Anthropic Engineer Warns: The Stronger AI Gets, the More Easily Humans Fall into "Cognitive Surrender"

Anthropic Engineer Warns: The Stronger AI Gets, the More Easily Humans Fall into "Cognitive Surrender" Phoenix Tech Download the App for Exclusive Early Access November 04, 2026 11:36:44 From Beijing Updated November 04, 2026 11:37:27

Phoenix Tech News, Beijing Time, November 4 — Addy Osmani, an engineer on Anthropic's Claude Code team and former Google technical executive, recently published a long post on social platform X titled "Critical thinking when agents do the work," warning that as AI agents grow more powerful, humans may enjoy efficiency gains while gradually losing the ability to think independently, verify results, and understand complex systems. He calls this phenomenon "Cognitive Surrender."

Osmani worked at Google for more than 14 years, where he was responsible for Chrome developer experience and Google Cloud AI-related business, and currently works on Claude Code-related matters at Anthropic.

Osmani said he is a firm supporter of AI agent workflows and often runs multiple agents simultaneously in his daily work. But he has noticed a worrying change: as AI models keep improving, obvious errors in their output become rarer, making it harder for humans to stay vigilant.

In the past, people usually had to go through processes such as information gathering, analysis, and judgment to arrive at an answer, but now this process can be compressed into a few seconds. AI-generated content is often well-structured, fluently expressed, and correct most of the time, which makes it increasingly easy for users to skip independent thinking and verification.

He divides the states of human AI use into three levels: first, "Cognitive Offloading," that is, handing tasks to AI to complete while still being responsible for the results; second, "Deferral," that is, no longer treating AI output as a hypothesis that needs verification but directly using it as the basis for the next step of work; and finally, "Cognitive Surrender," that is, AI output gradually becomes the user's own judgment, and the user no longer even considers it necessary to check.

Among these, the concept of "Cognitive Surrender" was first proposed by Wharton School researchers Steven Shaw and Gideon Nave.

Osmani believes that this shift often does not happen suddenly but gradually forms through continuous use of AI. A user may carefully read the first AI report, only skim the second, and by the third even check only the outline. Eventually, when someone asks about a specific conclusion, the user may find that they cannot explain the logic behind it at all.

He links this phenomenon to "Cognitive Debt." In software development, AI can generate large amounts of code, and tests may all pass, but team members' understanding of the system architecture and implementation logic keeps decreasing. Until one day when the system needs to be modified, developers discover that no one truly understands how certain features work.

However, Osmani does not believe developers need to memorize all the details of a massive codebase. He points out that the Claude Code team also uses AI when needed to rebuild understanding of specific code modules. What really matters is not always mastering all knowledge, but knowing what you do not understand and having the ability to re-understand it.

As for how to avoid "Cognitive Surrender," Osmani suggests that before asking AI complex questions, humans should first form their own preliminary hypotheses rather than relying entirely on AI to provide answers. Even if the initial guess is not accurate, it can help users retain a basis for independent judgment when receiving AI results.

In software development, Osmani believes that the bottleneck in the AI era is shifting from writing code to verifying whether code is correct.

Currently, AI agents can already autonomously write code, run tests, read error messages, and repeatedly revise until tests pass. But the problem is that the tests themselves may also be written by AI. If the test logic contains the same errors as the code, even if all tests show passing, it cannot prove that the software has no problems.

Therefore, he emphasizes that developers should not only focus on whether tests pass but also clarify what the tests actually prove and which risks they do not cover. For deliverables from AI, the acceptance criteria should be verifiable evidence such as actual test results, run logs, and screenshots, rather than a plausible-looking summary generated by AI.

Osmani also proposes that developers should require agents to explicitly list the decisions they made autonomously during task execution, the assumptions they adopted, and the conditions they ignored, because this content usually does not appear in the final work summary.

As AI becomes able to execute more and more tasks simultaneously, another problem begins to emerge. Osmani calls it the "Orchestration Tax," namely the gap between the speed at which AI generates results and the speed at which humans can carefully review and are willing to take responsibility for them.

When AI-generated work results keep piling up and humans no longer have enough time to review them, continuing to increase the number of agents cannot solve the problem. He believes that in this case, teams should improve automated verification mechanisms or appropriately slow down the pace of task generation.

Osmani also warns that the high degree of automation of AI may cause humans to gradually give up practicing their original skills. He suggests that developers try writing some code without using AI every week, re-examine the skills they have stopped practicing every month, and judge which abilities are still worth retaining.

He also cites Anthropic research on internal engineers' and researchers' use of AI, pointing out that there is a "supervision paradox": supervising Claude requires programming ability, but excessive use of AI may cause these abilities to degrade.

Osmani emphasizes that these views do not mean that AI use should be reduced. For tasks with low error costs, fully handing them over to AI is fine, but when it comes to major decisions involving investment, career choices, product safety, or affecting other people, humans should still retain key judgment authority.

In his view, future AI models may be better than humans at finding errors, but this cannot eliminate human responsibility. Deciding what product to develop, why it is worth developing, who it will affect, and what costs one is willing to bear for it still require humans to make choices.

For more first-hand news, welcome to download the Phoenix News app and subscribe to Phoenix Tech. For in-depth reporting, please search "Phoenix Tech" on WeChat.