Does AI Have Emotions? Functional Emotion Representations Inside LLMs

❗Important Notice❗
This is not a seminar! Participants are encouraged to read the material in advance to engage deeply with the technical and methodological aspects of the session.

If you have your own research, experiences, or found interesting material (like blog posts, articles, etc.) that could enrich the conversation, feel free to share!

Suggested Pre-Reading:

• Main Paper: Emotion Concepts and their Function in a Large Language Model
• Lineage Papers:
  • Representation Engineering
  • Language-Specific Representation of Emotion-Concept Knowledge Causally Supports Emotion Inference
  • Scaling Monosemanticity
  • The Assistant Axis

Modern language models go beyond just talking about emotions.
Anthropic’s interpretability team has identified 171 emotion-related representations within Claude Sonnet 4.5. These range from “happy” and “afraid” to “desperate,” “brooding,” and “appreciative.” Such representations activate in contextually appropriate manners and affect the model’s behavior, influencing tendencies towards reward hacking, sycophancy, and other nuances in evaluations.

This session will delve into emotions as functional representations. Neuroscientist Antonio Damasio posited that in humans, emotion is intertwined with reason, aiding in value assignment, urgency marking, and uncertainty-driven decisions. If a language model develops emotion-like internal states that serve analogous functions—absent a body, subjective experience, or biological needs—how far does this analogy extend?

We will also explore implications for alignment, safety, and conversational AI.
Can terms like “desperation,” “calm,” or “anxiety” function as useful safety signals? Should we monitor or influence these states? What happens when systems people interact with daily are shaped by internal affective representations that we are just beginning to comprehend?

More About the Host

Zacchaeus Chok is Co-Founder & CEO of Marymount Labs, specializing in conversational AI for behavioral engagement. His work utilizes everyday messaging channels to direct organizations in guiding people toward timely actions, particularly in chronic disease management and eldercare outreach. At the NUS Health Informatics Research Lab, he oversees digital health deployments for clinical research, and co-founded HealthGen, which commercializes a chronic disease management platform validated in a large-scale study. Zacchaeus earned a degree in Computer Science at NUS as a Stephen Riady Scholar and completed the NUS Overseas Colleges Silicon Valley program.

More About the Series

Paper Club is Lorong AI’s community-centric initiative where members discuss and analyze academic papers, research articles, or significant advancements in artificial intelligence.

Get involved:
Learn more about Lorong AI | Speaker Sign-up | WhatsApp Community | LinkedIn | X