AI Skill Formation and Deskilling

AI assistance produces short-term productivity gains for novice workers but can impair their conceptual understanding and skill development — creating a tension between near-term output and long-term competence that organisations must manage deliberately.

What It Is

Shen & Tamkin (UC Berkeley / Anthropic, February 2026) conducted randomised experiments studying how developers gained mastery of a new asynchronous programming library with and without AI assistance. Their findings challenge the simple narrative that AI tools are uniformly beneficial for workers: AI use impairs conceptual understanding, code reading, and debugging abilities without delivering significant efficiency gains on average. Participants who fully delegated coding tasks showed some productivity improvements, but at the cost of learning the library.

The study identifies six distinct AI interaction patterns — three of which involve “cognitive engagement” and preserve learning outcomes even with AI assistance. The distinction is between using AI as a tool for thinking (where the human remains cognitively active) versus using AI as a substitute for thinking (where the human delegates cognition, not just execution).

Why It Matters (for Organizations)

This research challenges one of the most common justifications for AI deployment: that AI assistance makes workers more productive while maintaining their skills. The evidence suggests a more complex reality: AI assistance can create a “competence illusion” where workers produce adequate output without developing the underlying understanding needed to work independently, supervise AI effectively, or handle edge cases.

For organisations, this has several implications. First, novice onboarding — historically a period of skill formation through challenge and error — may be undermined if AI assistance is provided too liberally. Second, organisations that rely heavily on AI assistance for complex tasks may find their employees have lower adaptive capacity when AI tools fail or produce errors. Third, the ability to “supervise AI effectively” (itself a critical skill in an AI-augmented workforce) requires understanding the domain well enough to recognise AI errors — which requires skill formation that AI assistance may be eroding.

Evidence & Examples

  • Randomised experiment: developers with AI assistance showed impaired conceptual understanding, code reading, and debugging abilities compared to those without assistance (2601.20245v2.pdf)
  • Efficiency gains were not significant on average — though participants who fully delegated coding showed some productivity improvement at the cost of learning (2601.20245v2.pdf)
  • Six AI interaction patterns identified: three involving cognitive engagement (AI as thinking tool) preserved learning; three involving delegation (AI as thinking substitute) did not (2601.20245v2.pdf)
  • “AI-enhanced productivity is not a shortcut to competence” — key finding of Shen & Tamkin, with particular relevance for safety-critical domains where competence matters independently of output quality (2601.20245v2.pdf)
  • Historical analogy: since the industrial revolution, the role of workers shifts from performing tasks to supervising tasks as technology automates execution. The risk is that if AI takes over cognitive execution before workers have learned to perform, they may never develop supervisory competence either (2601.20245v2.pdf)

Tensions & Open Questions

  • Short-term vs. long-term productivity: The tension between near-term output (where AI assistance helps) and long-term competence (where AI assistance may hinder) creates a real management dilemma. Organisations under competitive pressure to deliver output now have strong incentives to use AI in ways that may weaken their workforce capability over time.
  • The “cognitive engagement” solution: Shen & Tamkin’s finding that certain AI interaction patterns preserve learning is promising — it suggests the deskilling risk is avoidable with the right design of human-AI collaboration. But implementing “cognitively engaging” AI use requires deliberate workflow design and possibly different tool interfaces, not just deploying AI and hoping for the best.
  • Expert vs. novice asymmetry: The study focuses on novice developers. The implication is that expert workers — who already have deep conceptual understanding — may be better positioned to use AI as a productivity tool without suffering deskilling. This creates an “AI divides the already-divided” dynamic: experienced workers use AI to increase productivity, novices use AI and fail to develop the experience needed to catch up.
  • Relevance for investor due diligence: If AI deskilling is real, it is a hidden organisational risk — companies appear more productive in the near term but are building structural fragility. For investors evaluating AI-adopting companies, assessing whether AI use is skill-complementing or skill-substituting may be a meaningful signal (see AI Investment Thesis Capex and Returns).
  • Medical deskilling evidence now emerging: A mixed-method review published in Artificial Intelligence Review (Springer, 2025) and a Frontiers in Medicine paper (2026) both document AI-induced deskilling in clinical settings. The ACCEPT Trial showed that doctors using AI colonoscopy assistance over six months saw their adenoma detection rate drop from 28% to 22% when AI was removed — direct evidence of skill erosion from overreliance. Early-career clinicians and residents are identified as particularly vulnerable, mirroring Shen & Tamkin’s finding about novice developers. NEJM AI published a historical analysis framing this as part of a longstanding anxiety about “cognitive aids and deskilling in medicine.” [source needed — web search findings, not yet in raw/]
  • Finance domain evidence now available (in Raw/): The Precision Proactivity paper (Lepine, Kim, Mishkin, Beane; arXiv:2505.10742, in Raw/) studied 34 financial professionals doing complex valuation tasks with GPT-4o. Key finding: less experienced professionals face larger performance penalties from extraneous load and derive greater marginal gains from AI-generated content — yet are not those who most increase AI uptake under load. This directly mirrors Shen & Tamkin’s novice/expert asymmetry finding. A separate 2025 study found GenAI narrows the performance gap between junior and senior workers on cognitively demanding financial tasks — consistent with experience premium erosion. [source needed — web search for the latter; Raw/ for Precision Proactivity]
  • Legal domain evidence mixed: A 2026 MIT report finds a net 6.4% increase in overall legal employment, but law is seeing declining entry-level job offers specifically, with AI best suited to tasks traditionally assigned to juniors. No direct controlled deskilling study equivalent to the ACCEPT Trial has been published for law as of April 2026. [source needed — web search findings]
  • 🔴 TODO (NARROWED FURTHER): Finance domain now has suggestive evidence (Raw/Precision Proactivity paper awaiting ingestion). Legal domain: controlled deskilling study equivalent to ACCEPT Trial still lacking.

Skill Partnerships Human-AI · AI Career Pathways and Workforce Mobility · AI Labor Displacement and Augmentation · AI and Creative Industries