IT之家 11 月 4 日消息,Anthropic Claude Code 团队工程师、前谷歌工艺高管 Addy Osmani 在 X 上发表长文《当 AI 智能体完成工作时,如何保持批判性思维》。
他认为,随着 AI 智能体越来越强大,人类可能在逐步依赖 AI 的同时,逐步丧失独立思考、验证结果和理解复杂平台的能力,最终将 AI 的判断直接当作自己的判断。他将这种现象称为“认知投降”。
Osmani 曾在谷歌工作超过 14 年,负责过 Chrome 开发者感受及谷歌云 AI 相关业务,当下在 Anthropic 从事 Claude Code 相关工作。
他表示自己是 AI 智能体工作流的坚定兼容者,日常工作中经常同时运行多个 Agent,但发现了一个值得警惕的变化:随着 AI 模型不断进步,输出结果中的显著错误越来越少,人类反而更难保持警惕。
过去人们通常需要经历信息收集、分析、判断等过程才能得到答案,如今这一过程可能被压缩到几秒钟。AI 生成的内容往往结构完整、表达流畅,且大部分情况下是正确的,这使得使用者越来越容易跳过独立思考和验证环节。
Osmani 将人类使用 AI 的状态划分为三个层次,IT之家总结如下:
实际上,“认知投降”一词最初由宾夕法尼亚大学沃顿商学院研究人员 Steven Shaw 和 Gideon Nave 提出。两人在题为《思考 —— 快速、缓慢与人工智能:AI 如何重塑人类推理及认知投降的兴起》的研究论文中讨论了这一现象。论文将其描述为,人们在缺少充分检验的情况下接受 AI 输出,并让 AI 的判断取代自身的直觉和思考。
Osmani 认为,这种转变往往不是突然发生的,而是在不断使用 AI 的过程中逐步形成。使用者可能认真阅读第一份 AI 报告,对第二份只开展快速浏览,到了第三份甚至只检查大纲。最终当有人询问某个具体结论时,使用者可能发现自己根本无法解释其中的逻辑。
他将这一现象与“认知债务”联系起来。在软件开发领域,AI 可以生成大量代码,测试也可能全部借助,但团队成员对平台架构和达成逻辑的理解却在不断减少。直到某天需要修改平台时,开发者才发现已经没有人能真正理解某些代码的工作原理。
不过 Osmani 并不认为开发者需要记住庞大代码库的全部细节。他指出,Claude Code 团队也会在需要时借助 AI 重新建立对特定代码模块的理解。真正关键的不是始终掌握所有知识,而是知道自己不了解什么,并且具备重新理解它的能力。
对于如何避免“认知投降”,Osmani 建议在向 AI 提出复杂难题前,人类应该先形成自己的初步假设,而不是完全依赖 AI 给出答案。即使最初的猜测并不准确,也可以协助使用者在接收 AI 结果时保留独立判断的基础。他还建议使用者主动寻找可以推翻 AI 结论的证据,而不是只询问模型对自己答案有多大信心。
在软件开发领域,Osmani 认为 AI 时代的瓶颈正在从编写代码转向验证代码是否正确。AI 智能体已经可以自主编写代码、运行测试、读取错误信息并反复修改,直到测试借助。但难题在于测试本身也可能由 AI 编写,如果测试逻辑与代码存在相同错误,即使所有测试表明借助,也无法证明软件不存在难题。
因而他强调,开发者不仅应该关注测试是否借助,还需要明确测试究竟证明了什么、没有覆盖哪些风险。对于 AI 交付的成果,验收标准应该是实际测试结果、运行日志、截图等可以验证的证据,而不是 AI 生成的一段看起来合理的总结。他还提出,开发者应该要求 Agent 明确列出它在执行任务过程中自主作出的决定、使用的假设以及忽略的条件。
随着 AI 可以同时执行越来越多任务,另一个难题开始浮现。Osmani 将其称为“编排税”,即 AI 生成的速度与人类可以认真审核、并愿意为其负责的速度之间的差距。
当 AI 生成的工作成果不断堆积而人类没有足够时间审核时,继续增加 Agent 数量并不能应对难题。他认为团队应该改进自动验证机制,或者适当降低任务生成速度。
Osmani 同时警告,AI 的高度自动化可能使人类逐步放弃练习原有技能。他建议开发者每周尝试在不使用 AI 的情况下编写一些代码,每个月重新审视自己已经停止练习的技能,并判断哪些能力仍然值得保留。他引用 Anthropic 关于内部工程师和研究人员使用 AI 的研究,指出其中存在一种“监督悖论”:监督 Claude 需要具备编程能力,但过度使用 AI 又可能使这些能力退化。
Osmani 强调,这并不意味着应该减少 AI 的使用。在错误成本较低的任务中,完全交由 AI 完成没有难题,但涉及投资、职业挑选、机型安全或影响其他人的重大决定时,人类仍应保留关键判断权。
他在最后写道:“你可以将工作和大部分审核任务交给 AI,但不能将责任也一并交出去。”
本文来源:https://www.ithome.com/1/022/031.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT Home, November 4 — Addy Osmani, an engineer on the Anthropic Claude Code team and former Google engineering process executive, published a long article on X titled "How to Maintain Critical Thinking When AI Agents Complete the Work."
