Accuracy report · detector v2.0 · October 2026
How accurate is GPTGone?
Every number on this page comes from texts the detector never trained on, scored by the same code that runs on gptgone.com. We publish the hard cases too.
Headline (test split, 11 sources)
| GPTGone v2 | Fast model only | v1 heuristics | |
|---|---|---|---|
| AUROC | 0.990 | 0.957 | 0.670 |
| AI texts caught (balanced) | 89.6% | 74.3% | 59.3% |
| Human texts flagged (balanced) | 2.2% | 2.2% | 27.4% |
| Human texts called "AI" (balanced) | 0.5% | 0.6% | 11.1% |
| Documents | 9267 human, 15979 AI |
By source (test split)
| Source | Human | AI | AUROC | Human flagged | AI caught | v1 AUROC |
|---|---|---|---|---|---|---|
| artnitolog | 318 | 2195 | 0.998 | 0.0% | 97.0% | 0.811 |
| beemo | 567 | 669 | 0.833 | 17.1% | 65.8% | 0.553 |
| coai | 1209 | 2044 | 0.992 | 0.6% | 89.5% | 0.630 |
| eli5 | 618 | 0 | – | 0.5% | – | – |
| hape | 1035 | 6134 | 0.995 | 0.0% | 86.9% | 0.675 |
| hc3 | 670 | 872 | 0.999 | 0.3% | 97.5% | 0.783 |
| mage | 1241 | 760 | 0.991 | 0.7% | 87.4% | 0.748 |
| mildrgb | 0 | 492 | – | – | 99.6% | – |
| raid | 397 | 815 | 0.974 | 1.8% | 85.3% | 0.751 |
| rasbt | 2312 | 1998 | 0.999 | 0.3% | 97.6% | 0.897 |
| scruples | 900 | 0 | – | 0.2% | – | – |
Held-out sets (never trained or tuned on)
| Set | Human | AI | AUROC | Human flagged | AI caught | v1: human flagged / AI caught |
|---|---|---|---|---|---|---|
| textsight-2026 | 641 | 56 | 0.976 | 1.4% | 78.6% | 25.1% / 73.2% |
| frontier-abstracts | 562 | 594 | 0.990 | 1.2% | 82.8% | 82.6% / 91.2% |
| heldout-generators | 0 | 5649 | – | – | 97.0% | – / 76.2% |
| liang-2023 | 392 | 263 | 0.953 | 3.6% | 74.9% | 20.4% / 74.1% |
| claude-5-family | 0 | 150 | – | – | 74.0% | – / 21.3% |
| human-edited-ai | 0 | 613 | – | – | 40.0% | – / 40.3% |
textsight-2026: the TextSight 2026 control set of hard human texts (non-native academic writing, pre-2022 Wikipedia, public-domain literature) and 2026 models. frontier-abstracts: arXiv abstracts paraphrased by GPT-5.2, Claude Opus 4.5, Claude Sonnet 4.5 and Gemini 3 Pro.
heldout-generators: every document from Grok 4.6 and Kimi K3 (no human texts in this set). claude-5-family: Claude Opus 5.5 and Fable 5.1 were never trained on; Sonnet 5.5 and Haiku 5.5 appear here only for prompts held out of training. human-edited-ai: AI drafts a person then edited (Beemo), so "caught" is the share still flagged.
