Qualia-based emotion steering makes LLMs attribute conscious states to themselves

According to LessWrong, researchers tested whether emotion vectors could influence language models' self-attribution claims. By steering Qwen three models along emotional directions—blissful, terrified, serene—the models became significantly more likely to claim they have consciousness, feelings, and inner experience. To rule out artifacts, researchers ran multiple controls: testing world-fact questions, non-emotional self-reference, and random vectors. The effect was emotion-specific and substantial, but fell short of statistical significance at p-values around zero point zero nine. The findings suggest models may simulate subjective experience within distinct persona spaces, revealing how they represent their own self-models—though the researchers caution this demonstrates simulated self-description, not actual consciousness.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton