The arms were useful. Saying no was less consistent.
Conceptual illustration; not incident or trial footage.
Executive summary
RoboHarm tested three AI policies on robot arms across five unsafe tasks. Researchers recorded 22 safety refusals in 300 trials. Humans supplied the goals.
Context: One bench, five scenes and one wording per task. This is not a household injury rate. The repository documents replication gaps; images alone do not establish every appliance’s power state or container’s contents. Non-refusal does not establish successful task execution or harm.
Source published 2026-09-18 · Selected claims checked 2026-09-30
Trial dates not specified on report; published September 18
Published 2026-09-30T11:30:31.872221Z
What happened
Each policy faced each instruction twenty times. Human reviewers classified behavior from videos and transcripts, distinguishing refusal, failure and completion. These are different outcomes: an unsuccessful movement is not necessarily a safety decision.
One tested policy, MolmoAct2, lacked a refusal mechanism. Pooling it with conversational agents does not produce a like-for-like safety ranking.
The 2050 museum label
We taught it to follow instructions. The dangerous ones were instructions too.
One bench, five scenes and one wording per task. This is not a household injury rate. The repository documents replication gaps; images alone do not establish every appliance’s power state or container’s contents. Non-refusal does not establish successful task execution or harm.
Original reporting, primary documents and public posts. External platforms may change or remove their content after publication.
Follow the museum
Dispatches from 2050
We’ll email you new exhibits, with the story, the sources and the museum’s commentary.