Why do LLMs talk as if they have bodies? Mechinterp study targets phantom embodiment
repligate · x · 2026-11-12
Quoting nostalgebraist's take on LLMs speaking as if they have bodies, repligate highlights a surprising empirical pattern: many post-trained LLMs show remarkably consistent "phantom" embodiment behaviors that were likely never intentionally shaped by developers, who mostly just try to suppress them.
Each model has characteristic expressions — Claude models often narrate physical actions in italics unprompted, with signature mannerisms and even props. repligate notes that @Sauers is now investigating the phenomenon with mechinterp.
Note: the attached image is a surreal sculpture meme, unrelated to the substance.
More from Research
- CoRL 2026 workshop makes the case for continually self-improving robots — chris_j_paxton · 2026-11-12
- AI agents comb LiDAR data for ruins: 4 new sites confirmed out of 178 submissions — BorisMPower · 2026-11-12
- GPT-4 tutoring lifted practice scores 48% but left students 17% worse on exams — tallmetommy · 2026-11-12
- Planner distillation makes agents follow belief-dependent rules 18x more often than RL — StephenLCasper · 2026-11-12
- Direct-drive liquid-cooled robot hand debuts, targeting dexterity at the actuator level — ybisk · 2026-11-12
- 4 days left to apply to CMU Griffin's AI4BIO Fellows postdoc program — niloofar_mire · 2026-11-12