How it works
Lekh uses trained text classifiers and a separate reference set of human writing to look for signs of AI generation. Here is how the current research preview works and where it fails.
Models trained on examples
English uses a ModernBERT-large model trained on about 144,000 texts: human writing of many kinds, AI text from current models, and human documents with AI-written passages inside. It gives two scores: one for the text as a whole, and one for each stretch of about a dozen words. French, Spanish and Hindi use an earlier mmBERT-base model that scores the text as a whole only.
How the verdict is decided
Both English scores are compared with a separate reference set of about 16,000 human texts, grouped by kind of writing and by length. A text is flagged only if it scores higher than 99% of the human texts in every group of its length, so the target false-positive rate is 1%. That target holds for human writing that resembles the reference groups; it is not a promise for every style or population (see the limits below).
There are four results. AI-generated: most of the text reads as AI. AI-written passages found: parts read as AI and are marked. No AI signal found: this is an absence of evidence, not proof that a person wrote it. Can't tell or No verdict: the score sits between the lines, the text is under 60 words, or we have too few human reference texts of that length. Texts over 2,000 words are checked in parts, with a stricter line for each part.
Marked passages are where the per-word AI score is highest. On held-back academic test documents with known AI passages, about 9 in 10 marked characters were truly AI-written. The edges of a mark are approximate, and a single rewritten sentence is usually too little to detect.
Hidden tricks we check for
Some "humanizer" tools insert invisible characters or swap letters for look-alikes from other alphabets so detectors misread the text. We remove these before scoring and tell you we found them. For PDFs we also look for white or microscopic text, including hidden instructions aimed at AI reviewers. These findings are shown separately and never change the AI score.
Accuracy
English results for release 2.2, measured on 16 November 2026. Every rate uses the website's own decision rule and counts only texts of 60 words or more. "Final tests" were held back and opened once, after the release was frozen. "Development tests" had already been looked at while building, so they flatter the detector slightly.
| Final tests (opened once) | AI caught | Human texts wrongly flagged |
|---|---|---|
| Whole AI documents from 8 flagship models never used before | 94.9% (260 of 274) | — |
| Human documents with AI passages inside: books, journal papers, official reports | 90.0% (385 of 428) | 2.3% (10 of 428), all in central-bank reports |
| Research papers with AI-written or AI-rewritten parts | 81.0% (655 of 809) | 1.3% (12 of 897) |
| Single paragraphs of a paper rewritten by AI | 37.7% (26 of 69) | 0% (0 of 264) |
| One or two sentences rewritten by AI inside a paper | 1.4% (2 of 142) | — |
| Development tests | AI caught | Human texts wrongly flagged |
|---|---|---|
| Whole AI documents, 33 kinds of writing, 30 models | 95.0% (797 of 839) | 0.42% (20 of 4,767) |
| Research papers with an AI paragraph or section inside | 75.8% (1,020 of 1,346) | 0.65% (16 of 2,471) |
| Chat answers from 50 models | 99.2% (4,831 of 4,872) | — |
| Outside benchmark: essays, learner writing, speeches | 84.7% (593 of 700) | 1.7% (6 of 362) |
| Same benchmark, after commercial "humanizer" tools | 8% to 80% depending on the tool | — |
| French, Spanish, Hindi (earlier model) | 96.4% (270 of 280) | 0.61% (3 of 495) |
Human text that AI only proofread or lightly polished was flagged 2.1% (3 of 142) and 3.7% (11 of 296) of the time in the final tests; human text that AI substantially rewrote was flagged 63% (44 of 70). Across all final tests, 1.4% of human texts were flagged (22 of 1,589), above the 1% target; 19 of those 22 were texts over 900 words.
Known limits
- False positives happen. Formulaic official writing is the worst case we have measured: 10 of 49 central-bank regional reports were flagged. In the final tests, texts over 900 words were flagged more often (19 of 443) than shorter ones (3 of 1,146).
- Editing moves results in both directions. Copyediting, grammar tools and translation can make human text look more AI-like or AI text look more human. Proofreading by AI was almost never flagged in our tests; a substantial AI rewrite usually was.
- Small AI edits are invisible. One or two AI-written sentences inside a human text are almost never detected, and a single rewritten paragraph is caught less than half the time.
- Commercial "humanizer" tools defeat the detector often: between 8% and 80% of their output was caught, depending on the tool.
- "No AI signal found" is not a human verdict. In an outside benchmark that includes humanizer output, 1 in 5 AI texts received it.
- Short texts (under 60 words) get no verdict.
- New AI models change how AI text looks. We re-test when major models are released.
- Hindi is in beta, and Nepali and Hinglish are not supported yet. French, Spanish and Hindi do not have passage-level detection.
- Never use a detector as the only evidence. Talk to the writer, look at drafts and version history.
Single models and councils
The menu under the text box lets you choose which of our models checks the text. Every option is a model we fine-tuned ourselves, named by its code name, and runs on our own servers.
- lekh-en-passage-v22 is the current release described above and the default. It is the only option that marks AI passages inside human writing at the passage level.
- Earlier models such as lekh-en-vanguard-v21 and lekh-en-ettin-1b-v21 are offered for comparison. Each has its own calibration on human texts. They check English only; other languages always go to the multilingual model.
- Councils will combine the scores of several of our models into one result. A council is added only after it is tested at the same 1% false-positive line on texts it has never seen, and it replaces the default only if it does better than the best single model.
Your answers under each result tell us which option was right, so the menu is also how we compare models on real texts.
Privacy
We don't store the text, files or links you check. For each check we log only the model you chose, the verdict, the length, the language and the time it took, so we can see which models work. If you send feedback under a result, that text, the result and your answer are saved so we can fix mistakes. Links are fetched once, only from public websites, and the page is discarded after analysis. All models run on our own servers; no text is sent to outside AI services.