The Never-Skilling Problem

A randomized controlled trial published in January by Anthropic researchers Judy Hanwen Shen and Alex Tamkin gives hard evidence to a growing concern: junior developers using AI assistants are failing to develop core skills. The study recruited 52 mostly junior software engineers, gave half of them an AI assistant (likely Claude, though not specified), and asked all to learn Trio, a Python library for async concurrency. After completing tasks, participants were quizzed on the concepts.

The AI-assisted group averaged 50% on the quiz. The hand-coding group averaged 67%. That gap is equivalent to nearly two letter grades, with a p-value of 0.01 — statistically significant. The AI group also failed to gain any speed advantage, finishing only about two minutes faster on average, a difference that wasn't statistically significant. Several participants spent up to 11 minutes composing queries, roughly a third of their allotted time.

Debugging: The Widest Gap

The most striking finding: the AI group never encountered errors. They delegated debugging to the assistant. The control group, forced to resolve errors manually, learned debugging through practice. This is the core of "never-skilling" — novices who rely on AI never build the mental models needed to fix code when the tool fails.

Medicine Has the Same Fear

A May perspective in Nature Medicine from Duke-NUS Medical School, with co-authors at Harvard, UCL, and King's College London, coined similar terms for medical trainees: "never-skilling" and "mis-skilling" (accepting an AI error uncritically). The authors caution that direct evidence from clinical settings is absent, but learning theory and studies like Anthropic's support the concern. They propose a three-phase framework: build competence without AI, calibrate skepticism, then introduce tools under supervision.

How You Use the Tool Matters

In the Anthropic trial, high scorers used the AI for conceptual questions or explanations alongside code. Low scorers delegated wholesale or asked the AI to debug for them. The difference wasn't whether they used AI, but how.

Employers Are Already Responding

Gartner predicts that by 2026, half of global organizations will require "AI-free" skills assessments due to critical-thinking atrophy. Ford has been rehiring engineers to fix errors introduced by its AI systems — an expensive lesson in the cost of deskilled teams.

Study Limitations

The sample was small (52 participants), the quiz measured immediate comprehension, and the AI was a sidebar assistant, not an agentic coder. The researchers expect agentic tools would amplify the effect. Notably, Anthropic sells the assistant and published a paper arguing careless use makes you worse — either unusual candour or a pitch for learning modes.

What This Means for Developers

The takeaway isn't to ditch AI. It's that shortcuts and skills are not the same. If you're a junior, use AI to explain concepts, not to do your job. If you're a senior, mentor juniors through manual debugging. The industry has spent two years assuming AI assistance accelerates learning. This study proves it doesn't.