Honest numbers, your texts stay yours
Nothing here is invented for marketing: every figure below comes from the results section and the open technical report.
Your texts never leave your computer
The analysis runs on your PC and, after the first launch, works offline. The internet is needed only for model files, updates and licence-key checks.
Open data and code
Data, tests and limitations are described in the technical report, and the code is on GitHub. You can check everything yourself.
We show the weak spots too
The drop on short MAGE texts (−6.0 pts) and on StealthGPT (−2.5 pts) is published above. No detector is perfect: false alarms are 0–2%.
Clear payment and refunds
Payments go through Polar. Personal and Team keys: full refund within 14 days, no questions asked.
History stays on your computer
Checks, folders and texts are stored only on your computer, with automatic clean-up. The free version needs no key at all; a paid key, once activated, also works without internet.
progress: 1.1 → 1.3.1
Share of texts flagged as AI. For AI-written texts higher is better (more caught); for human texts lower is better (fewer false alarms). Measured in Sensitive mode — the “AI / not AI” line is drawn so that about 1% of ordinary human texts are flagged (on our test sets 0.2–2.0%, higher on non-native writing). Exact figures are in the tables below.
| Test set | 1.1 | 1.3 | 1.3.1 | change vs 1.3 |
|---|---|---|---|---|
| AI-written texts — caught, % (higher is better) | ||||
| HumanizerBench: 16 rewriting tools, 1,675 texts (kept only for testing, never used in training) | 69.9 | 80.8 | 81.0 | +0.2 |
| AI Humanizer Benchmark, Oct 2026: 363 texts (separate check set, not used in training) | 81.8 | 81.3 | 84.0 | +2.7 |
| Three unseen generators: 889 texts | 93.0 | 94.6 | 95.5 | +0.9 |
| Chicago Booth, plain AI | 99.7 | 99.8 | 99.8 | 0.0 |
| Chicago Booth, after StealthGPT | 82.9 | 86.6 | 88.0 | +1.4 |
| MAGE: short texts, old small models | 44.5 | 45.9 | 46.6 | +0.7 |
| Human-written texts — wrongly flagged, % (lower is better) | ||||
| TOEFL exam essays (non-native), 91 | 4.4 | 6.6 | 5.5 | −1.1 |
| Chicago humans: news, books, reviews | 0.2 | 0.1 | 0.2 | +0.1 |
| School ELL essays (ELLIPSE), 1,500 | 1.1 | 0.5 | 0.4 | −0.1 |
| Adult ESL (PELIC), 1,000 | 0.7 | 0.4 | 0.4 | 0.0 |
| Adult ESL (W&I;), 280 | 1.8 | 0.4 | 0.4 | 0.0 |
| MAGE humans, 1,400 | 1.1 | 0.9 | 2.0 | +1.1 |
Same rewritten texts, different detectors
HumanizerBench gave every detector the same rewritten texts. Share flagged as AI — higher is better.
Competitor verdicts as published by the benchmark operators; their thresholds are unknown, so false-positive rates are not comparable.
| Chicago Booth test, strict setting (1% false alarms) | plain AI text | after a rewriting tool |
|---|---|---|
| Pangram | 99.9 | 98.1 |
| Linda Assay 1.3.1 | 99.8 | 88.0 |
| Linda Assay 1.3 | 99.8 | 86.6 |
| Linda Assay 1.1 | 99.7 | 82.9 |
| GPTZero | 98.6 | 44.3 |
| Originality.ai | 94.2 | 29.1 |
Competitor numbers come from the benchmark authors’ repository (2025 runs). Pangram is clearly ahead on humanizer text; we do not claim to beat it. This is our own evaluation, not an independent audit. Both humanizer benchmarks are run by humanizer vendors or rankers, who publish all texts and verdicts. Details and caveats: technical report.
Russian and Polish: what changed in 1.3.1
Russian and Polish now run on the same models as English. On our held-out test sets (350 AI + 350 human texts per language) the verdict “AI” catches about 79% (Polish) and 80% (Russian) of AI texts at 1.4% and 0.6% false alarms (version 1.1: 39% and 57%; Polish false alarms were 0.9% in 1.3). Texts rewritten by another model are caught less often (Gemini-rewritten: Polish 56%, Russian 59%). The sets are small, so expect ±5 points. We did not compare Russian or Polish against other detectors.
Linda Loupe next to Linda Assay
The same test sets and the same rule for the verdict.
| Test set | Linda Assay | Linda Loupe |
|---|---|---|
| HumanizerBench: 16 rewriting tools, 1,675 texts (kept only for testing, never used in training) | 81.0 | 69.4 |
| AI Humanizer Benchmark, Oct 2026: 363 texts (separate check set, not used in training) | 84.0 | 75.8 |
| Three unseen generators: 889 texts | 95.5 | 83.0 |
| Polish AI texts found | 79.4 | 72.2 |
| Russian AI texts found | 80.3 | 77.3 |
| Human-written texts — wrongly flagged, % (lower is better) | ||
| TOEFL exam essays (non-native), 91 | 5.5 | 2.2 |
| MAGE humans, 1,400 | 2.0 | 0.4 |
| Chicago humans: news, books, reviews | 0.2 | 0.0 |
Linda Loupe is a smaller model: it finds fewer rewritten texts, and much fewer if the text was deliberately disguised by a stealth rewriter; false alarms stay low (0.1 to 2% on our human test sets). Own measurement, not an independent audit.
How fast is it
The whole check on one text in the Windows app. Test computer: Ryzen 7 7800X3D processor, Radeon RX 9070 XT graphics. Older laptops are slower.
1,000 words
3,000 words
0.25 s
0.76 s
2.7 s
8.4 s
5.7 s
20 s
CPU: 6 threads. Model loading at start takes 5–10 s. The GPU runs the models in half precision, so scores differ slightly from the CPU (about 0.1 on a scale where the AI line is near 0); verdicts matched on our checks. Measured on version 1.3; 1.3.1 has the same architecture.
news & updates
Release notes and model updates.
Language-aware text cleaning, more accurate sentence splitting and clear warnings for unsupported languages. Model package label 1.3.4.1; weights unchanged, no accuracy improvement is claimed.
Full changelog · 2.0.7.0