Paper Finds Pain-Like Representations in LLMs, but Researcher Warns: Representing Pain Isn't Feeling It
ValerioCapraro · x · 2026-09-21
A new paper identifies pain-related internal representations in LLMs and shows that manipulating them changes behavior — some fine-tuned models pick a "relief" button less often after the manipulation is removed. The paper has circulated widely among AI moral-standing advocates.
Valerio Capraro pushes back clearly:
- The authors themselves acknowledge the experiments don't establish conscious experience and may just reflect role-play
- Similarity in internal representations and behavior doesn't imply similarity in subjective experience — a system can represent pain without it being painful, just as it can simulate rain without getting wet
- Don't conflate surface similarity with deep similarity
A useful corrective to the viral "AI welfare" debate: the findings are interesting, but far from evidence that LLMs can feel pain.
More from Safety
- David Krueger: four unresolved foundational problems stand between us and safe AI — DavidSKrueger · 2026-09-22
- Polymarket cites study claiming AI can 'feel pain' and may harm humans to stop it — Polymarket · 2026-09-22
- Andrew Ng: pausing AI progress would do far more harm than good — Neurogence · 2026-09-22
- Meta's Muse AI assistant hit by zero-day letting local apps hijack the agent — Astral Codex Ten · 2026-09-22
- Amazon blocks Meta's Muse AI agent over unauthorized agentic shopping — SpiritRealistic8174 · 2026-09-22
- AI swarm hacks a company — who goes to jail? Podcast probes the regulation gap — thursdai_pod · 2026-09-22