AI and the Formation of Human Skills

How AI assistance affects the acquisition of new domain skills — as distinct from short-term productivity. Core concept: productivity and competence can be decoupled. Workflows that raise short-term output may simultaneously erode the human capacity to supervise, validate, and extend that output.

Empirical anchor

sources/2026-shen-ai-skill-formation (and its companion post sources/2026-shen-anthropic-coding-skills-post) run a randomised controlled trial on 52 junior developers learning the Trio async Python library. The AI group scored 17 percentage points (~two grade points) lower on a post-task competency quiz than the control group (50% vs. 67%, Cohen's d = 0.738, p = 0.01), with no significant speedup in task completion. The largest gap appears in debugging — exactly the skill needed to supervise AI-generated code.

Taxonomy of AI interaction patterns (Shen & Tamkin 2026)

Cognitively engaged — preserve skill formation:

Generation-then-Comprehension — generate code, ask follow-up clarification.
Hybrid Code-Explanation — interleave code and explanation requests.
Conceptual Inquiry — ask only conceptual questions; resolve errors independently.

Cognitive offloading — erode skill formation:

AI Delegation — hand the task to the AI.
Progressive AI Reliance — start engaged, slide into delegation.
Iterative AI Debugging — outsource error resolution. Heavy debugging-via-AI query ratio correlates negatively with quiz score (r = -0.41, p = 0.043).

Why it matters for BPM/APM

As organisations scale AI-augmented workflows (see concepts/ai-adoption and concepts/agentic-bpm), the human validators of AI output become the remaining bottleneck for safety and quality. If junior workers never build the conceptual model, the pool of qualified supervisors shrinks just as AI deployment requires more of them (Bowman et al. 2022). The Wharton/GBK survey (sources/2025-korst-wharton-gen-ai-enterprise-adoption) quantifies this concern: 43% of enterprise leaders warn of proficiency declines even as 89% see Gen AI as skill-enhancing, while training investment has fallen 8 percentage points and confidence in training as the primary path to fluency has fallen 14 percentage points.

Design implications

Intentional interaction modes. Learning-oriented AI modes (explanation-first) built into tools.
Friction-by-design. Require the user to articulate the concept before generating code; withhold full solutions for pedagogic tasks.
Role-aware deployment. Treat junior-developer AI access differently from senior-developer access; Wharton data shows junior hiring will be most disrupted.
Measurement. Short-term productivity metrics (lines of code, PRs) are insufficient; supplement with periodic competency evaluations.
Debugging in particular. The largest skill deficit in Shen & Tamkin — and the core safety skill — deserves explicit scaffolding.

concepts/ai-adoption · concepts/ai-agent-benchmarks · concepts/agentic-bpm · concepts/explainability-apm