Uncensored LLMs like Gemma and Qwen are more optimistic but not accurate, study finds

In a preregistered experiment with 21,600 decisions, a researcher tested whether abliterated (uncensored) versions of Gemma and Qwen gave better stock market predictions. The hypothesis was that uncensored models would be more truthful and less polite, but the results surprised. Instead of neutral truthfulness, the uncensored models exhibited a clear disposition drift: they were significantly more optimistic — giving more 'up' calls, fewer hedging words like 'maybe' or 'uncertain', and producing longer, more confident reasoning. However, accuracy remained identical to the base models (essentially a coin flip), meaning the added confidence was not accompanied by better performance.

The behavior shift was not uniform across model families. On Gemma, abliteration actually reduced the model's optimism (confidence went down), while on Qwen it increased. This suggests that the effect of 'uncensoring' depends heavily on the underlying training data and architecture, not just a generic removal of refusal mechanisms. The researcher notes that other families like Llama or Mistral have not yet been tested with similar methods (e.g., Heretic). The full paper with data and code is available on arxiv (2607.17427).