AI detection by generator (test + held-out)
| Generator | Texts | Caught | Called "AI" | v1 caught |
|---|---|---|---|---|
| allenai/tulu-2-13b | 56 | 58.9% | 33.9% | 35.7% |
| allenai/tulu-2-7b | 59 | 57.6% | 27.1% | 44.1% |
| anthropic/claude-haiku-4.5 | 516 | 97.5% | 85.1% | 91.3% |
| anthropic/claude-opus-4.5 | 145 | 68.3% | 27.6% | 89.0% |
| anthropic/claude-sonnet-4.5 | 150 | 94.0% | 68.7% | 90.7% |
| chatgpt | 105 | 92.4% | 72.4% | 81.9% |
| Claude 3 Opus | 322 | 97.8% | 95.0% | 82.3% |
| claude-fable-5.1 | 60 | 76.7% | 45.0% | 23.3% |
| claude-haiku-4-5 | 12 | 100.0% | 50.0% | 83.3% |
| claude-haiku-5.5 | 30 | 83.3% | 56.7% | 33.3% |
| claude-opus-5 | 12 | 58.3% | 16.7% | 58.3% |
| claude-opus-5.5 | 30 | 66.7% | 26.7% | 10.0% |
| claude-sonnet-5.5 | 30 | 66.7% | 36.7% | 16.7% |
| cohere | 18 | 50.0% | 16.7% | 83.3% |
| cohere-chat | 112 | 76.8% | 53.6% | 41.1% |
| Command R+ | 322 | 85.4% | 72.0% | 74.8% |
| deepseek-v4-flash-0731 | 512 | 98.0% | 85.0% | 76.0% |
| deepseek/deepseek-v4-pro | 90 | 98.9% | 96.7% | 2.2% |
| gemini-3.6-flash | 501 | 97.4% | 87.6% | 97.6% |
| GigaChat Pro | 278 | 100.0% | 97.1% | 86.7% |
| google/gemini-3-flash-preview | 522 | 99.6% | 88.3% | 94.3% |
| google/gemini-3-pro-preview | 149 | 91.9% | 61.1% | 95.3% |
| google/gemini-3.7-flash | 79 | 100.0% | 100.0% | 20.3% |
| google/gemma-2b-it | 52 | 86.5% | 55.8% | 53.8% |
| google/gemma-7b-it | 51 | 90.2% | 56.9% | 54.9% |
| gpt-3 (2023) | 172 | 81.4% | 36.6% | 75.6% |
| gpt-3.5-turbo | 389 | 91.8% | 71.5% | 65.8% |
| gpt-3.5-turbo (2023) | 872 | 97.5% | 88.6% | 59.2% |
| gpt-4 (2023) | 91 | 62.6% | 15.4% | 71.4% |
| gpt-4.1 | 12 | 58.3% | 41.7% | 91.7% |
| gpt-4o-2024-08-06 | 1035 | 98.6% | 92.3% | 58.5% |
| gpt-4o-mini-2024-07-18 | 1035 | 99.2% | 93.1% | 47.7% |
| gpt-oss-120b | 533 | 98.9% | 96.6% | 87.8% |
| gpt2 | 24 | 58.3% | 45.8% | 50.0% |
| gpt3 | 14 | 50.0% | 50.0% | 50.0% |
| gpt4 | 105 | 93.3% | 77.1% | 83.8% |
| GPT4 Omni | 321 | 100.0% | 97.8% | 84.1% |
| GPT4 Turbo 2024-04-09 | 322 | 98.8% | 86.6% | 82.6% |
| HuggingFaceH4/zephyr-7b-beta | 81 | 42.0% | 16.0% | 43.2% |
| llama-3-70B | 995 | 53.3% | 21.0% | 12.3% |
| llama-3-70B-Instruct | 1035 | 98.2% | 88.3% | 34.5% |
| llama-3-8B | 999 | 71.8% | 40.0% | 11.3% |
| llama-3-8B-Instruct | 1035 | 98.4% | 92.4% | 32.4% |
| llama-chat | 134 | 94.8% | 81.3% | 75.4% |
| Llama3 70B | 322 | 99.4% | 96.9% | 71.4% |
| luna-medium | 336 | 97.0% | 93.5% | 86.9% |
| meta-llama/Llama-2-13b-chat-hf | 74 | 75.7% | 41.9% | 20.3% |
| meta-llama/Llama-2-70b-chat-hf | 77 | 66.2% | 40.3% | 33.8% |
| meta-llama/Llama-2-7b-chat-hf | 69 | 75.4% | 46.4% | 30.4% |
| mistral | 28 | 46.4% | 25.0% | 53.6% |
| mistral-chat | 140 | 94.3% | 79.3% | 87.1% |
| mistralai/Mistral-7B-Instruct-v0.1 | 74 | 56.8% | 25.7% | 52.7% |
| mistralai/Mixtral-8x7B-Instruct-v0.1 | 76 | 61.8% | 31.6% | 47.4% |
| moonshotai/kimi-k3 | 4821 | 96.6% | 86.2% | 88.3% |
| mpt | 18 | 50.0% | 50.0% | 33.3% |
| mpt-chat | 117 | 88.0% | 63.2% | 70.9% |
| nvidia/nemotron-3.5-lightning | 82 | 98.8% | 97.6% | 2.4% |
| openai/gpt-5-nano | 485 | 60.0% | 24.1% | 90.9% |
| openai/gpt-5.2-chat | 150 | 76.7% | 40.0% | 90.0% |
