Anthropic engineer warns AI agents can erode critical thinking

In an essay posted Tuesday on X, Osmani described three ways AI can diminish users’ thinking. “Cognitive offloading” means handing tasks to AI while remaining responsible for the result; he said he does it every day and that others should, too. “Deferral” occurs when an AI response becomes the user’s next input. The most extreme form, “cognitive surrender,” is complete trust in AI output. Osmani called that lazy and warned that repeated surrender builds “cognitive debt.”

The phrase “cognitive surrender” was coined by two researchers at the Wharton School. Whether AI is making people less intelligent or changing how they think remains unclear. Researchers interviewed by Business Insider pointed to concerning signs but cautioned against firm conclusions.

For Osmani, the right level of reliance depends on the stakes. “Where the stakes are low, feel free to delegate entirely,” he wrote. But if someone cannot explain an answer, even through their agent, “you’ve ceded ownership.” He suggested asking questions such as, “Did I formulate a hypothesis before asking?” and “What am I no longer practicing?” Help and offloading can both be acceptable, he added, provided users know which one they are doing.

Osmani also urged engineers to make checking AI-generated code a habit. They should build feedback loops and ask agents what they can and cannot verify. Skills fade when they go unused, he argued, comparing coding to navigating without a phone app or remembering friends’ numbers without relying on contacts. “Keep doing the ones you’re not willing to let go,” he wrote.