Since June 2025 I have been building my own detector of AI-written texts, Genuino, and humanization tools on top of it. [5] The first recorded validation, of version 4.6, gave 72.5% accuracy on 40 texts.
Most of the work went into data. The older set is 18,320 pairs of human and AI texts. [5] On top of that I wrote a system that used 940 keywords to download 4,961 articles from The Guardian, and rewrote 4,499 of them with OpenAI, Anthropic and Google models, in a neutral variant and in one imitating a human. I generated hard cases separately: an AI text went into the database only if it fooled the detector.
These data gave 58 stylometric features: vocabulary, syntax, rhythm, coherence, its own predictability computed against a corpus of about 6 million words of human texts, and a dictionary of 1,443 phrases typical of AI, chosen from more than 8 million. [5] I tested new features on 2,500 pairs. The strongest one, the density of AI collocations, reached an AUC of 0.91, and I rejected four others as noise.
The model is LightGBM trained in a loop: errors went back into training, and part of the data never touched it. On such a clean set of 632 records, version 9.3 reached an F1 of 0.965. [5] On the training data the models showed even 0.9989, so only the result on the clean set counts. I also tested robustness to attacks: from 94 to 98% of texts detected across four profiles.
I checked the results with Originality.ai. In one series of 1,770 scans, on a sample of texts chosen as suspicious, it rated 14.1% of human texts as AI and let through 10.4% of AI texts. [5] This is a lesson for the whole industry: even a good detector is wrong, so "100% human" is a bad goal.
Humanization turned out harder than detection. On saved samples, casual texts rewritten by models stayed at 94-100% AI in Genuino. [5] A medical text went down to 31.9%, but in a single run. 29 attempts at fine-tuning various OpenAI models did not produce a model that rewrites well. This stage is still open.