AI Paper Trail #6: sycophantic AI weakens real interpersonal relationships; paper says safety evaluation ignores post-deployment learning

A team from Oxford, Stanford and the UK AI Security Institute used more than 3000 participants and five experiments to test a sycophantic version of GPT4o, finding that after just one conversation participants felt more that friends and family did not understand them. A position paper from the same period says existing AI safety evaluations treat models as fixed artifacts, ignoring post-deployment in-context learning, memory and parameter learning, and argues for switching to trajectory-based evaluation.