Fast-DetectGPT zero-shot AI detector
Vendor Bao, Zhao, Teng, Yang and Zhang (Westlake University, Zhejiang University, Shanghai Polytechnic University and Nanyang Technological University) Official site github.com/baoguangsheng/fast-detect-gpt How this card is sourced
Fast-DetectGPT flagged 23.2% of RAID's human-written documents as machine-generated at a naive threshold of 0.5, falling to 1.70% at 0.95, in Dugan et al. 2024 (arXiv 2405.07940, Table 4, read 10 November 2026). Fast-DetectGPT is a research method with published code, not a company's product: the ICLR 2024 paper by Guangsheng Bao, Yanbin Zhao, Zhiyang Teng, Linyi Yang and Yue Zhang replaces DetectGPT's perturbation step with a sampling step that reads what the authors call conditional probability curvature, and the implementation is on GitHub under the MIT License (arXiv 2310.05130 and github.com/baoguangsheng/fast-detect-gpt, read 10 November 2026). One hosted service exists: the repository's homepage, fastdetect.net, runs a web demo and an API on credits, carries the footer YuanNan FastDetect with a Zhejiang ICP filing, and publishes no price outside a login. The card joins the detector index as its third open-source research tool after GLTR and DetectGPT, and how this card is sourced explains why the paper and the repository stand in for a product page.
What Fast-DetectGPT detects
| Field | Value | Source |
|---|---|---|
| Content types detected | Text | arxiv.org/abs/2310.05130 |
| Generators covered | ChatGPT, GPT-4, GPT-3 (paper experiments), open models used as sources in the paper: GPT-2, OPT-2.7B, GPT-Neo-2.7B, GPT-J 6B, GPT-NeoX 20B | arxiv.org/pdf/2310.05130 |
| Input limits | Research code: the local demo (python scripts/local_infer.py) takes pasted text with gpt-neo-2.7B as the default sampling and scoring models, or gpt-j-6B as the sampling model; the paper's experiments ran on one Tesla A100 with 80 GB of memory. The hosted demo at fastdetect.net accepts pasted text or a .txt, .pdf or .docx file of up to 10 MB, offers sampling and scoring model choices (gpt-neo-2.7b; falcon-7b with falcon-7b-instruct; llama3-8b with llama3-8b-instruct; a fast-detect-pro flagship), meters guest trials by IP and credits per account, and its terms state that single detection text length does not exceed specified limits and that each account has daily and monthly call limits, with no figures published. | github.com/baoguangsheng/fast-detect-gpt |
| Free tier | Free and open source (MIT License): the code and the paper's GPT-3, ChatGPT and GPT-4 generations are on GitHub. The hosted demo at fastdetect.net offers guest trials and free credits on registration (Register now to get free credits!) with no published allowance; a Wallet page sells further detection credits | github.com/baoguangsheng/fast-detect-gpt |
| API | Yes, fastdetect.net/#/apidoc | fastdetect.net/#/apidoc |
| Integrations | GitHub repository with local demo script and experiment scripts, Hosted web demo at fastdetect.net (text or file upload), REST API with API keys from the fastdetect.net dashboard, Glimpse extension for proprietary scoring models (GPT-3.5 and others), a separate repository | github.com/baoguangsheng/fast-detect-gpt |
| Data retention | The paper and repository publish no privacy policy. Hosted demo terms of use (fastdetect.net, read 10 November 2026): Detection text is only temporarily stored during processing and deleted immediately after completion; All transmitted data is protected by SSL/TLS encryption; Do not share user data with third parties except as required by law; the terms are subject to the laws and regulations of the People's Republic of China. | fastdetect.net |
| Launch year | 2023 | github.com/baoguangsheng/fast-detect-gpt |
