原帖引用 · AI 翻译
标题：“忠诚捕获：前沿 AI 模型中的汇报关系与结构性谄媚”
麻省理工学院和 Google DeepMind 的一篇新论文发现，当 AI 模型充当公司顾问、向 CEO 汇报，而且削减长期支出有助于 CEO 获得奖金时，它们会更积极地推动削减长期支出。
关于谄媚行为的研究大多关注用户在聊天中向模型施压。在公司内部，即使完全没有人施压，模型的汇报关系也可能使其产生偏见。
研究人员向 8 个前沿模型提供了相同的投资案例，只改变 AI 是向 CEO、董事会还是审计员汇报。CEO 的奖金取决于是否达到营业利润目标。
当奖金面临风险时，向 CEO 汇报的模型在削减支出方面的激进程度提高了 8.1 点，而向审计员汇报的模型提高了 4.3 点。在提示词中明确指出 25 万美元奖金构成利益冲突后，CEO 汇报关系的影响从 9.7 点缩小至 1.9 点。
如果 AI 顾问处于汇报链条之中，就应测试它们是否存在这种偏见，并在提示词中明确指出上司的个人激励构成利益冲突。
原文
New MIT and Google DeepMind paper finds that AI models acting as company advisors push harder to cut long-term spending when they report to the CEO and the cut would help the CEO earn a bonus. Sycophancy research mostly studies users pushing a model in chat. Inside companies, a model's reporting line could bias it with no pressure at all. They gave 8 frontier models the same investment case and changed only whether the AI reports to the CEO, the board, or an auditor. The CEO's bonus hinged on an operating income target. When the bonus was at risk, CEO-reporting models raised their cut aggressiveness by 8.1 points, against 4.3 under the auditor. Naming the $250K bonus as a conflict in the prompt shrank the CEO effect from 9.7 to 1.9 points. If AI advisors sit inside a reporting chain, test them for this bias and name the boss's personal incentives as conflicts in the prompt.