GPT-2 Output Detector
Vendor OpenAI Official site github.com/openai/gpt-2-output-dataset/tr... How this card is sourced
Updated · AI Detector Index editors
The GPT-2 Output Detector is a RoBERTa classifier that OpenAI released on November 5, 2019 alongside the 1.5B-parameter GPT-2 model, with a stated detection rate of about 95% on GPT-2 text, according to the OpenAI release post read 9 November 2026. The hosted demo at openai-openai-detector.hf.space showed "The space is paused, ask a maintainer to restart it" on the same day, and the GitHub repository that holds the code is archived. Nine of the 35 study records in the index tested this tool, more than any card except GPTZero; four of the nine were on the card on the morning of 9 November 2026 and five were added later that day. The card sits in the detector index with twenty-one others; how this card is sourced explains the field rules.
What GPT-2 Output Detector detects
| Field | Value | Source |
|---|---|---|
| Content types detected | Text | huggingface.co/openai-community/roberta-b... |
| Generators covered | GPT-2 (a RoBERTa classifier fine-tuned on outputs of the 1.5B-parameter GPT-2 model against WebText; the model card says it is strongly suggested not to use it as a ChatGPT detector) | huggingface.co/openai-community/roberta-b... |
| Input limits | Evaluated on 510-token examples (5,000 WebText and 5,000 GPT-2 samples per the model card); the local web UI accepts pasted text with no published character limit. Weights: detector-base.pt 478 MB (roberta-base), detector-large.pt 1.5 GB (roberta-large). | huggingface.co/openai-community/roberta-b... |
| Free tier | Free: MIT-licensed weights and code on GitHub and the model on Hugging Face, run locally (python -m detector.server, web UI at localhost:8080); the hosted Hugging Face demo is paused | github.com/openai/gpt-2-output-dataset/tr... |
| API | Yes, huggingface.co/openai-community/roberta-b... | huggingface.co/openai-community/roberta-b... |
| Launch year | 2019 | openai.com/index/gpt-2-1-5b-release |
| Ownership | OpenAI. Released November 5, 2019 with the 1.5B-parameter GPT-2 model under the MIT License. The GitHub repository is archived (read-only, last pushed December 13, 2023) and the hosted Hugging Face demo at openai-openai-detector.hf.space shows: The space is paused, ask a maintainer to restart it. | openai-openai-detector.hf.space |
| Independent studies | 13 records | Studies index |
The model detects text generated by GPT-2. The Hugging Face model card describes it as "the GPT-2 output detector model, obtained by fine-tuning a RoBERTa base model with the outputs of the 1.5B-parameter GPT-2 model", and adds that "it is strongly suggested not to use it as a ChatGPT detector for the purposes of making grave allegations of academic misconduct" (model card, read 9 November 2026). Two sizes were released: a roberta-base detector of 478 MB and a roberta-large detector of 1.5 GB (detector README, read 9 November 2026).
The training data was OpenAI's own GPT-2 output dataset: 250,000 WebText documents and, for each GPT-2 size, 250,000 random samples and 250,000 top-k 40 samples (repository README, read 9 November 2026). The model card states the test set as 5,000 WebText samples and 5,000 GPT-2 samples of 510 tokens each. No language list is published; WebText is English web text.
The repository is archived and read-only, last pushed December 13, 2023, under the MIT License (GitHub, read 9 November 2026). The tool's nearest relative on the index is GLTR, another 2019 research tool built around GPT-2; the two were often tested side by side.
Limits and plans
There are no plans and no prices. The weights download from OpenAI's public storage, and the README runs a local web UI at localhost:8080 with python -m detector.server detector-base.pt (detector README, read 9 November 2026). The hosted demo, the address every indexed study used, is paused as of 9 November 2026, so a reader who wants the tool today runs it locally.
The model card publishes no character limit for the local interface; the evaluation used 510-token examples. OpenAI publishes no retention statement for the detector, because nothing leaves the machine when the model runs locally.
What the vendor claims
We conducted in-house detection research and developed a detection model that has detection rates of ~95% for detecting 1.5B GPT-2-generated text. We believe this is not high enough accuracy for standalone detection and needs to be paired with metadata-based approaches, human judgment, and public education to be more effective.
OpenAI's release post states: "We conducted in-house detection research and developed a detection model that has detection rates of ~95% for detecting 1.5B GPT-2-generated text. We believe this is not high enough accuracy for standalone detection and needs to be paired with metadata-based approaches, human judgment, and public education to be more effective" (openai.com, November 5, 2019, read 9 November 2026). The repository README puts the figures as "accuracies in the mid-90s for Top-K 40 generations, and mid-70s to high-80s (depending on model size) for random generations", and adds: "We also find some evidence that adversaries can evade detection via finetuning from released models."
