AI Coding Tools Risk Blocking Path to Expertise, Studies Show

The article begins with a quote from Sam Altman envisioning intelligence as a commodity utility, then pivots to explore why that vision poses a hidden threat to the next generation of developers. The core insight comes from research by JetBrains, a major player in developer tools, which conducted a detailed study of junior and novice developers in live coding sessions. The findings were stark: participants believed they were using AI as a personal tutor, but JetBrains' data showed the opposite was true.

Developers who relied heavily on AI assistance often skipped crucial planning stages, finished with an "illusion of competence" rather than true understanding, and became lost when required to reason through problems. Conversely, developers who mitigated AI usage succeeded because they developed "negative expertise"—the ability to identify and ignore incorrect suggestions—and used AI to accelerate work they had already intended to create. Most striking: the novice developers with the fewest restrictions on AI assistance "had skipped crucial steps in the programming problem-solving process, and were now lost." Those who performed best were the ones who greatly limited or ignored the AI assistance altogether.

The article frames this as an "inverted learning" problem. LLMs work best for those with deep expertise because experienced developers can steer, audit, and verify outputs. For novices, the tools work in reverse: less knowledge means more potential for the model to mislead. When exploring unfamiliar territory, inexperienced developers often do not even know what questions to ask. One participant with solid planning habits suddenly abandoned problem-solving planning when using Copilot, jumped directly to code, and then had to rely on the LLM to fix errors the LLM itself created. The article emphasizes that LLMs lack judgment, empathy, and pedagogical intent; they are pattern interpolators from training data, not teachers.

A University of Pennsylvania study of 1,000 students learning mathematics with unrestricted AI assistance found they scored 17% worse than students using only a textbook—yet believed they were excelling. However, the same UPenn study tested a "Tutor" version where students asked for help and then independently solved problems. That group performed 127% better in practice sessions. Anthropic's 2026 study "How AI assistance impacts the formation of coding skills" reached parallel conclusions: "Cognitive effort—and even getting painfully stuck—is likely important for fostering mastery."

The article argues that true expertise—what the Germans call Fingerspitzengefühl (fingertip feeling)—is built through friction: tracing obscure errors, experiencing performance tradeoffs, rewriting approaches that will not scale. This struggle builds intuition. By automating away the mechanics of struggle, that intuition is never formed. David Cramer, co-founder of Sentry, skeptically notes that the bet being made is that LLMs will eventually fix all accumulated junk—a vision he dismisses as "a science experiment," arguing that code generation without deep understanding creates broken systems.

The article concludes that the path forward depends on a shift toward pedagogical use of AI systems. Rather than prioritizing code generation, the focus should be on using AI as a Socratic sparring partner that prompts reflection and maintains cognitive friction. Joel Spolsky's "Law of Leaky Abstractions" is cited: developers should not start with abstractions like Spring Boot, React, or Tailwind—and similarly should not begin with LLMs as code generators. The real skill lies in learning the underlying mechanics that those abstractions hide.

The article presents a fundamental tension in how AI coding tools are being deployed in the industry. While Sam Altman's vision of intelligence as a utility is widely embraced, the practical reality for junior developers is far more complicated. The research from JetBrains—conducted through detailed analysis of live coding sessions—reveals that the promise of AI-as-tutor does not match observed behavior. Novices leaning heavily on assistance skip planning stages that build intuition and judgment, while those who used AI sparingly performed best. This echoes findings from Anthropic's 2026 study, which emphasizes that "cognitive effort—and even getting painfully stuck—is likely important for fostering mastery."

The paradox the article identifies is urgent: industry messaging tells junior developers that avoiding AI tools will leave them behind, yet the tools themselves can prevent the friction-driven learning that builds expertise. The UPenn mathematics study of 1,000 students underscores the scale of the risk—students using unrestricted AI assistance scored 17% worse than textbook-only learners and yet were unaware of their underperformance. The article argues this creates a "skilled orchestrator paradox": the skills needed to manage AI agents safely are the same skills that atrophy through continued reliance on the AI. Only developers with years of prior experience (predating AI tooling) have the bedrock knowledge to wield these tools responsibly. Novices, by contrast, face an impossible choice: use tools designed for experts, or fall behind peers who do.

No comments yet. Be the first to share your thoughts!