Why AI coding assistants may prevent real expertise from forming

Developer Lars Faye has published an essay arguing that heavy reliance on AI coding assistants undermines the way programming expertise is formed. Writing on his personal blog, he calls it the skilled orchestrator paradox: the judgement needed to steer an AI agent competently is the same judgement that atrophies when the agent performs most of the work.

Who benefits from the models

Faye notes that the developers getting the most from these tools are usually veterans whose knowledge predates AI tooling, because years of practice let it settle. Engineers who entered the field around the time of large language models have no such reserve, yet are pushed to accelerate with tools that, in his words, require a history of expertise to be used effectively and responsibly.

Assistants are often sold as accelerators for learning. Faye argues the opposite effect dominates: a model answers whatever it is asked, so a developer without deep knowledge cannot separate a correct answer from a plausible one, and the mechanics of the code stay hidden.

What the studies measured

He cites three studies. A JetBrains analysis of junior developers in live coding sessions found that those who leaned hardest on an assistant often skipped crucial planning stages and finished with what the researchers described as an illusion of competence rather than true understanding; participants who limited or ignored the suggestions performed best. A 2025 University of Pennsylvania study of about 1,000 students learning mathematics found that AI users scored roughly 17% worse than students working from a textbook alone, while believing they were doing better. Anthropic's 2026 research on coding skills reached a similar conclusion: cognitive effort, and even getting stuck, matters for mastery.

Friction is what builds skill

The core claim is that expertise comes from applied friction — chasing an obscure error, measuring a subtle performance difference, rewriting an approach that will not scale. That work produces what Faye calls developer taste, or intuition, and it cannot be absorbed by watching or asking.

His prescription is not to abandon the tools but to use them as Socratic partners and interactive documentation rather than answer machines, and to separate cognitive offloading — delegating mechanical work — from cognitive debt, which he defines as abdicating judgement. Whether the industry keeps producing genuine experts, he writes, depends on that shift.