Open-source Humanizer model claims 95% human-written rewrite rate, but tests show mixed results
A free, open-source AI model called Humanizer—fine-tuned on 28,598 AI-to-human text pairs—claims to rewrite AI-generated content into human-like writing with 95% success, per its developer. The 12-billion-parameter model, built on Google’s Gemma 4 12B, has been downloaded over 143,000 times on Hugging Face and supports local deployment via tools like llama.cpp or Ollama. Its developer reports…
Key points
- Humanizer, a 12B-parameter open-source model, claims 95% of its rewrites pass Originality.ai’s human-written filter, but tests show inconsistent results
- Developed on Google’s Gemma 4 12B, it avoids AI detectors in training, focusing on factual accuracy and natural phrasing instead
The author noted inconsistency in outcomes and questioned detector reliability. Humanizer runs offline on Windows, macOS, or Linux with GPU support, offering an alternative to paid cloud-based humanizers. The project’s GitHub and Hugging Face pages provide setup guides for non-developers, though advanced users can fine-tune parameters like context window and Flash Attention for performance.
Can This Open-Source Model Beat AI Detectors? Let’s Test It Locally
KDnuggets · 28 October 2026
Loading the full article…
This text was published by KDnuggets and written by Abid Ali Awan. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
Comments
via GitHub Discussions