What AI Detector Checker measures
All 10 guides describe the same check. Paste at least 100 words of English and the frozen V4 engine returns an AI-writing signal score from 1 to 99, the zone it falls into (Likely human-written, Uncertain or Likely AI-written, with a Moderate or Strong signal chip in the AI zone), a score stability range, and one sentence of evidence for that length from the sealed test. Scores of 65 or more are treated as AI-written, at every text length.
What it does not return: a model name, a per-sentence verdict, a percentage of the document that is AI, or anything about who typed it. The detector never attributes text to ChatGPT, Claude, Gemini, DeepSeek, Kimi, Qwen, GLM, MiniMax, Llama, Copilot or any other product. What each guide gives you instead is the sealed-test result for texts known to come from that family: how many were caught, by version and by length, and where the engine is weakest.
The sealed results at a glance
Texts of 150 words or more, scored once on 17 October 2026. Each guide carries the versions, the counts and the Wilson 95% intervals.
| Family | Caught by V4 | Texts | Guide |
|---|---|---|---|
| ChatGPT | 100.00% | 525 | ChatGPT AI Detector |
| Claude | 99.71% | 1,048 | Claude AI Detector |
| Gemini | 100.00% | 315 | Gemini AI Detector |
| DeepSeek | 98.08% | 261 | DeepSeek AI Detector |
| Kimi | 99.04% | 209 | Kimi AI Detector |
| Qwen | 99.61% | 255 | Qwen AI Detector |
| GLM | 98.71% | 233 | GLM AI Detector |
| MiniMax | 99.40% | 333 | MiniMax AI Detector |
| Llama (Meta) | 97.27% | 990 | Llama AI Detector |
| Copilot | not tested separately | see ChatGPT and Claude | Copilot AI Detector |
Models we never trained on
Two open-weights families were kept out of training entirely so the sealed test includes a genuine unseen-model measurement. The engine caught 98.50% of 998 gpt-oss texts across all lengths (Wilson 95% interval 97.53 to 99.09), and 98.55% of 69 Gemma texts (a small sample; interval 92.24 to 99.74). V3 caught 16.53% and 27.54% of the same texts. These two families are a measurement of two families, not of every model the engine has not seen; there is no guide page for them because nothing about their writing was studied beyond these counts.
Why outputs from different models overlap
These guides describe tendencies, not fingerprints, and the tendencies overlap heavily. The current assistants are trained on overlapping data and tuned towards the same qualities, clear, balanced, evenly paced prose, so a summary from one reads much like a summary from another. The same product produces different writing depending on the instruction it was given, so a single prompt change moves the output further than a change of model would. A person editing a draft removes exactly the patterns a detector relies on, while a person writing formal, templated or translated English produces them without any assistance at all. And every model changes with each release, so a pattern that held six months ago may not hold now. That is why no page here converts a score into a product name, and why the guides are worth reading as descriptions of what was measured rather than as identification keys.
AI model detector FAQ
Can AI Detector Checker identify which AI model wrote my text?
No. AI Detector Checker measures general AI-writing signals. It does not identify ChatGPT, Claude, Gemini, DeepSeek, Kimi, Qwen, GLM, MiniMax, Llama, Copilot, a particular model family, or a model version. Similar outputs can come from different systems, and editing can obscure or introduce the same patterns.
Why provide model guides if the detector cannot identify the source?
Because the sealed test did measure, family by family and version by version, how often texts known to come from each model were caught. That is useful if you know or suspect which tool was used: it tells you how much a result can and cannot be expected to show for that family. It does not turn a score into model-specific proof.
Do the model pages use different detectors?
No. Every page describes the same V4 engine and the same check. The overall result comes from a single document-level classifier; no individual sentences are marked, and no page performs source-model attribution.
How much text should I check?
At least 100 words, and preferably 150 or more: short texts are the least reliable band at every family. A longer sample offers more structure, rhythm and variation to measure, but it still does not make the result proof of authorship.
Can human-written text be flagged as AI-like?
Yes. In the sealed test about 1 in 700 human texts of 150 words or more reached 65 or more, and about 1 in 100 at 100 to 149 words; modern formal speeches and non-native English were flagged more often than that. Treat every result as a review signal and examine the wider context.