He argues that as AI agents become increasingly powerful, humans may gradually lose their ability to think independently, verify results, and understand complex platforms while simultaneously relying more on AI, ultimately accepting AI's judgments directly as their own. He terms this phenomenon "cognitive surrender."
Osmani worked at Google for over 14 years, where he was responsible for Chrome developer experience and Google Cloud AI-related businesses, and currently works at Anthropic on Claude Code.
He states that he is a strong proponent of AI agent workflows and often runs multiple Agents simultaneously in his daily work, but has noticed a change worthy of caution: as AI models continue to advance, significant errors in their outputs are becoming increasingly rare, making it actually harder for humans to remain vigilant.
In the past, people typically had to go through processes such as information gathering, analysis, and judgment to get answers; today, this process may be compressed into just a few seconds. AI-generated content is often structurally complete, fluent, and largely correct, making it increasingly easy for users to skip independent thinking and verification steps.
Osmani divides the state of human AI usage into three levels, which IT Home summarizes as follows:
In fact, the term "cognitive surrender" was originally coined by Steven Shaw and Gideon Nave, researchers at the Wharton School of the University of Pennsylvania. The two discussed this phenomenon in a research paper titled "Thinking — Fast, Slow, and Artificial Intelligence: How AI Reshapes Human Reasoning and the Rise of Cognitive Surrender." The paper describes it as accepting AI output without sufficient scrutiny and allowing AI's judgments to replace one's own intuition and thinking.
Osmani believes that this shift often does not happen suddenly but is gradually formed through the continuous use of AI. A user might carefully read the first AI report, merely skim the second, and by the third, only check the outline. Ultimately, when asked about a specific conclusion, the user may find themselves completely unable to explain the logic behind it.
He links this phenomenon to "cognitive debt." In software development, AI can generate massive amounts of code, and testing may also be entirely delegated to AI, but team members' understanding of the platform architecture and the logic behind how it was achieved is constantly diminishing. Until one day when the platform needs modification, developers discover that no one truly understands how certain code works.
However, Osmani does not believe that developers need to remember every detail of a massive codebase. He points out that the Claude Code team also uses AI to rebuild their understanding of specific code modules when needed. What is truly critical is not always mastering all knowledge, but knowing what one does not understand and possessing the ability to re-learn it.
Regarding how to avoid "cognitive surrender," Osmani suggests that before posing complex problems to AI, humans should first form their own initial hypotheses rather than completely relying on AI for answers. Even if the initial guess is inaccurate, it helps users maintain a foundation for independent judgment when receiving AI results. He also suggests that users actively seek evidence to refute AI's conclusions rather than merely asking the model how confident it is in its answers.
In software development, Osmani believes the bottleneck in the AI era is shifting from writing code to verifying its correctness. AI agents can already autonomously write code, run tests, read error messages, and repeatedly modify the code until the tests pass. However, the difficulty lies in the fact that the tests themselves may also be written by AI; if the test logic shares the same flaw as the code, even if all tests pass, it cannot prove that the software is free of issues.
Therefore, he emphasizes that developers should not only focus on whether tests pass but also clarify exactly what the tests prove and which risks they fail to cover. For AI-delivered results, acceptance criteria should be verifiable evidence such as actual test results, runtime logs, and screenshots, rather than a seemingly plausible summary generated by AI. He also proposes that developers should require Agents to explicitly list the autonomous decisions made during task execution, the assumptions used, and the conditions ignored.
As AI can execute an increasing number of tasks simultaneously, another difficulty has emerged. Osmani calls this the "orchestration tax," which is the gap between the speed at which AI generates work and the speed at which humans can carefully review it and are willing to take responsibility for it.
When AI-generated work continuously piles up and humans lack sufficient time to review it, simply increasing the number of Agents cannot solve the problem. He believes teams should improve automated verification mechanisms or appropriately reduce the speed of task generation.
Osmani also warns that AI's high level of automation may cause humans to gradually abandon practicing their original skills. He suggests that developers try writing some code without using AI every week, and review the skills they have stopped practicing monthly to determine which abilities are still worth retaining. Citing Anthropic's research on internal engineers' and researchers' use of AI, he points out a "supervision paradox": supervising Claude requires programming skills, but overusing AI may cause these skills to degrade.
Osmani emphasizes that this does not mean AI usage should be reduced. For tasks with a low cost of error, there is no problem with leaving them entirely to AI; however, for major decisions involving investments, career choices, machine safety, or affecting others, humans should still retain the right to critical judgment.
He concluded by writing: "You can hand over the work and most of the review tasks to AI, but you cannot hand over the responsibility along with it."
