Op-ed: Internal representations don't mean LLMs have emotions
ValerioCapraro · x · 2026-08-27
Responding to an Anthropic paper finding emotional representations in Claude, the author argues that interpreting this as evidence of emotion is fundamentally flawed. It conflates similarity in representation and output with similarity in underlying processes.
The post notes that human emotions involve deep reorganization of attention, decision-making, and processing systems, whereas models currently only simulate outputs. Additionally, since consistent neural signatures for emotions haven't even been found in humans, the attempt to prove LLM emotions via stable signatures is invalid.
More from AGI Musings
- Ex-OpenAI researcher: AI's next steps are form factors and the science/cyber frontier — inductionheads · 2026-10-15
- Frontier labs paying massive sums for proprietary data to move into verticals — sarahookr · 2026-10-15
- Frontier labs are paying obscene sums for old Slack threads as real-world data becomes the moat — sarahookr · 2026-10-15
- Researchers say AI-assisted science boosts joy and curiosity, not paper mills — IgorCarron · 2026-10-15
- Miles Cranmer's "How to avoid AI Brain Rot" talk: AI pulls your work quality to the population mean — MilesCranmer · 2026-10-15
- Less than 0.001% of Africans use AI to full potential as $100-200/mo plans lock them out — weskambale · 2026-10-15