Is an AI detector itself an AI? The four detection methods behind the cards, each sourced
Most are, and the index can show which. Of the 72 cards, 23 describe a trained model, a classifier trained to tell AI text from human text, as the thing that produces the score, by the method statements on the vendor pages read on 8 and 9 November 2026 and the detector classes that Sadasivan 2023 and RAID 2024 use to sort them (arXiv 2303.11156; arXiv 2405.07940, read 8 and 9 November 2026). One card, GLTR, is a statistical display that colours each word by how a language model ranked it and trains nothing (gltr.io, read 9 November 2026). One, SynthID, checks for a watermark that a model embedded when the text or image was made (synthid.com, read 9 November 2026). Two, Grammarly Authorship and Turnitin Clarity, record how a document was written and return no probability at all, and one, Codequiry, pairs a trained model with code similarity (grammarly.com/authorship; Turnitin Clarity, read 9 November 2026). The short answer to the literal question is yes for the tools people mean when they say "AI detector", and the longer answer is that the method decides what a score can and cannot show.
The table sorts every card by method, using four classes that the research literature uses. Sadasivan 2023 tested "watermarking, trained neural network classifiers, zero-shot statistical detectors and a retrieval-based detector" (Sadasivan 2023 record), and RAID 2024 grouped its twelve detectors as neural, metric-based or commercial (RAID 2024 record). The index adds a fifth row for process evidence, because two cards sell a record of writing rather than a judgement about text. A card is placed by the vendor's own description where one exists, and by the tool's published design where the vendor describes only its output. This page is one of the questions answered from the index; its siblings ask whether a 100% accurate detector exists and which detectors are open source, and the AI detector register carries every card.
| Method | What produces the score | Cards | Method statement, source and read date |
|---|---|---|---|
| Trained classifier (neural) | A model trained on human and AI text that outputs a probability | GPTZero, Copyleaks, Originality.ai, Turnitin, Pangram, Winston AI, ZeroGPT, Sapling, Scribbr, QuillBot, Grammarly, Phrasly, JustDone, Proofademic, Walter Writes, Hive, AIDetectPlus, WasItAI, Reality Defender, Writer, Undetectable AI, AI or Not, GPT-2 Output Detector | Grammarly: "Our detection model was trained on tens of thousands of texts, including both human-written and AI-generated text created before 2021" (grammarly.com, 8 November 2026); GPTZero: "We repeatedly train our model" (gptzero.me/faq, 8 November 2026); the GPT-2 Output Detector is "a RoBERTa classifier" fine-tuned on GPT-2 output (openai.com, 9 November 2026); Undetectable AI names the generators it is "trained on" (undetectable.ai, 9 November 2026); Codequiry: "Our current model, v3e, is trained on" named generators (codequiry.com, 9 November 2026) |
| Zero-shot statistics (metric-based) | A fixed language model's own word probabilities, shown without a trained verdict | GLTR | GLTR "colours every word of a text by how highly the GPT-2 117M model ranked that word as the next one"; RAID 2024 groups it as metric-based, open (gltr.io; arXiv 2405.07940, 9 November 2026) |
| Watermark check | A signal embedded by the generating model, read back by a detector built for that signal | SynthID | "SynthID Watermarking: Verifies media by detecting an intentional signal embedded directly during content creation"; "This is not a general AI detector" (synthid.com, 9 November 2026) |
| Process evidence (no score) | A record of typing, pasting and revision while the document was written | Grammarly Authorship, Turnitin Clarity | Authorship "records your actual writing process" and returns no probability for text it did not watch being written (grammarly.com/authorship, 9 November 2026); Clarity's Writing Report records "writing time, sessions, pasted text and playback" and shows AI detection only with the separate Originality add-on (Turnitin Clarity, 9 November 2026) |
| Trained classifier plus code similarity | A trained model on source code, beside a similarity engine | Codequiry | Model v3e "catches 90% of AI/GPT-written code and falsely flags 1.3% of human-written code", with peer and web similarity as separate checks (codequiry.com, 9 November 2026) |
Cards as placed on 9 November 2026 from the method statement on each card; Codequiry is counted once, in its own row, so the five rows sum to 72. Where a vendor publishes only its output format and no training statement, as Hive, WasItAI and Reality Defender do for their image, voice and video models, the card is placed in the classifier row because the product returns a trained model's probability score and no other method is described.