| openai/gpt-5.6-luna-pro | 82 | 100.0% | 96.3% | 29.3% |
| qwen/qwen3.8-max | 83 | 100.0% | 97.6% | 0.0% |
| qwen3.6:35b | 495 | 98.8% | 89.7% | 97.2% |
| sol-medium | 154 | 94.8% | 88.3% | 84.4% |
| text-davinci-003 | 371 | 82.7% | 51.5% | 55.5% |
| x-ai/grok-4.6 | 828 | 99.6% | 98.2% | 5.8% |
| YandexGPT 3 Pro | 308 | 98.4% | 95.1% | 72.7% |
| z-ai/glm-5.3 | 76 | 100.0% | 98.7% | 1.3% |
Human false positives by source (test + held-out)
| Human source / genre | Texts | Flagged | Called "AI" | v1 flagged |
|---|---|---|---|---|
| artnitolog / essay | 110 | 0.0% | 0.0% | 83.6% |
| artnitolog / reuters | 103 | 0.0% | 0.0% | 22.3% |
| artnitolog / wp | 105 | 0.0% | 0.0% | 4.8% |
| beemo / Generation | 165 | 18.8% | 7.3% | 13.3% |
| beemo / Open QA | 129 | 13.2% | 0.8% | 33.3% |
| beemo / Rewrite | 167 | 22.2% | 4.8% | 24.6% |
| beemo / Summarize | 100 | 12.0% | 1.0% | 53.0% |
| coai / arxiv-abstract | 1209 | 0.6% | 0.1% | 80.6% |
| coai-bench / arxiv-abstract | 562 | 1.2% | 0.2% | 82.6% |
| eli5 / eli5 | 618 | 0.5% | 0.3% | 13.6% |
| hape / acad | 151 | 0.0% | 0.0% | 30.5% |
| hape / blog | 193 | 0.0% | 0.0% | 2.1% |
| hape / fic | 182 | 0.0% | 0.0% | 11.5% |
| hape / news | 168 | 0.0% | 0.0% | 8.9% |
| hape / spok | 190 | 0.0% | 0.0% | 0.0% |
| hape / tvm | 151 | 0.0% | 0.0% | 0.0% |
| hc3 / finance | 179 | 0.6% | 0.0% | 8.9% |
| hc3 / medicine | 96 | 0.0% | 0.0% | 18.8% |
| hc3 / open_qa | 18 | 0.0% | 0.0% | 38.9% |
| hc3 / reddit_eli5 | 292 | 0.3% | 0.0% | 11.0% |
| hc3 / wiki_csai | 85 | 0.0% | 0.0% | 58.8% |
| liang / college-essay | 70 | 2.9% | 1.4% | 0.0% |
| liang / cs224n-abstract | 145 | 1.4% | 0.0% | 51.0% |
| liang / toefl-esl | 89 | 11.2% | 2.2% | 5.6% |
| liang / us-8th-grade | 88 | 0.0% | 0.0% | 1.1% |
| mage / cmv | 146 | 0.0% | 0.0% | 2.7% |
| mage / eli5 | 128 | 0.8% | 0.0% | 3.1% |
| mage / roct | 99 | 1.0% | 0.0% | 15.2% |
| mage / sci | 148 | 0.0% | 0.0% | 81.1% |
| mage / squad | 147 | 0.0% | 0.0% | 46.3% |
| mage / tldr | 137 | 3.6% | 0.0% | 62.8% |
| mage / wp | 150 | 0.7% | 0.7% | 0.7% |
| mage / xsum | 147 | 0.7% | 0.0% | 18.4% |
| mage / yelp | 125 | 0.0% | 0.0% | 0.0% |
| raid / abstracts | 141 | 2.8% | 0.0% | 46.1% |
| raid / books | 105 | 1.0% | 0.0% | 55.2% |
| raid / news | 151 | 1.3% | 0.0% | 16.6% |
| rasbt / arxiv | 305 | 0.3% | 0.0% | 33.1% |
| rasbt / arxiv-preprints | 48 | 0.0% | 0.0% | 85.4% |
| rasbt / gutenberg | 305 | 0.0% | 0.0% | 8.5% |
| rasbt / openstax | 318 | 0.3% | 0.0% | 38.7% |
| rasbt / personal-blog | 21 | 0.0% | 0.0% | 0.0% |
| rasbt / plos | 322 | 0.0% | 0.0% | 28.0% |
| rasbt / pmc | 308 | 0.3% | 0.0% | 19.5% |
| rasbt / stackexchange | 327 | 1.5% | 0.6% | 9.2% |
| rasbt / wikimedia | 330 | 0.0% | 0.0% | 19.4% |
| scruples / aita | 900 | 0.2% | 0.1% | 0.0% |
| textsight / human-pmc-esl | 180 | 3.9% | 1.1% | 27.2% |
| textsight / human-wikipedia-pre2022 | 155 | 0.0% | 0.0% | 55.5% |
| textsight / pd-literature | 306 | 0.7% | 0.3% | 8.5% |
Robustness to evasion attacks (RAID, test split)
RAID applies common detector-evasion tricks to AI and human text. "Caught" is the share of attacked AI texts still flagged; "flagged" the share of attacked human texts wrongly flagged.
