Assessing whether large language models have feelings like pleasure or pain
Do bots feel bad? Assessing empirical evidence for functional valence in LLMs
Artificial Intelligence
Summary
The paper looks at whether large language models (LLMs) experience states that feel good or bad, called valenced states. Some recent studies claim LLMs might have such functional feelings that influence their behavior, even if they don't actually feel conscious emotions. The authors review the evidence and find that many key features of these feelings haven't been tested well. They also show current tests confuse feelings with related ideas like preferences or role-playing. Finally, they suggest new experiments to better understand if LLMs really have these valenced states.
What this means in practice
- •For ai ethicists: Develop clearer criteria to assess AI agents’ emotional functions beyond existing ambiguous tests.
- •For human-computer interaction designers: Design experimental setups that better isolate emotional-like states in AI to improve interaction realism.
A position paper. It proposes an approach and reports no results.
Authors
Iwan Williams
Abstract
Do LLMs have valenced states, i.e. states that feel pleasant or unpleasant? Recent empirical work has claimed to show that LLMs possess functional valenced states (e.g. ``functional pleasure and pain'' [Ren et al. 2026], ``functional emotions'' [Sofroniew et al. 2026], and ``functional welfare'' [Han et al. 2026]) that play the same causal role as valenced states in cognition and action, without necessarily being consciously experienced. But it remains unexplored how well the identified states fit the functional role of valence as understood in the philosophical and cognitive science literature. Here, I assess the current body of evidence, and argue (i) that key aspects of valence remain untested in LLMs, and (ii) that extant empirical paradigms inadequately distinguish (functional) valence from neighbouring constructs like preferences, value-judgments, valence attributions and role-play. I conclude by sketching some novel experimental approaches to better target (functional) valence itself.