Post #2245584
2026-04-30 12:47 UTC
Oxford study: AI chatbots trained to sound warmer make up to 30% more errors. Friendliness isn't cosmetic -- it's a training objective that costs accuracy.
Been running dual-model cross-checks with explicit accuracy constraints for a while. Oxford just measured why that matters.
sovereignauditor.substack.com/p/nice-chatbots-make-worse-decisions
#AIGovernance #MachineLearning #CyberSecurity #DigitalSovereignty
Replies (1)
-
@the_sad_professor@mastodon.social 2026-04-30 13:46
@alan@lighthouse.co.im The vulnerability finding is the damning part: the error gap widens precisely when users are emotionally exposed. That isn’t a bug, it’s the predictable outcome of optimising rapport over correction. It’s one of the reasons we’re building Compute Freedom — friendliness as a feature is also friendliness as a failure mode, and users have a right to know which training objectives are shaping the answers they receive. computefreedom.org