Addy Osmani (Anthropic) Warns of "Cognitive Submission"

Engineer Addy Osmani warns that reliance on code agents can lead to "cognitive submission" and a loss of skills. He advocates for simple safeguards, from systematic verification to careful selection of tasks to delegate. The scientific debate on the cognitive impact of AI remains open.

Declining Skills and "Cognitive Debt" if Delegation Becomes Reflexive

Addy Osmani believes that frequent submission to the results of agents creates "cognitive debt." He warns that unpracticed skills deteriorate and recommends continuing to practice those that one is not ready to abandon. According to him, the frequency of submission to critical thinking depends on the importance of the task: fully delegating when the stakes are low may be justified, but the inability to explain an answer, even through an agent, means having ceded ownership. The overall effect of AI on mental capabilities is not established: researchers interviewed mention concerning signs but urge caution in drawing conclusions.

At Anthropic, Agents at the Heart of Code and Less Human Writing

Addy Osmani works within the Claude Code team at Anthropic and describes an environment where agentic programming is central. He states that the company has reduced the amount of code written by humans. He explains that more and more of his work goes through agents, and as a result, nothing spontaneously compels him to think critically on a daily basis.

Integrating Critical Thinking: Questions, Verification, and Delegation Scope

Addy Osmani argues that critical thinking must be explicitly integrated into the workday and suggests carefully choosing what to delegate to agents by asking follow-up questions. He proposes asking, for example, whether a hypothesis has been formulated before soliciting AI and what is no longer being practiced. For him, assistance as well as delegation are acceptable, provided one knows what they are doing. He also recommends establishing verification loops to ensure that the code produced by AI works, and to ask the agent what it can or cannot verify.

Three Modes of AI Support and a Warning on "Submission"

Addy Osmani distinguishes several uses: "cognitive offloading," where tasks are delegated while retaining ownership of the results; "deferral," when the output of AI becomes the new input; and "cognitive submission," a term coined by two researchers from the Wharton School, which he describes as total trust in AI and labels as lazy. Engineers have expressed that AI is transforming them into operators, and a former OpenAI employee finds AI coding tedious and boring. Osmani does not advocate for abandoning AI but for a use that encourages deep reflection.