AI tool registry
AI tools are bought on the strength of a sentence on a website: “99 % accuracy”, “eight times faster”. What was measured, on which texts and in which version is usually not said, and such a figure cannot be argued with. The registry exists so that a buyer has a simple sign: the vendor explains how the tool works, or does not.
What this is
A list of AI tools with cards. Each card has four things: which model version, by which rule a decision is made, on which languages and text lengths it works, and on which sample accuracy was measured.
If a vendor does not publish one of these, the card says so. Order is alphabetical: position in the list means nothing.
The Russian section of this registry also covers thirteen systems used in hiring decisions — those cards are published in Russian only, because the products, their methods and the procurement around them are Russian. Русская часть реестра.
Six AI-text detectors
Checked 02.11.2026 from public pages. Where a vendor publishes a method, we say what it contains; where it publishes a figure, we say on what sample.
Pangram
Method. Published in full: model card v4.1 of 19.10.2026 with error tables by language and confidence intervals. Minimum 50 words, 24 languages. There is no numeric threshold; instead a target false-positive rate, about one in ten thousand texts.
Accuracy. False positives 0.0090 % (70 of 781,981 English texts), misses 0.3652 % (1,899 of 520,000 generations from 26 models).
Russian. In the language list; false-positive rate for Russian 0.0052 %. Checked 02.11.2026. The vendor’s model card →
Copyleaks
Method. Published in full: Testing Methodology V11, test of 10.09.2026, published 28.09.2026 — datasets, error matrix and three sensitivity levels with their false-positive and miss rates. Minimum 350 characters, over 30 languages.
Accuracy. On 500,000 texts: correctly detected 0.989, correctly identified human 0.999; separately, 326,120 human and 128,225 machine texts at accuracy above 0.99.
Russian. Listed among languages; the published tests cover English only, and the method says so itself. Checked 02.11.2026. The vendor’s methodology →
GPTZero
Method. Partially disclosed: how their tests are built is described, there is a length requirement (minimum 250 characters) and version numbers; the model itself and the threshold are not disclosed.
Accuracy. On the external RAID test — 95.7 % of machine texts at 1 % false positives. Own figures: precision 99.93 %, recall 99.60 %. The “99.9 %” on the front page is given without a sample.
Russian. Not declared. Checked 02.11.2026.
Originality.ai
Method. Partially disclosed: the design and version history with figures are described (5.0, October 2026); there is no separate methodology document and no threshold. Recommends from 100 words, 31 languages.
Accuracy. “99 %+” without a sample. The false-positive rate differs between places: 0.03 % on the front page and about 1.56 % in a help article — we give both.
Russian. Declared, named explicitly. Checked 02.11.2026.
ZeroGPT
Method. Partially disclosed: DeepAnalyse is described, with an honest note that weights and implementation details are not disclosed; the model number and threshold are not published, limit 15,000 characters.
Accuracy. 98.4 % and under 1 % false positives — without a sample and without naming a language.
Russian. Not declared. Checked 02.11.2026.
Tekstovod
Method. Not published: colour indication and a breakdown by fragment; no version, no threshold, no probability calculation, language list not published.
Accuracy. 85 % “on a large sample” — the size and composition of the sample are not described.
Russian. The service’s main language; no formal statement. Checked 02.11.2026. ⚠ The site did not open that day — the domain did not respond; data taken from an archived snapshot of 04.10.2026.
What is not in the registry, and why
- no quality grades or rankings: “better” or “worse” is written nowhere here;
- no paid places, partner cards or priority order;
- no hidden rules: there is one criterion and it is published above;
- no names of people — only companies and tools.
The registry is assembled from public data, without prior consent of the companies. If a vendor objects to inclusion, write to us and we will remove the card. Formal letters about inclusion are sent after state registration.
All the Institute’s registries — on the Registries page.