Anthropic and Amazon show how AI at work erodes skills

A new analysis of so-called AI deskilling brings together several troubling signals from software development, medicine, and corporate IT. The main point is straightforward: if generative AI spends too much time doing the most complex mental work for a human, productivity might grow on the surface, but real skills—weaken.

How the risk manifests

One of the most illustrative examples is the story of software consultant Josh Anderson, who tried to build the Road Trip Ninja application almost entirely with AI. At first, the experiment looked promising, but as the project grew, everything hit a scaling wall: the codebase ballooned to over 100 thousand lines, conversations with the bot began consuming hours, and progress nearly stalled. Formally, the person could have intervened and finished the system themselves, but in practice, unpacking a large mass of poorly structured AI-generated code proved too difficult.

The problem is not limited to one case. The article cites observations from developers who felt a sharp drop in their own efficiency during recent Claude outages: when the assistant disappeared, their usual work pace broke down. Against this backdrop, an internal Anthropic study is particularly telling: the company found that generative programming tools can degrade debugging and code comprehension skills. If a specialist has grown accustomed to the model handling draft architecture, template functions, and part of the checks, regaining full control over the task turns out to be harder than it seems.

Why skills fade

For this effect, working terms have already emerged. One of them is the "AI rebound effect": the system appears to boost productivity, but simultaneously shrinks the mental models that a person uses in complex work. Researchers describe similar logic as "cognitive debt": an employee closes tasks faster today but stops training the critical skills needed tomorrow. In the end, a dangerous mixture emerges of tool dependency and false confidence in results.

"When automation takes on the details, situational awareness dulls."

• A person less frequently debugs problems manually and sees the root of errors worse
• The habit of checking intermediate steps, not just the final answer, declines
• Dependency grows on the model's availability, its updates, and the external vendor
• Model errors are easier to miss because the user has less hands-on practice

Where business pays the price

The article's authors point to several studies across different fields. In accounting, automation was already linked to skill erosion and declining critical thinking. In a Carnegie Mellon study supported by Microsoft, 319 knowledge workers reported losing some critical skills when actively using generative AI. In medicine, the conclusion is similar: assistants can speed up individual doctor actions, but if the system is removed, work quality can drop below the previous norm because part of professional routine has already stopped being trained regularly.

For business, this is not just a matter of convenience, but of resilience. If an AI service becomes unavailable, changes its operating rules, or simply fails at a critical moment, the team still needs people capable of taking the task on. This is precisely where the Amazon example surfaces: after layoffs and strengthening of remaining AI teams, the company, according to the author, lacked the engineering expertise to prevent and quickly resolve major outages.

There is a second risk as well: models need constant retraining on quality data, and that data is created by qualified specialists. If their level declines, the AI itself eventually becomes worse.

What this means

The main conclusion is not that AI should be turned off, but that it is dangerous to turn it into a replacement for thinking. While companies measure the effect by task execution speed, they may not notice the slow loss of competencies inside teams. If this trend takes hold, the market will get workers who know how to work well only alongside an assistant, but handle themselves worse without it—precisely where errors cost the most.