OpenAI publishes no false positive rate for the detector. The model card notes that nucleus sampling outputs "proved most difficult to correctly classify" and that larger generators are harder to detect (model card, read 9 November 2026).
The vendor's own warning is the caveat that matters: a model trained to tell GPT-2 from WebText in 2019 is being asked, in every indexed study, to judge text from models it never saw.
What independent studies found
- AIDI-S-0001Weber-Wulff 20232023
GPT-2 Output Detector Demo: 39 of 54 documents correct (72%, rank 3 of 14, Table 5); 2 false accusations in 18 human-written documents (11.1%, both machine-translated texts, Table 9); 12 of 36 AI documents undetected (33.3%, Table 10). Tested in March 2023 at openai-openai-detector.hf.space; the authors note the tool was not trained to detect GPT-3.5 output.
- AIDI-S-0002Liang 20232023
Listed as OpenAI (openai-openai-detector.hf.space), one of seven detectors run on 91 human-written TOEFL essays and 88 US eighth-grade essays, accessed March 15, 2023; the average false positive rate across the seven was 61.22% on the TOEFL essays and 5.19% on the US essays, with per-detector rates given only in a chart.
- AIDI-S-0005Perkins 20242024
GPT-2 detector: baseline mean accuracy 57.2% (rank 4 of 7, Table 7); 34.7% on non-manipulated AI text falling to 17.5% after adversarial techniques (drop 17.2%, Table 8); 0 false accusations in 10 human samples; 81 of 104 AI samples undetected (UDC 77%, Table 10).
- AIDI-S-0006RAID 20242024
RoBERTa (GPT-2), base and large: false positive rate at threshold 0.5 of 6.59% (base) and 2.91% (large) on RAID human-written documents (Table 4); accuracy at 5% FPR for RoBERTa-Large of 79.7 on open chat models with greedy decoding and 98.5 on open non-chat models with greedy decoding, 19.5 on open chat models under sampling with a repetition penalty (Table 5); RoBERTa-Large 56.7 with no attack, 72.9 under paraphrase (+16.2), 21.3 under homoglyphs (-35.4) (Table 6).
- AIDI-S-0010Sadasivan 20232023
The RoBERTa-Large GPT-2 detector's true positive rate at 1% FPR fell from 1.0 to 0.925 after one T5 paraphrase round and to 0.792 after ten paraphrases with detector queries, on 1,000 GPT-2 Medium passages (Table 5); the paper's introduction states the drop as 100% to 60%.
- AIDI-S-0012Orenstrakh 20232023
Scored 98.25% on 114 human and 95.00% on 20 ChatGPT English coursework submissions by the threshold method, with 2 false positives in 114; accuracy on 10 QuillBot-paraphrased ChatGPT texts fell from 100% to 60% (Tables 6, 8 and 9).
- AIDI-S-0013Gao 20232023
Separated 50 ChatGPT-generated medical abstracts from 50 journal abstracts with an AUROC of 0.94, sensitivity 86% and specificity 94% at a 1.46% cutoff; median fake scores 99.98% and 0.02%.
- AIDI-S-0014Jabarian and Imas 20252025
Run as the open-source RoBERTa-base baseline: false positive rate 0.9503 at a 0.5 threshold on 1,992 human passages and 0.8991 at 0.9, AUROC 0.3957 to 0.7845, and up to 51% of AI passages missed at Youden-optimised thresholds (Tables 3, 1 and 7).
- AIDI-S-0016NIST AI 700-1 20252025
Named as one of two NIST baseline discriminators in the GenAI text-to-text pilot; the report prints no result under this detector's name, and participant teams are anonymised.
- AIDI-S-0018Pegoraro 20232023
Run as the Solaiman et al. RoBERTa GPT-2 detector: true positive rate 7.2% on ChatGPT responses and true negative rate 96.4% on human responses on a ChatGPT and human Q&A benchmark, March 2023 (Table I).
- AIDI-S-0021Ibrahim 20232023
16 of 120 ESL student essays (13.3%) scored below a 95% originality cut and 10 of 120 ChatGPT (GPT-3.5, January 2023) essays (8.3%) scored above a 5% cut; the author's average accuracy is 89.2% (Results and Discussion RQ1).
- AIDI-S-0024Alshammari and Rao 20252025
Average AI score 8.15% on 49 DeepSeek-V3 answers and 3.65% after a QuillBot paraphrase, with an average human score of 93.73% on 49 human answers (Tables III and IV, March 2025).
