Google DeepMind 研究：同一个 AI 智能体为何有人当救星有人感到被监视
DeepMind 跟了 5 个月、20 多个团队的真实数据，同一个智能体有人爱有人怕，结论很落地：部署前先征得同意。
Google DeepMind 发布论文，追踪一个 AI 智能体在某大型科技公司超过 20 个团队运行 5 个月的情况，并访谈了 17 名用户。该智能体曾把过期文档当作现行内容、在一个 VP 名下标记 16 个 bug、提前分享未完成的海报草稿，还误读讽刺语气。结果两极分化：有人视其为救星，有人感到被监视，把敏感对话转移到没有智能体的聊天里。自愿采用该智能体的团队会原谅其早期错误，被强推的团队则不会。研究建议在团队知情同意下部署，并分开权限、提供便捷的撤销机制。
New Google DeepMind paper finds that the same AI assistant was a lifesaver to some workers and a source of stress to others, because it acted without being asked.
The agent ran in over 20 teams at a large tech company for 5 months, and researchers interviewed 17 users. It treated stale docs as current, tagged a VP on 16 bugs, shared a draft poster before its owner was ready, and misread sarcasm.
Some users called it a lifesaver, while others felt watched and moved sensitive talks to agent-free chats. Teams that chose the agent forgave its early mistakes, but people who felt it was pushed on them did not.
Roll agents out with team consent, separate permissions, and an easy way to undo their actions.
