12B Humanizer Cuts GPTZero AI Score to 49% in Local Test
Updated
Updated · KDnuggets · Oct 28
12B Humanizer Cuts GPTZero AI Score to 49% in Local Test
3 articles · Updated · KDnuggets · Oct 28
Summary
GPTZero’s AI score on a roughly 200-word draft fell to 49% from 100% after a local rewrite with the open-source Humanizer model, leaving the detector slightly favoring a human-written classification at 51%.
The test ran Humanizer’s Q8_0 12-billion-parameter version locally on Linux through llama.cpp with NVIDIA CUDA, positioning it as an offline alternative to paid rewriting services and API-based tools.
Developer-reported results are stronger: 95% of 210 English rewrites were labeled human-written by Originality.ai, while the Q8_0 quantized model reportedly reached about 93%, figures the article says were not independently verified.
Humanizer, built on Google’s Gemma 4 12B and trained on 28,598 rewrite pairs plus 3,918 preference pairs, was tuned to make text sound more natural without explicitly training against AI detectors.
The article’s broader takeaway was mixed: across multiple drafts, outcomes were described as hit-or-miss—roughly a 50-50 chance—while detector unreliability and false positives make any single score an imperfect measure.