AI crutches for coding: Anthropic confirms that speed kills understanding

We got used to thinking of neural networks as an exoskeleton for the brain that makes us stronger and faster. But a fresh Anthropic study forces a different view. It seems we are less putting on Iron Man's armor than sitting down in a wheelchair that carries us toward intellectual atrophy. While CTOs around the world celebrate shorter development cycles, a crisis is brewing inside the teams themselves, and it will surface in a couple of years.

Anthropic set out to test how using AI affects learning and the acquisition of new skills. It took 52 junior developers and gave them a task: implement a feature using a technology they had never encountered before. The subjects were split into two groups. The first was allowed to use Claude and other tools; the second relied only on documentation and their own brainpower. The outcome was as predictable as it is alarming for the industry's future.

Those who worked alongside AI closed tickets in record time. They looked like productivity superstars, turning out finished solutions in minutes. But the devil was in the final assessment. When participants were asked to explain how the code they had written actually works, the "with AI" group fell apart. It turned out they hadn't dug into the architecture, didn't understand why certain methods were chosen, and had simply acted as a relay between the chatbot and the code editor. Meanwhile the second group, which struggled with the documentation and made mistakes, showed a deep understanding of the topic and could apply that knowledge in other contexts.

The problem isn't AI itself but the dopamine trap of instant results. When a solution appears at the press of a button, the brain stops spending energy on deep processing of information. Why bother understanding pointers or asynchrony if Claude has already produced a working piece of code? In the long run, this creates a generation of "surface-level" programmers. They can assemble a prototype in an evening, but they will be completely helpless when the AI makes a subtle architectural error or when a project demands a non-standard solution that isn't in the model's training data.

This resembles what happened to our sense of direction after GPS arrived. We stopped memorizing roads and landmarks, trusting the blue dot on the screen completely. But with a navigator, the price of a mistake is an extra lap around the neighborhood. In software development, the price is security holes and unmaintainable legacy code that nobody at the company fully understands. We risk a situation where there is simply nowhere to get senior developers, because the road from junior to middle used to run through pain, mistakes and deep digging in other people's code, and has now been replaced by an endless Ctrl+C and Ctrl+V.

Companies that today encourage thoughtless use of AI to hit quarterly targets are effectively taking out a loan at enormous interest from their own future. Yes, performance rises, but expertise inside the team drains away. If a developer doesn't understand why the code works, they aren't a developer but a typewriter operator with extended functions. The irony is that AI will replace exactly these "operators" first, as soon as it becomes a bit more stable.

The bottom line: AI is an excellent tool for those who already know how to program, but a toxic crutch for those who are still learning. Will the industry manage to introduce "coding hygiene" rules in time, or are we doomed to software that nobody understands?