilistic basis is the source of their central, well-documented limitation, and it needs to be stated plainly here because overclaiming detector reliability causes real harm to accused students and researchers.
The false-positive problem is real, documented, and disproportionate
Vendor-claimed accuracy figures are consistently higher than what independent testing finds. Turnitin has publicly claimed roughly 98% accuracy with a false-positive rate under 1% for documents where more than 20% of the text is AI-generated; independent 2024-2025 testing has generally found lower accuracy on unedited generative-AI output and false-positive rates climbing meaningfully higher — reported in some studies in the range of 5-12% — specifically on non-native-English writing, heavily edited drafts, and technical prose. OpenAI’s own AI Text Classifier, launched January 2023, was withdrawn by OpenAI itself in July 2023, citing a ‘low rate of accuracy’: OpenAI’s own testing found it correctly identified only about 26% of AI-written text while mislabelling roughly 9% of genuinely human-written text as AI-generated.
The consequences are not hypothetical. Orion Newby, an Adelphi University student with learning/neurological disabilities, was accused of academic dishonesty after an AI-detection flag; a New York state court ultimately ruled in his favour, reversed the disciplinary findings, and ordered his record expunged. Vanderbilt University disabled Turnitin’s AI-writing detection feature in 2023 over false-positive risk at scale; Curtin University announced it would disable the same feature from January 2026, citing reliability concerns. Several institutions have stopped using these tools, or use them only as a prompt for human conversation rather than as evidence of misconduct on their own. See why does my paper say AI-detected? and Turnitin AI detection vs. standalone AI detectors for more detail, and treat any single detector flag as a starting point for human review, never as standalone proof.
How this differs from related AI-band terms
- vs. watermarking: detection is forensic and after-the-fact, working (unreliably) on any content regardless of source cooperation; watermarking is proactive and requires the generator to have embedded a signal, which makes it more reliable when present but inapplicable when the generator didn’t cooperate.
- vs. AI provenance: provenance is the broader documentation goal; detection tools are one (unreliable) attempt to reconstruct provenance information that was never actually recorded.
Also known as
AI detector · GPT detector
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="Detection tool (AI-generated)"
vocab-term-identifier="https://casrai.org/dictionary/term/detection-tool-ai-generated" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/detection-tool-ai-generated",
"name": "Detection tool (AI-generated)",
"identifier": "https://casrai.org/dictionary/term/detection-tool-ai-generated",
"description": "A software system that estimates the probability that a given piece of content -- typically text, sometimes images -- was produced by a generative AI system, usually by statistically analysing patterns in the content itself without any cooperation from or signal embedded by the original generator, as distinct from watermarking, which requires generator-side cooperation at the point of creation.",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/genai-disclosure#set",
"url": "https://casrai.org/dictionary/term/detection-tool-ai-generated",
"sameAs": [
"AI detector",
"GPT detector"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"author": {
"@id": "https://casrai.org/#editorial-team"
},
"datePublished": "2026-05-21T01:55:32",
"dateModified": "2026-08-22T14:52:05",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
