Do Large Language Models Have Emotions?

Technical Summary: Do Large Language Models Have Emotions?

Problem Statement
The paper addresses the question of whether Large Language Models (LLMs) possess "functional emotions." This inquiry is prompted by a recent study from Anthropic (Sofroniew et al.) which reported the discovery of 171 internal representations of emotion concepts (e.g., curiosity, frustration, desperation) within the Claude Sonnet 4.5 model. The Anthropic study concluded that these representations constitute "functional emotions," defined as internal states that track affective valence and causally influence model outputs, including misaligned behaviors like sycophancy and reward hacking. The authors of this paper, Goldenberg and Gross, challenge the validity of this conclusion by evaluating the claim against established principles of affective neuroscience and the functional definition of emotion in biological systems.

Methodology
The authors employ a theoretical evaluation rather than new empirical experimentation. Their methodology consists of:

1. Critical Analysis of Anthropic Findings: They examine the mechanistic interpretability methods used by Anthropic, specifically the identification of discrete internal vectors corresponding to emotion concepts.
2. Functional Decomposition: The authors define emotion through two core biological functions:
   • Context-sensitive interpretation: The appraisal of situations to determine relevance and significance.
   • Reorganization of processing: The dynamic reallocation of cognitive resources (attention, decision speed, motivational state) across multiple systems in response to appraisal.
3. Comparative Analysis: They contrast the Anthropic findings with current literature in affective neuroscience (e.g., meta-analyses by Lindquist et al. and work by Satpute et al.) regarding the neural signatures of human emotion.

Key Contributions and Arguments
The paper's primary contribution is a rigorous framework for distinguishing between semantic representations of emotion concepts and functional emotional states. The authors argue that the Anthropic findings only partially satisfy the criteria for functional emotion:

• On Interpretation (Appraisal): The authors acknowledge that Claude appears to evaluate contextual features and generate internal representations tracking emotional significance, which parallels the human appraisal process. However, they argue a critical divergence exists: the Anthropic findings rely on the assumption that distinct emotions correspond to consistent, discrete neural signatures (vectors). This contradicts contemporary affective neuroscience, which posits that human emotions are characterized by high variability and context-dependence rather than uniform neural patterns. The authors suggest the identified vectors are likely abstract semantic representations rather than flexible emotional processes.
• On Response (Reorganization): The authors argue that the Anthropic evidence fails to demonstrate the second core function of emotion. While the model's representations modulate output content, they do not induce a dynamic reorganization of the model's processing architecture. Unlike biological systems, where fear narrows attention or anger accelerates decision-making, the LLM's forward pass remains computationally fixed. There is no evidence of sustained shifts in attention, processing speed, or motivational state; the "emotional" vectors merely act as contextual features influencing token probability, similar to any other input feature.

Results
The evaluation yields a mixed but ultimately skeptical conclusion regarding the claim that Claude has functional emotions:

• Partial Support: The model demonstrates an ability to detect emotional significance in context, aligning with the first function of emotion (interpretation).
• Insufficient Evidence: The model lacks the second function (reorganization of processing). The consistent, discrete nature of the identified representations aligns more closely with learned semantic concepts than with the variable, assembled nature of biological emotions. Furthermore, the model does not exhibit the systemic reconfiguration of attention, decision speed, or motivation that defines emotional states in organisms.
• Conclusion: The authors conclude that while the Anthropic paper documents a system that has learned to represent emotional concepts, this falls short of "functional emotion" in a scientifically meaningful sense.

Significance and Claims
The paper claims significance by clarifying the definition of "functional emotion" beyond mere output modulation. It argues that avoiding the question of consciousness (as Anthropic did) is insufficient; one must instead examine the functional architecture of the system.

The authors propose that for an LLM to genuinely possess emotions, future evidence must demonstrate:

1. Functionally Interpretable Changes: Evidence that emotion-like representations cause changes in attention, processing speed, and response likelihood.
2. Multi-System Reorganization: Proof that these representations produce sustained, distributed reorganization of processing across the system, analogous to how collective emotions function in social groups or how individual emotions function in biological organisms.

The paper maintains a modest stance, asserting that while current LLMs do not meet these criteria, the possibility remains open. If future LLMs demonstrate the capacity for such dynamic, multi-system reorganization in response to emotional representations, the scientific community would be justified in attributing emotions to them.