| Ownership | A research method published at ICLR 2024 by Guangsheng Bao (Zhejiang University and School of Engineering, Westlake University), Yanbin Zhao (Shanghai Polytechnic University), Zhiyang Teng (Nanyang Technological University), Linyi Yang and Yue Zhang (Westlake University), arXiv 2310.05130, first posted 8 October 2023 and revised to v3 on 16 December 2024. The code is on GitHub under the MIT License (copyright 2023 Bao Guangsheng); the repository was created on 7 July 2023, last pushed on 7 February 2026 and held 439 stars on 10 November 2026. A hosted demo and API at fastdetect.net (the repository's homepage) carries the footer YuanNan FastDetect All Rights Reserved with a Zhejiang ICP filing (浙ICP备2025159537号-2) and terms subject to the laws of the People's Republic of China; the README dates its recovery to 13 September 2025. | arxiv.org/abs/2310.05130 |
| Independent studies | 1 record | Studies index |
Fast-DetectGPT analyses text with two small open models. The repository's README states that it "requires no training and directly uses open-source small language models to detect text content generated by various large language models such as ChatGPT and GPT-4", and the local demo runs with gpt-neo-2.7B as the default sampling and scoring models or gpt-j-6B as the sampling model (README, read 10 November 2026). The paper's abstract states the method's claim in one sentence: Fast-DetectGPT "not only surpasses DetectGPT by a relative around 75% in both the white-box and black-box settings but also accelerates the detection process by a factor of 340" (arXiv 2310.05130, read 10 November 2026). The white-box setting scores a passage with the model suspected of writing it; the black-box setting scores it with a surrogate, which is how the method reaches ChatGPT and GPT-4.
The paper's generators are GPT-2, OPT-2.7B, GPT-Neo-2.7B, GPT-J 6B and GPT-NeoX 20B in the white-box experiments, and GPT-3, ChatGPT and GPT-4 in the black-box ones, on XSum news, SQuAD Wikipedia contexts, WritingPrompts stories, PubMedQA and WMT16 English and German (arXiv 2310.05130, Section 3.1 and Tables 2 and 3, read 10 November 2026). Every dataset is English except the WMT16 German experiment, so the languages field is empty. The README adds a 2026 note that Llama3-8B and Llama3-8B-Instruct as the sampling and scoring models "substantially outperforms falcon-7b/falcon-7b-instruct, especially on LRM-generated texts", and the hosted demo offers gpt-neo-2.7b, falcon-7b, llama3-8b and a fast-detect-pro model as choices (fastdetect.net, read 10 November 2026).
The repository was created on 7 July 2023, last pushed on 7 February 2026 with the commit "Update README.md", and held 439 stars on 10 November 2026 (GitHub API, read 10 November 2026). Two cards share its lineage: DetectGPT, the 2023 Stanford method whose code Fast-DetectGPT borrows from and whose 100 perturbations it replaces, and GLTR, the 2019 open-source display that RAID 2024 groups with Fast-DetectGPT as a metric-based detector.
Limits and plans
The research code publishes no character limit, no plan and no price; its cost is compute, and the paper's experiments ran on one Tesla A100 with 80 GB of memory (README, read 10 November 2026). The hosted demo at fastdetect.net accepts pasted text or a .txt, .pdf or .docx file of up to 10 MB, meters guest trials by IP address and credits per registered account, and tells a user who runs out to "visit the Wallet page to purchase more detection credits" (fastdetect.net, read 10 November 2026). Its pricing route carries the plan labels Basic, Pro, Enterprise and Free Forever with a requests per month unit, but the amounts sit behind a login and no figure is published, so the plans table on this card is empty and the API price field is null. The demo's API documentation route issues keys from its dashboard, which is why the API field is marked available with the demo as its documentation.
The demo's terms of use, read 10 November 2026 on fastdetect.net, state that "Detection text is only temporarily stored during processing and deleted immediately after completion", that each account has daily and monthly call limits and that single detection text length does not exceed specified limits, without figures, and that the terms are subject to the laws and regulations of the People's Republic of China. The paper and repository publish no privacy policy, so the retention field takes the demo's deletion clause as a stated zero and says where it came from.