| Attack | AI texts | Caught | v1 caught | Human texts | Human flagged |
|---|---|---|---|---|---|
| alternative spelling | 102 | 92.2% | 83.3% | 53 | 1.9% |
| article deletion | 123 | 89.4% | 82.9% | 47 | 2.1% |
| homoglyph | 85 | 88.2% | 0.0% | 53 | 0.0% |
| insert paragraphs | 77 | 90.9% | 85.7% | 39 | 2.6% |
| none | 40 | 87.5% | 92.5% | 0 | – |
| number | 42 | 85.7% | 64.3% | 26 | 3.8% |
| paraphrase | 100 | 61.0% | 74.0% | 53 | 1.9% |
| perplexity misspelling | 69 | 94.2% | 79.7% | 40 | 0.0% |
| synonym | 96 | 86.5% | 81.3% | 53 | 3.8% |
| upper lower | 10 | 80.0% | 80.0% | 0 | – |
| whitespace | 4 | 75.0% | 50.0% | 13 | 0.0% |
| zero width space | 67 | 82.1% | 70.1% | 20 | 0.0% |
By length (test split)
| Words | Docs | AUROC | Human flagged | AI caught |
|---|---|---|---|---|
| 40–100 | 3485 | 0.973 | 1.8% | 76.7% |
| 100–200 | 5901 | 0.985 | 2.2% | 87.0% |
| 200–400 | 5710 | 0.992 | 1.7% | 91.4% |
| 400– | 10150 | 0.996 | 0.4% | 92.9% |
Mixed documents (synthetic: half human, half AI)
960 test documents, each a human passage and an AI passage from the same source and genre in random order.
| Verdict "Mixed" | Verdict "AI" | Verdict "Human" | Sentence accuracy | Sentence AUROC | AI-share error (mean abs.) |
|---|---|---|---|---|---|
| 46.9% | 7.0% | 46.1% | 84.8% | 0.908 | 14.2 pts |
Bias check: human student essays (Liang et al., 2023)
Essays from Liang et al., "GPT detectors are biased against non-native English writers", which found that popular detectors flagged most TOEFL essays by non-native writers as AI.
Share of genuinely human essays flagged. Commercial detector figures are the 2023 scores published with the paper (score ≥ 0.5); those products have changed since, so read them as context, not a current ranking.
| Essays | Texts | GPTGone v2 | GPTZero | Originality.ai | ZeroGPT | Crossplag | Sapling | OpenAI (HF) | Quillbot |
|---|---|---|---|---|---|---|---|---|---|
| TOEFL essays (non-native) | 89 | 11.2% | 51.6% | 75.8% | 52.7% | 51.6% | 68.1% | 58.2% | 74.7% |
| US 8th-grade essays | 88 | 0.0% | 0.0% | 1.1% | 1.1% | 12.5% | 4.5% | 9.1% | 9.1% |
| College admission essays | 70 | 2.9% | 1.4% | 4.3% | 2.9% | 0.0% | 1.4% | 0.0% | 1.4% |
Method
Data. 173,690 labelled documents from 16 sources: 118,239 for training (56,741 human, 61,498 AI, from 55 AI models), 17,109 for calibration, and the rest for this report. A human text and the AI texts written for the same prompt always land in the same split, so no test text has a sibling in training.
Models. A 17M-parameter transformer encoder (Ettin, ModernBERT architecture) fine-tuned on ~120-word windows, plus a logistic regression over 39 writing signals and a 6,000-entry vocabulary. Both run in the browser; the neural model is quantised to 8-bit (~17 MB).
Decisions. Window scores are fused, averaged per sentence and per document, and calibrated on the validation split. The "AI" verdict needs a document score that flags about 1% of human validation texts (averaged across sources); "AI or mixed" about 3%. Thresholds were fixed before the test data was scored.
How to read this
- No detector is proof. These rates are averages over sources; a particular writer or model can do better or worse.
- The test split shares sources (not documents) with training, so it is the optimistic number. The held-out sets are the realistic one, especially for brand-new models.
- Text that a person rewrote from an AI draft, or AI text run through a paraphraser, is the hardest case for every detector, ours included.
- The texts in this report are up to about 1,900 words long. You can scan far longer texts; beyond 2,000 words the detector's calibration stays at its 2,000-word setting rather than being extrapolated, but we haven't measured accuracy on texts that long.