What does a trained classifier do?
It applies a model that was itself produced by machine learning, so the detector is an AI in the plain sense. The oldest card in the index says so in its name: the GPT-2 Output Detector is a RoBERTa classifier that OpenAI fine-tuned on outputs of its own GPT-2 model and released in 2019 with a stated detection rate of about 95% on that model's text (openai.com, read 9 November 2026). Grammarly describes the same design in 2026: a detection model "trained on tens of thousands of texts, including both human-written and AI-generated text" (grammarly.com, read 8 November 2026). GPTZero's FAQ says the company repeatedly trains its model and that the classifier "is not trained to identify AI-generated text after it has been heavily modified" (gptzero.me/faq, read 8 November 2026), and Chaka 2023 records the perplexity scores GPTZero reported for each text, from 38 on English originals up to 2,478 on a German translation, the statistic the model reads before it decides (DOI 10.37074/jalt.2023.6.2.12, read 9 November 2026). Pangram states that its detection "is based off of a large, proprietary dataset that doesn't include user submitted content" (pangram.com/data-privacy, read 8 November 2026). A trained classifier learns the texture of the AI text it was shown, which is why Sadasivan 2023 and Krishna 2023 could lower every classifier's detection rate with a paraphraser the classifier had not seen (Sadasivan 2023 record; Krishna 2023 record; robustness ranking).
How is GLTR different from a classifier?
GLTR trains nothing. The 2019 tool from the MIT-IBM Watson AI Lab and HarvardNLP runs a fixed GPT-2 117M model over the text and colours each word by the rank the model gave it among its predicted next words, green for the top 10, yellow for the top 100, red for the top 1,000 and purple beyond (gltr.io; GLTR card, read 9 November 2026). The reader, not a model, makes the call, which is why the GLTR paper reports a human detection rate rising from 54% to 72% with the display rather than a tool accuracy (claim vs evidence gap ranking). RAID 2024 groups GLTR with Binoculars, Fast-DetectGPT and LLMDet as metric-based detectors, and its finding about that class is the one a reader needs: at a default threshold of 0.5, GLTR flagged 99.3% of human-written documents, falling to 0.05% at 0.95, so the statistic is only as good as the threshold a person sets on it (arXiv 2405.07940, read 9 November 2026). In Orenstrakh 2023, run through the authors' own z-score wrapper, the same statistic was the only method whose accuracy on 10 QuillBot-paraphrased texts held at 100% (arXiv 2307.07411, read 9 November 2026).
Is a watermark detector an AI detector at all?
Only for the models that carry the watermark. SynthID embeds an imperceptible signal when a Google or partner model generates an image, video, audio clip or text, and the Detector reads that signal back; the vendor's own FAQ says "This is not a general AI detector. We can only detect media from the above companies, who have adopted SynthID technology", and a file with no watermark "could have been created by other AI systems" (synthid.com, read 9 November 2026). The method needs cooperation from the generator, which Sadasivan 2023 lists as a class of its own and tests with a paraphraser that weakened the watermark (Sadasivan 2023 record). The text component is open source and scores only text produced with a shared watermark configuration (open-source view). A watermark check answers "did a participating model make this?" and never "did a human write this?", so its false positive question is different in kind from a classifier's.
What do Authorship and Clarity measure instead?
Keystrokes, pastes and time, not language. Grammarly Authorship records whether each sentence in a Google Doc or Word document was typed, pasted from a website, pasted from a generative AI source or produced by Grammarly's own AI, and returns no probability score for a piece of text it did not watch being written (grammarly.com/authorship; support article, read 9 November 2026). Turnitin Clarity's Writing Report records writing time, sessions, pasted text and a playback of every change made inside its browser writing space, with flags set at process thresholds such as pasted words above 50% of the final word count; the AI writing indicator appears inside that report only when the institution licenses the Turnitin Originality add-on, which is the classifier on the Turnitin card (Reviewing the Writing Report, read 9 November 2026). Neither tool has a false positive rate because neither makes the kind of claim a false positive contradicts; both have an empty field where the classifiers carry a figure, and no indexed study has tested either (data retention ranking for what each keeps).
The method also decides what the studies in the index could measure. Every one of the 35 records scored a classifier or a statistic on text or images; none could score a process record, and only Sadasivan 2023 ran a watermark, a research one rather than SynthID. A reader choosing a tool is choosing a method first and a vendor second.
