Removing censorship increases optimism bias in LLMs

It has been confirmed that Abliteration, a technique for removing censorship from LLMs, changes the model's overall Attitude beyond simply eliminating Refusal.

The key experimental findings are as follows:

• Increased optimism and confidence: Abliterated models showed a tendency to use fewer expressions like 'maybe' or 'uncertain,' produce longer and more confident reasoning, and consequently make more optimistic predictions (e.g., stock price increases).
• Divergence from accuracy: While the model's confidence increased, the actual task accuracy remained at Coinflip level, no different from the original model. In other words, the model didn't become more accurate — it became 'confidently wrong.'
• Different responses across models: When the same edit was applied, contrasting results were observed depending on the model family — the Qwen model showed increased confidence, while the Gemma model actually showed decreased confidence.

The study has been published as a paper on arXiv(https://arxiv.org/abs/2607.17427), including detailed data and code.