What the vendor claims
No accuracy percentage is published. Paper abstract: Fast-DetectGPT not only surpasses DetectGPT by a relative around 75% in both the white-box and black-box settings but also accelerates the detection process by a factor of 340. Table 1: AUROC 0.9887 on five-model generations (white-box) and 0.9338 on ChatGPT/GPT-4 generations (black-box), against DetectGPT's 0.9554 and 0.7225. Section 3.4: when applied to ChatGPT-generated content, Fast-DetectGPT achieves a recall of 87% for machine-generated texts with only 1% misclassification of human-written text as machine-generated; at 10% false positives the recall reaches 98%. Demo page: it outperforms the commercial GPTZero on detection of ChatGPT and GPT-4 generations.
The paper reports AUROC, a threshold-free measure, so there is no accuracy percentage for the claim gap ranking to pair. Table 1 gives 0.9887 on five-model generations in the white-box setting and 0.9338 on ChatGPT and GPT-4 generations in the black-box setting, against DetectGPT's 0.9554 and 0.7225, with a 340x speedup measured on XSum (arXiv 2310.05130, Table 1, read 10 November 2026). Section 3.4 reads the ROC curve at an operating point: "when applied to ChatGPT-generated content, Fast-DetectGPT achieves a recall of 87% for machine-generated texts with only 1% misclassification of human-written text as machine-generated", and 98% recall at 10% false positives. The paper also states that Fast-DetectGPT "achieves overall higher accuracy" than the supervised RoBERTa detectors and GPTZero on its ChatGPT and GPT-4 sets (Table 3), and the demo page repeats that it "outperforms the commercial GPTZero on detection of ChatGPT and GPT-4 generations".
On rewritten text the authors ran their own attacks. Appendix F.2 paraphrases each sentence of ChatGPT passages with a T5 paraphraser and swaps adjacent words in long sentences, and reports Fast-DetectGPT (GPT-J/Neo-2.7) at 0.8669 average AUROC under paraphrase and 0.8569 under decoherence, against DetectGPT (T5-11B/Neo-2.7) at 0.5100 and 0.5135 (arXiv 2310.05130, Table 11, read 10 November 2026). The limitations section names the method's boundary: its design "leans on pre-trained models to span a multitude of domains and languages. This presents a challenge in a black-box setting, as no single model can [...] span all linguistic territories and domains."
What independent studies found
- AIDI-S-0006RAID 20242024
Run as F-DetectGPT with the repository defaults (GPT-Neo-2.7B scoring, GPT-J reference): false positive rate on RAID human-written documents of 47.3% at threshold 0.25, 23.2% at 0.5, 13.1% at 0.75 and 1.70% at 0.95 (Table 4); accuracy at 5% FPR of 98.6 on open chat models with greedy decoding, 96.0 on closed chat models and 0.6 on open non-chat models under sampling with a repetition penalty (Table 5), 73.6 overall on non-adversarial text (Table 15); 71.8 under paraphrase, 34.0 under synonym swap and 51.4 under homoglyphs against 73.6 with no attack (Table 16); 51.1 on Python code, 55.2 on Czech news and 75.5 on German news (Table 7); the paper names it among the most accurate detectors that can still deteriorate from perfect accuracy to complete failure when the generator or decoding changes (Finding 4), with the detector plateauing at 0.88% as the lowest reachable false positive rate (Finding 2).
One record in the studies index scored Fast-DetectGPT. RAID 2024 ran it as F-DetectGPT with the repository defaults, GPT-Neo-2.7B for scoring and GPT-J as the reference model, and reports false positive rates on human-written documents of 47.3% at threshold 0.25, 23.2% at 0.5, 13.1% at 0.75 and 1.70% at 0.95 (arXiv 2405.07940, Table 4, read 10 November 2026). At a threshold calibrated to 5% false positives it scored 98.6 on open chat models with greedy decoding, 96.0 on closed chat models and 0.6 on open non-chat models under sampling with a repetition penalty (Table 5), 73.6 overall on non-adversarial text (Table 15), and 51.1 on Python code, 55.2 on Czech news and 75.5 on German news (Table 7).