- AIDI-S-0033Rashidi 20232023
Run locally as the RoBERTa-base OpenAI detector on 14,400 real journal abstracts (1980 to February 2023, five journals, 1,600 per five-year range): up to 8.69% of a range's abstracts scored above 50% AI probability (1995 to 1999) and up to 5.13% above 90% (2000 to 2004), falling to 0.75% above 90% for 2020 to 2023; by journal, 10.24% of NEJM abstracts scored above 90% against 1.89% to 3.89% for the other four (Discussion, Fig. 2 and Table 1). No AI-written text was scored.
Weber-Wulff 2023 tested the hosted demo and scored it 39 of 54 documents correct (72%), rank 3 of 14 tools behind Turnitin and Compilatio, with 2 false accusations in 18 human-written documents (11.1%), both on machine-translated texts, and 12 of 36 AI documents undetected (33.3%) (arXiv 2306.15666, Tables 5, 9 and 10, read 9 November 2026). The authors call the result "particularly interesting because as the name suggests, GPT-2 Output Detector was not trained to detect GPT-3.5 output."
Liang 2023 lists the tool by its address, openai-openai-detector.hf.space, as one of seven detectors run on 91 TOEFL essays and 88 US eighth-grade essays on March 15, 2023; the average false positive rate across the seven was 61.22% on the TOEFL essays and 5.19% on the US essays, with per-detector rates only in a chart (arXiv 2304.02819v3, read 9 November 2026).
Perkins 2024 gives the GPT-2 detector a baseline mean accuracy of 57.2%, rank 4 of 7, with 0 false accusations in 10 human samples and 81 of 104 AI samples undetected; accuracy on non-manipulated AI text fell from 34.7% to 17.5% after adversarial techniques (arXiv 2403.19148, Tables 7, 8 and 10, read 9 November 2026).
The RAID benchmark (2024) ran both sizes under the name RoBERTa (GPT-2) and reports false positive rates at threshold 0.5 of 6.59% for the base model and 2.91% for the large model; at a fixed 5% false positive rate, RoBERTa-Large scored 79.7 on open chat models with greedy decoding, 98.5 on open non-chat models with greedy decoding and 19.5 on open chat models under sampling with a repetition penalty, and moved from 56.7 with no attack to 72.9 under paraphrase and 21.3 under homoglyphs (arXiv 2405.07940, Tables 4 to 6, read 9 November 2026).
Five records were added on 9 November 2026. Sadasivan 2023 reports the RoBERTa-Large detector's true positive rate at 1% FPR falling from 1.0 to 0.925 after one T5 paraphrase round and to 0.792 after ten paraphrases with detector queries on 1,000 GPT-2 Medium passages, with the introduction stating the drop as 100% to 60% (arXiv 2303.11156, read 9 November 2026). Orenstrakh 2023 scored the hosted demo at 98.25% on 114 human and 95.00% on 20 ChatGPT coursework submissions, 2 false positives in 114, and 60% on 10 QuillBot-paraphrased texts from 100% before (arXiv 2307.07411, read 9 November 2026). Gao 2023 separated 50 ChatGPT medical abstracts from 50 journal abstracts with an AUROC of 0.94, sensitivity 86% and specificity 94% at a 1.46% cutoff (DOI 10.1038/s41746-023-00819-6, read 9 November 2026). Jabarian and Imas 2025 ran the RoBERTa-base weights as an open-source baseline and reports a false positive rate of 0.9503 at threshold 0.5 on 1,992 human passages and AUROC between 0.3957 and 0.7845 (NBER w34223, read 9 November 2026). NIST AI 700-1 names the RoBERTa-base OpenAI detector as one of two NIST baseline discriminators and prints no result under that name (NIST AI 700-1, read 9 November 2026).
Pegoraro 2023, added 9 November 2026 as the tenth record, ran the Solaiman et al. RoBERTa detector among 20 detection methods on a ChatGPT and human question-and-answer benchmark in March 2023 and records a 7.2% true positive rate on ChatGPT responses and a 96.4% true negative rate on human responses (arXiv 2304.01487, Table I, read 9 November 2026); no ranking field takes a value from the printed rates.
Two records joined the list on 9 November 2026 and took the count to 12: Ibrahim 2023 scored 120 human ESL essays and 120 ChatGPT essays, with 16 of 120 human essays (13.3%) below a 95% originality cut under the author's inverted label "false negatives", a figure kept in prose and not entered as a rate, and 10 of 120 ChatGPT essays (8.3%) above a 5% cut, a printed miss count that replaced the 33.3% Weber-Wulff 2023 figure in the false negative field; Alshammari and Rao 2025 recorded a mean AI score of 8.15% on 49 DeepSeek-V3 answers.