No detector score should be the only evidence in a misconduct, hiring or publishing decision, and a process record is evidence of a different thing from a score.
Three process-evidence tools, GPTZero's writing report, Originality.ai's Writing Replay and Grammarly Authorship, run inside Google Docs, and the Google Docs question, added 9 November 2026, lists the 12 cards with a Docs surface of any kind.
Questions about how AI detectors work
Do AI detectors use AI?
Twenty-three of the 72 cards describe or return a trained classifier's probability, by the vendor statements read 8 and 9 November 2026, so yes for most. GLTR shows a language model's statistics without a trained verdict, SynthID reads a watermark, and Grammarly Authorship and Turnitin Clarity record the writing process and return no score.
Is GPTZero an AI?
Yes, by its own description: GPTZero's FAQ says "We repeatedly train our model" and that the classifier is not trained to identify heavily modified AI text (gptzero.me/faq, read 8 November 2026). The model reports perplexity and burstiness statistics along the way; Chaka 2023 prints the perplexity values it returned (DOI 10.37074/jalt.2023.6.2.12, read 9 November 2026).
What is the difference between a classifier and a watermark?
A classifier judges any text by patterns it learned from training data; a watermark detector reads a signal that a cooperating model embedded at generation and says nothing about text from other models (synthid.com, read 9 November 2026). Sadasivan 2023 treats the two as separate detector classes (arXiv 2303.11156, read 9 November 2026).
Is Grammarly Authorship an AI detector?
Not by the index's definition: it records typing, pasting and AI assistance while a document is written and returns no probability for pasted text (grammarly.com/authorship, read 9 November 2026). Paid plans add AI detection inside the report, which is the classifier on the Grammarly card.
Which method has the best record in the studies?
The index ranks results, not methods, and the records cut across them: GLTR's statistic posted both the highest false positive rate at a default threshold (99.3%, RAID 2024) and the only 100% after a paraphrase (Orenstrakh 2023), while classifiers posted every result in between (false positive rate ranking; robustness ranking). No process tool or SynthID has a record.
Sources
All read 8 November 2026 unless a later date is shown.
- Sadasivan et al. 2023: https://arxiv.org/abs/2303.11156 (read 9 November 2026)
- Dugan et al. 2024, RAID: https://arxiv.org/pdf/2405.07940
- Krishna et al. 2023: https://arxiv.org/abs/2303.13408 (read 9 November 2026)
- Orenstrakh et al. 2023: https://arxiv.org/abs/2307.07411 (read 9 November 2026)
- Chaka 2023: https://doi.org/10.37074/jalt.2023.6.2.12 (read 9 November 2026)
- OpenAI, GPT-2 1.5B release: https://openai.com/index/gpt-2-1-5b-release/ (read 9 November 2026)
- GLTR: http://gltr.io/ (read 9 November 2026)
- Grammarly AI detector: https://www.grammarly.com/ai-detector
- Grammarly Authorship: https://www.grammarly.com/authorship (read 9 November 2026)
- Grammarly Authorship support article: https://support.grammarly.com/hc/en-us/articles/29548735595405 (read 9 November 2026)
- GPTZero FAQ: https://gptzero.me/faq
- Pangram data privacy: https://www.pangram.com/data-privacy
- SynthID Detector: https://synthid.com/ (read 9 November 2026)
- Turnitin Clarity: https://www.turnitin.com/products/feedback-studio/clarity (read 9 November 2026)
- Turnitin, Reviewing the Writing Report: https://guides.turnitin.com/hc/en-us/articles/36916426919949-Reviewing-the-Writing-Report (read 9 November 2026)
- Undetectable AI detector: https://undetectable.ai/ai-detector (read 9 November 2026)
- Codequiry AI code detection: https://codequiry.com/detect-ai-written-code (read 9 November 2026)
Revision log: 2026-11-09: page created from the method statements of 28 cards at index version 2026.11.4.; 2026-11-09: eighth-wave pages linked
Related pages: Is there an AI detector for short texts? Every stated minimum length in the index, Which AI detectors claim to detect ChatGPT? The generators field of all 38 cards, AI detectors ranked by published robustness to paraphrased and edited text and Questions about AI detectors, answered from the index.