How AI assistance impacts the formation of coding skills

Source: Anthropic Research Date: 2026-01-29 URL: https://www.anthropic.com/research/AI-assistance-coding-skills

Summary

RCT with 52 junior software engineers: AI-assisted vs. hand-coding on tasks requiring mastery of a new Python library (Trio). AI group scored ~17% lower on skill assessments (≈2 letter grades), despite completing tasks ~2 minutes faster (not statistically significant). Interaction style mattered: participants who used AI to build understanding through follow-up questions retained significantly more than those delegating entirely.

Implications

The deskilling concern in the AI skill formation thread now has RCT-level evidence, not just survey data. Two letter grades of knowledge deficit from AI assistance is a large effect for a small productivity gain. The interaction style finding is the actionable result: AI-as-tutor vs. AI-as-delegatee produces dramatically different learning outcomes, and the difference is in user behavior, not tool capability. This has direct implications for engineering education and onboarding design — companies can preserve skill formation by structuring AI use around comprehension-seeking interactions. Watch whether coding bootcamps, CS programs, and enterprise onboarding programs respond to this with explicit AI-use guidelines.