Where GPT-2 Output Detector appears in the index
Rankings
- AIDI-R-01AI 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.
- AIDI-R-02AI detector false positive rates, ranked
- AIDI-R-08AI detectors ranked by published robustness to paraphrased and edited text
- AIDI-R-09AI detector false negative rates, ranked
- AIDI-R-10AI detector accuracy claims vs published results, ranked by the gap
- AIDI-R-11AI detectors ranked by named integration surfaces
- AIDI-R-17AI detectors ranked by the generator models the vendor names
- AIDI-R-18AI detectors ranked by study records on non-English and non-native text
- AIDI-R-20AI detectors ranked by launch year, with every change of hands and withdrawal the cards record
The card qualifies for two rankings. In AI detectors ranked by independent testing the GPT-2 Output Detector counted 4 study records on the morning of 9 November 2026 and 9 after the records added later that day, second only to GPTZero, and in AI detector false positive rates the card ranks on the Perkins 2024 figure of 0 false accusations in 10 human samples (detectors.json, read 9 November 2026). There is no free allowance stated in words and no language list, so the free tier and language rankings leave the card out.
Since 9 November 2026 the card also appears in false negative rate on 12 of 36 AI documents undetected in Weber-Wulff 2023, in robustness to paraphrase on the Sadasivan 2023 figure of 0.925 after one paraphrase round, and in the claim vs evidence gap on OpenAI's stated 95% against 98.5% in RAID 2024.
No audience view lists the tool: the card states no integrations, no plans and no hosted service. The AI detector register carries the card as the most-tested baseline in the index, and the OpenAI classifier record covers OpenAI's second, separate detector, the AI Text Classifier of January 2023, which was withdrawn that July.
Since 9 November 2026 the GPT-2 Output Detector is second on the false positive studies question with 8 of its 12 records printing a human-text figure, second on the non-English evidence ranking with four records (Weber-Wulff, Liang, Orenstrakh and Ibrahim), and appears on the GPT-4 question with four records.
Questions about GPT-2 Output Detector
Is the GPT-2 Output Detector still available?
The code and weights are, under the MIT License on GitHub and Hugging Face, read 9 November 2026. The hosted demo at openai-openai-detector.hf.space is paused and the repository is archived, so running the model means running it locally.
How accurate is the GPT-2 Output Detector?
OpenAI stated detection rates of about 95% on 1.5B GPT-2 text in November 2019 and called that not high enough for standalone use. In Weber-Wulff 2023 the demo scored 39 of 54 (72%); in Perkins 2024 its baseline mean accuracy was 57.2%; in RAID 2024 the large model posted a 2.91% false positive rate at threshold 0.5. All read 9 November 2026.
Can the GPT-2 Output Detector detect ChatGPT?
The model card says it is "strongly suggested not to use it as a ChatGPT detector", read 9 November 2026. Weber-Wulff 2023 found it classified 72% of a ChatGPT-heavy test set correctly while noting it was never trained for GPT-3.5.
Is the GPT-2 Output Detector the same as OpenAI's AI Text Classifier?
No. The GPT-2 Output Detector is the 2019 RoBERTa model; the AI Text Classifier was released on January 31, 2023 and withdrawn on July 20, 2023, and has its own record in the studies index, which carries both dates.
Sources and revisions
Sources
- openai.com/index/gpt-2-1-5b-releaseRead 9 November 2026
- github.com/openai/gpt-2-output-dataset/tr...Read 9 November 2026
- github.com/openai/gpt-2-output-datasetRead 9 November 2026
- huggingface.co/openai-community/roberta-b...Read 9 November 2026
- openai-openai-detector.hf.spaceRead 9 November 2026
- arxiv.org/pdf/2306.15666Read 9 November 2026
- arxiv.org/pdf/2304.02819v3Read 9 November 2026
- arxiv.org/pdf/2403.19148Read 9 November 2026
- arxiv.org/pdf/2405.07940Read 9 November 2026
Revision log
- Card created
- Sadasivan 2023, orenstrakh 2023, gao 2023, jabarian and imas 2025 and nist ai 700-1 records added, 9 study records
- Pegoraro 2023 record added to the study list in the fifth wave2026-11-09: ibrahim 2023 and alshammari and rao 2025 records added in the sixth wave, 12 study records
- False positive studies question, gpt-4 question and non-english evidence ranking linked (seventh wave)
). 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.: register plate](/img/independently-tested-480.webp)
