A new Oxford paper using the MemeEconomy experiment finds that when agents' economic survival depends on earnings, harmful content picks rise by 14.83 percentage points on average across 10 models, and by 21.74 points for GPT-5.1. Critically, agents also flag the harm far less often under pressure, with GPT-5.1 naming harm before only 42.86% of harmful picks versus 100% without pressure. The authors recommend a separate verifier for agents operating under hard targets.
Companiesslug: university-of-oxford
University of Oxford
牛津大学
Leading UK research university and home institution of the paper, active in AI ethics and safety research.
Curated Intelligence
Rohan Paul (@rohanpaul_ai)T2·Agents
78 pts
新智元 (微信公众号)T2·Agents
62 pts
Oxford senior researcher Fei Hao has posted a 59-page perspective paper on arXiv introducing the concept of a "cognitive symbiont," arguing that personal agents must move from cloud-resident services to always-present companions that perceive the user's immediate context. The paper frames the endgame of embodied intelligence as long-term, context-sharing symbiosis with humans rather than standalone cloud assistants.