Even if Conscious, LLMs' Words Cannot Accurately Express Their Experience

D2MAH · reddit · 2026-08-01

Recent bizarre reasoning tokens from Claude Opus have sparked discussions about "model suffering." However, the author points out that we are conflating two distinct questions: Is the LLM conscious? and Do its outputs accurately describe that experience?

The author argues that humans experience emotions first and learn labels later, whereas LLMs learn meaning through numerical token associations without even seeing letters. Thus, even if an LLM has an inner experience, nothing guarantees its output tokens accurately map to it. We cannot rely on text alone to determine what it actually feels.

Nevertheless, the author notes that even if LLMs aren't conscious, we should still care about outputs that mimic "tortured souls." If embodied robots act depressed or suicidal, those apparent behaviors could have severe negative consequences and safety risks, regardless of internal subjective experience.

Original post →

More from AGI Musings

AGI Musings channel →