计算器和GPS是“认知卸载”的典型案例，通常提高效率并带来明显收益：计算器可提升学生数学成绩和自信，而GPS让驾驶更便捷并降低迷路和糟糕交通的风险。与此并行的是代价同样可测的损失，2019年LaCour 及其同事发现，多项计算任务中，大学生往往对故意篡改的错误计算答案缺乏怀疑，甚至对明显荒谬的结果也不加验证；哈佛与麦吉尔的研究表明，终身GPS使用更多与更差的空间记忆相关，步行者若用手机导航也更容易走更长路线并停靠更多次数。
这一趋势也出现在信息检索：所谓“Google效应”指出，人们会更少记忆那些自己可以上网查到的内容，说明外部认知工具可替代记忆。随着AI以更强能力、对话式界面和高置信度语气介入，这种替代可能升级为“认知投降”，Shaw与Nave的实验发现：AI回答正确时，使用者优于仅凭自身判断组；AI回答错误时，使用者显著落后于对照组，说明其自我校验在下降。
对管理者而言，问题是如何防止这种下沉式依赖。研究显示，高“认知需求”倾向能部分降低风险，若配合“是否正确”的即时反馈与金钱激励，用户在遇到错误AI输出时更可能抵制错误；INSEAD与同事对200余名国际象棋俱乐部学生的研究也显示，若可随时点按获取AI提示的组，性能增益不到无法自主触发提示组的一半，意味着“可得性过强”会抑制技能形成。
Calculators and GPS are classic examples of cognitive offloading, usually improving efficiency: calculators can improve students’ math performance and confidence, while GPS makes driving easier and reduces getting lost or hitting bad traffic. Yet costs are measurable, as shown in a 2019 study by LaCour and colleagues where undergraduates often failed to question intentionally wrong calculator outputs—even implausible ones—and researchers at Harvard and McGill linked greater lifetime GPS use to poorer spatial memory, while phone navigation users took longer routes and made more stops than physical-map users.
The pattern also appears in online search, where the “Google effect” shows people remember less when information is expected to be available online. With AI’s stronger capabilities, conversational interface, and confident tone, this may become cognitive surrender: in Shaw and Nave’s experiment, participants using AI outperformed a control group when the AI was correct, but performed much worse when the AI was wrong, indicating weaker self-checking and excessive reliance.
For managers, the implication is that cognition must be actively protected, not left to drift. Individuals high in need for cognition were more resistant, and adding monetary rewards plus immediate correctness feedback made people override wrong AI answers more often, though still not to the level of independent judgment; an INSEAD study of over 200 chess-club students found that on-demand AI tips yielded less than half the performance gains of students who had AI help restricted in timing, implying unrestricted access can weaken skill development.
Source: The danger of cognitive surrender
Subtitle: How much should managers let AI do the thinking?
Dateline: The Economist May 2nd 2026