---
title: "How AI Assistance Impacts the Formation of Coding Skills (Anthropic research post)"
type: source
tags: [ai-skill-formation, cognitive-offloading, software-engineering, anthropic, blog]
authors: [Shen, Judy Hanwen; Tamkin, Alex]
year: 2026
venue: "Anthropic Research blog"
kind: webpage
raw_path: "raw/AI Capabilities & Adoption/How AI assistance impacts the formation of coding skills \\ Anthropic.webloc"
sources: []
key_claims:
  - "Companion research post to the arXiv paper; summarises the randomised controlled trial with 52 junior software engineers learning the Trio Python library."
  - "AI group averaged 50% on the competency quiz vs. 67% for the hand-coding group (Cohen's d = 0.738, p = 0.01) — roughly two letter grades."
  - "AI-assisted participants completed tasks only ~2 minutes faster on average — not statistically significant."
  - "Largest performance gap was on debugging questions, suggesting debugging is the skill most at risk from over-reliance."
  - "Three AI interaction patterns preserve learning (Generation-then-Comprehension, Hybrid Code-Explanation, Conceptual Inquiry); three erode it (AI Delegation, Progressive AI Reliance, Iterative AI Debugging)."
  - "Recommends deliberate workflow design and learning-oriented AI modes (e.g., explanatory modes) to keep junior engineers able to validate AI-generated code."
created: 2026-04-20
updated: 2026-04-20
---

# How AI Assistance Impacts the Formation of Coding Skills (Anthropic post)

## Summary
Anthropic research blog post (Shen & Tamkin, January 2026) pointing to the full paper [[sources/2026-shen-ai-skill-formation]]. The webloc file in `raw/` resolves to `https://www.anthropic.com/research/AI-assistance-coding-skills`. The post distils the RCT for a broader audience:

- **Population:** 52 junior software engineers.
- **Task:** learn and use the Trio async Python library; post-task closed-book assessment.
- **Key result:** AI group 50% average, control 67% — nearly two letter grades lower, large effect size (d = 0.738, p = 0.01).
- **Productivity:** AI users finished ~2 minutes faster — not statistically significant.
- **Skill most affected:** debugging (largest gap).
- **Taxonomy:** six AI-interaction patterns — three preserve learning (Generation-then-Comprehension, Hybrid Code-Explanation, Conceptual Inquiry), three undermine it (AI Delegation, Progressive AI Reliance, Iterative AI Debugging).

**Takeaways for practice.** Learning-oriented AI modes (explanation-first tools) and deliberate workflow design can preserve skill formation. For organisations deploying AI into junior developer workflows, the costs of unmanaged adoption may be invisible in short-term productivity metrics but material for supervisory competence — particularly in safety-critical deployments where humans must validate AI-generated code.

## Connections
- Direct companion to the peer-reviewed-style paper [[sources/2026-shen-ai-skill-formation]]; both feed [[concepts/ai-skill-formation]].
- Aligns with enterprise-level skill-atrophy warning in [[sources/2025-korst-wharton-gen-ai-enterprise-adoption]] (43%).
- Methodological contrast with [[sources/2025-becker-metr-ai-developer-productivity]] (experienced developers, productivity outcome, not learning).
- Authors: [[entities/judy-hanwen-shen]], [[entities/alex-tamkin]].