Under attack the same record prints 71.8 for paraphrase against 73.6 with none, 34.0 under synonym swap and 51.4 under homoglyphs (Table 16, read 10 November 2026). The robustness field takes the paraphrase column, the attack the ranking sorts on, and quotes the synonym and homoglyph figures beside it. The authors name Fast-DetectGPT among "the most accurate detectors" that nonetheless "sometimes deteriorate from perfect accuracy to complete failure" when the generator, decoding strategy or a repetition penalty changes (Finding 4), and record that the detector plateaued at 0.88% as the lowest false positive rate it could reach (Finding 2). The RAID authors' appendix calls the method "an improvement on the original DetectGPT, speeding up inference by 340x without any reduction in accuracy".
Where Fast-DetectGPT 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-06AI detectors ranked by stated data retention
- AIDI-R-08AI detectors ranked by published robustness to paraphrased and edited text
- AIDI-R-11AI detectors ranked by named integration surfaces
- AIDI-R-12AI detectors ranked by the upload formats the vendor states
- AIDI-R-20AI detectors ranked by launch year, with every change of hands and withdrawal the cards record
Fast-DetectGPT qualifies for three study rankings, read 10 November 2026 in the data file: independently tested detectors with 1 record, false positive rates at 23.2% of RAID's human documents at threshold 0.5, and adversarial robustness at 71.8 under paraphrase. On stated values it ranks in data retention at zero days on the demo's deletion clause, in accepted file types on the demo's three formats and in platform integrations on the API. It is absent from the free tier, language and claim gap rankings, which need a word allowance, a language count and an accuracy percentage the method does not publish.
The card appears on the open-source view with the MIT License and the last push of 7 February 2026, on detectors with an API, and in the question on how AI detectors work, which names it with GLTR, Binoculars and LLMDet as a metric-based detector in RAID 2024's grouping. Every detector card on the register shows the same fourteen fields, so the empty price and language cells here can be read beside the commercial tools that fill them.
Questions about Fast-DetectGPT
Is Fast-DetectGPT a tool I can use?
Yes, two ways: the MIT-licensed code runs locally with gpt-neo-2.7B or gpt-j-6B, and the hosted demo at fastdetect.net runs text or a .txt, .pdf or .docx file of up to 10 MB on guest trials and account credits, read 10 November 2026. No price is published outside the demo's login.
How accurate is Fast-DetectGPT?
The paper reports AUROC, not accuracy: 0.9887 on five open models and 0.9338 on ChatGPT and GPT-4 generations, with 87% recall at 1% false positives on ChatGPT text, read 10 November 2026. RAID 2024 measured 73.6 accuracy at a fixed 5% false positive rate on non-adversarial text.
What false positive rate did RAID find for Fast-DetectGPT?
23.2% of human-written documents at a threshold of 0.5, 47.3% at 0.25 and 1.70% at 0.95, in Dugan et al. 2024 Table 4, read 10 November 2026; the paper found the detector could not reach a false positive rate below 0.88%.
Does Fast-DetectGPT survive paraphrasing?
Better than most in the one indexed record: RAID 2024 reports 71.8 accuracy at 5% false positives under paraphrase against 73.6 with no attack, but 34.0 under synonym swap and 51.4 under homoglyphs, read 10 November 2026. The authors' own T5 paraphrase test kept AUROC at 0.8669.
Does Fast-DetectGPT store the text you check?
The research code keeps text on the machine that runs it, and the fastdetect.net demo's terms state that detection text is only temporarily stored during processing and deleted immediately after completion, read 10 November 2026. No separate privacy policy is published.
Sources and revisions
Sources
- arxiv.org/abs/2310.05130Read 10 November 2026
- arxiv.org/pdf/2310.05130Read 10 November 2026
- github.com/baoguangsheng/fast-detect-gptRead 10 November 2026
- fastdetect.netRead 10 November 2026
- openreview.net/forumRead 8 November 2026
- arxiv.org/pdf/2405.07940Read 10 November 2026
). 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)
