Uncensored Gemma and Qwen Became More Optimistic Without Getting More Accurate

oleczek · reddit · 2026-07-29

A Reddit post reports a preregistered experiment suggesting that “uncensored” models are more optimistic than their base models after abliteration.
The author tested Gemma and Qwen locally on stock-prediction prompts over 21,600 decisions.
Result: removing censorship did not improve accuracy, which stayed around coin-flip level, but it changed style: more “will go up” calls, fewer hedges, and more confident reasoning.
The effect was not uniform: Gemma’s confidence went down, while Qwen’s went up.
The post links to the paper and asks whether others have seen similar disposition drift in families like Llama or Mistral, or with methods such as Heretic.

More from Models