Subjective emotion judgements adhere to principles of Bayesian inference and efficient representation

Emotion judgements are central to wellbeing and mental health, yet the fundamental principles determining how they arise remain poorly characterised. A key challenge is that repeated emotion ratings show substantial variability, typically treated as noise. Here we propose that emotion judgements, like perceptual judgements, arise from probabilistic inference under representational constraints. Using behavioural and computational approaches, we show that this variability is structured and meaningful. Emotion judgements are biased toward prior expectations when evidence is weak, consistent with Bayesian inference, and representational precision scales with the frequency of experienced intensities, consistent with efficient coding. Confidence decreases as uncertainty increases, indicating partial awareness of this uncertainty. A single model combining efficient encoding with Bayesian decoding captures these behavioural signatures. Notably, these signatures were largely preserved across variation in anxiety symptoms. These findings link emotion self-report to the principles of perceptual inference, providing a computational account of how uncertainty structures emotion judgements.