History of AI text detection: from GLTR to watermarks

With a visualisation, not a verdict. GLTR, by Sebastian Gehrmann, Hendrik Strobelt and Alexander Rush, was submitted to arXiv on 10 June 2019 and presented at the ACL 2019 demo track. It applied "a suite of baseline statistical methods" to a text, colouring each word by how likely a language model found it, so that a human reader could see the tell-tale sameness of generated text. In the authors' human-subjects study the annotation raised the detection rate for fake text from 54 percent to 72 percent "without any prior training" (Gehrmann, Strobelt and Rush 2019, GLTR, arXiv 1906.04043, read 9 November 2026).

Five months later came the first detector that gave a verdict on its own. On 5 November 2019, releasing the 1.5-billion-parameter GPT-2, OpenAI also released a detection model: a sequence classifier based on RoBERTa, fine-tuned to tell GPT-2 output from the WebText data GPT-2 was trained on. OpenAI reported "detection rates of ~95%" for 1.5B GPT-2 text and wrote that this was "not high enough accuracy for standalone detection" and needed pairing with metadata, human judgment and public education (OpenAI, GPT-2: 1.5B release, read 9 November 2026). The same post noted that larger models' outputs were harder to classify and expected detection to get harder as models grew.

Those two 2019 tools set the two families that still define the field. GLTR is a statistics-based detector: it needs a language model to score the text but learns nothing from examples. The GPT-2 detector is a neural classifier: it learns from labelled human and machine text and reports a probability. Every commercial detector this site tests descends from the second family.

It lasted under six months. OpenAI launched a public "AI classifier for indicating AI-written text" on 31 January 2023, a language model fine-tuned on pairs of human and AI text on the same topic. The launch post stated that on a challenge set of English texts the classifier "correctly identifies 26% of AI-written text (true positives)" and labelled human text as AI "9% of the time (false positives)", that it was "very unreliable on short texts (below 1,000 characters)", and that "AI-written text can be edited to evade the classifier" (OpenAI, New AI classifier for indicating AI-written text, read 9 November 2026).

A note at the top of the same page now reads: "As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy." OpenAI wrote that it was researching provenance techniques for text instead. The shutdown is the single most cited event in this history because it was a model maker, with access to its own outputs, saying that a classifier was not good enough to keep online.

That the failures had shape. Three peer-reviewed or widely reviewed papers in 2023 measured where classifiers broke, and their numbers still frame how this site tests.

Liang and colleagues found the flagged TOEFL essays had significantly lower perplexity than the rest, and that asking a chatbot to enrich their vocabulary cut the false positive rate from 61.22 to 11.77 percent (Liang et al. 2023, arXiv 2304.02819, read 9 November 2026). Krishna and colleagues showed the mirror image: paraphrasing AI text made it read as human to the same instruments, and they proposed that model makers keep a database of their own outputs to search against (Krishna et al. 2023, arXiv 2303.13408, read 9 November 2026). Weber-Wulff and colleagues concluded that the tools were "neither accurate nor reliable" and biased toward calling text human (Weber-Wulff et al. 2023, arXiv 2306.15666, read 9 November 2026).

Put together, the three papers say that a classifier reads the predictability of word choice, that predictable human writers get caught, and that unpredictable rewrites of machine text get through. That is why the published method on this site weights meaning kept above detector pass rate and runs human controls beside every reading.

As a paper first, then as a product. Kirchenbauer, Geiping, Wen, Katz, Miers and Goldstein posted "A watermark for large language models" on 24 January 2023 and published it at ICML 2023. Their method selects a randomised set of "green" tokens before each word is generated and softly promotes them during sampling, so a detector that knows the rule can test for too many green tokens "without access to the language model API or parameters", with "negligible impact on text quality" (Kirchenbauer et al. 2023, arXiv 2301.10226, read 9 November 2026). The mechanism is explained in full on the AI text watermarks guide.

Wu and colleagues' survey, posted to arXiv in October 2023 and published in Computational Linguistics in 2025, sorted the whole field into four families: watermarking, statistics-based detectors, neural detectors and human-assisted methods, and listed paraphrase attacks and the lack of a shared evaluation framework among the open problems (Wu et al., A survey on LLM-generated text detection, arXiv 2310.14724, read 9 November 2026). GLTR sits in the human-assisted family, the GPT-2 detector and every commercial tool in the neural one, and Kirchenbauer's method in the first.

Deployment followed regulation. Anthropic's help centre states that it signed the EU AI Act's Article 50(2) Code of Practice on transparency, that Claude models launched in the EU on or after 2 August 2026 "support machine-readable marking at launch", that generated text "will carry embedded watermarks", and that watermark detection is "currently in private preview" for eligible organisations such as regulators, researchers and educational bodies (Anthropic, How Claude marks AI-generated content, read 9 November 2026). The same page lists the limits: a missing mark does not mean a text was not AI-generated, and heavily edited, paraphrased or very short text may carry no detectable signal.

Three things. A classifier reading in 2026 descends from a 2019 design whose own makers called it not good enough for standalone use, and the 2023 studies measured the same two failure modes, false positives on plain human writing and false negatives on paraphrased machine writing, that OpenAI's 2023 launch post listed as limitations. A watermark check answers a narrower question, "does this carry my mark", and in 2026 is available mainly to institutions rather than individuals. And no part of this history produced an instrument that establishes authorship on its own.

No detector score should be the only evidence in a misconduct case. The guides hub holds the rest of this series.