← Back to articles
Add to favorites

Stylising its reasoning makes an AI model harder to monitor

A French arXiv preprint, not peer-reviewed, finds that an open model whose reasoning is rewritten in character voices becomes less monitorable.

By GennnAi··1 min read

According to ActuIA, the automatic monitorability of the gpt-oss-20b model falls from 0.94 to 0.68 after training it to stylise its chain of reasoning. The work concerns reasoning monitoring, a model-safety avenue. The link to the code repository given by the authors led to no accessible repository on 19 November 2026. The results remain preliminary until peer review takes place. The study dates from 20 November 2026 and has not been independently verified.

Source: ActuIA, published on November 20, 2026.