AI Dependency in Coding Is Not Just a Junior Problem

A study published in February by METR, a nonprofit that measures AI capabilities, found that most software developers across experience levels would not complete a limited set of coding tasks without AI assistance. The researchers had set out to update a well-known 2025 study on AI coding productivity. Instead they found something different.

With AI, the task completion rate across seniority was 85%. Without AI, only 22% of participants finished within the time limit. That 22% were almost entirely engineers with more than 10 years of experience. Mid-level engineers with 3 to 6 years showed a completion rate of 12% without AI. Junior engineers managed 3%.

The finding points to a different problem than skill atrophy in new hires. It suggests an entire profession is quietly restructuring its relationship to a tool in a way nobody planned or announced.

Anthropic ran a controlled trial earlier this year that has been widely cited as evidence for the junior-atrophy thesis. The headline number: developers using Claude completed tasks 25% faster on average. Engineers who had relied on AI for six months or more, when forced to work without it, took 40% longer than engineers who had never used it. Their debugging accuracy dropped by a third.

Anthropic also measured what they called “explanation ability.” Engineers who used AI for a given module were less able to describe its architecture or the trade-offs in their implementation choices. The model filled those gaps during the task, then the model was removed, and the engineers had nothing to say. The researchers called this “context substitution” and noted it grew worse the longer the tool was used on the same codebase.

Some companies are reacting. Google and Microsoft have introduced “no-AI Fridays” or partial code-review sessions without tool assistance, the report said. The goal is to maintain baseline debugging skills. METR’s data suggests the skill is not being maintained at a steady state. It is decaying, and the rate of decay accelerates with time spent in AI-assisted workflows.

“The dependency is not a junior problem,” one of the METR researchers told Reuters. “Every engineer below the top quartile of experience now defaults to AI as the primary tool for debugging and code comprehension. The seniors who succeed without AI are the ones who built mental models of the codebase before the tools existed. That base of knowledge cannot be acquired through a prompt interface.”