The Friction Requirement

In the AI Story

The experimental foundation for the friction requirement spans decades. Motor learning research shows that blocked practice of a single skill produces better immediate performance than interleaved practice of multiple skills, but interleaved practice produces better long-term retention and transfer. Guidance during practice produces better immediate performance than self-directed practice, but self-directed practice produces better independent capability. Immediate feedback produces better immediate performance than delayed feedback, but delayed feedback produces better learning. In every case, the condition that produces more friction in the moment produces more learning over time.

The AI-specific instantiation is direct. In pre-AI software development, the process of writing a function involved productive failures: error messages, hypothesis generation, documentation reading, iterative refinement. All four Ericsson conditions were present — effort was required (the problem was novel), the work was at capability boundaries (the developer could almost do it but not quite), feedback was specific (error messages named the gap), and refinement was iterative (try, fail, adjust, try again). Each debugging session deposited a thin layer of understanding. Claude removes all four conditions simultaneously: the developer describes the function, Claude writes it, it works, she moves on. The output is preserved — perhaps improved. The four conditions for development are eliminated together.

The standard objection is that the frustration was unnecessary, that better tools should eliminate it as better surgical instruments eliminated certain mechanical difficulties. The objection misunderstands what the difficulty was doing. The frustration was not caused by inadequate tools; it was caused by the gap between current understanding and the understanding the problem demanded. This gap is the condition for growth. When the tool closes the gap by handling the problem without requiring understanding, the tool has not removed an obstacle to growth — it has removed the condition for growth itself.

The 2025 evidence on professional deskilling is precisely what the friction requirement predicts. Endoscopists lose polyp-detection capability when AI is removed. Students underperform peers when GPT-4 access is withdrawn. Physicians become AI-dependent in diagnostic tasks. In each case, the practitioner's current performance was elevated by tool use and the underlying capability was eroded by the absence of the friction that builds it. The framework does not merely describe the phenomenon after the fact; it predicted it from first principles before the evidence accumulated.

Origin

The four-conditions specification originated in the 1993 Ericsson, Krampe, and Tesch-Römer paper on violinists at the Berlin Academy and has been refined across subsequent work. The insight that struggle is the mechanism rather than the obstacle was implicit in the earlier chess expertise literature but was made explicit and generalized in the 1993 paper's framework.

The application to AI-mediated work is contemporary, drawing on both the Ericsson framework and the convergent Bjork-tradition findings on desirable difficulties to explain why tools optimized for performance produce the developmental deficits currently being documented.

Key Ideas

Four jointly necessary conditions. Effort, boundary, feedback, refinement — remove any one and development stalls.

Frustration as signal. The subjective experience of struggle is evidence that cognitive adaptation is occurring, not that practice is failing.

Tool elimination of conditions. AI tools by design remove all four conditions simultaneously in the default mode of operation.

Predictive framework. The conditions specification generated predictions about AI deskilling that have been confirmed by 2023-2025 empirical work.

Design implications. Tools could in principle be designed to preserve the conditions (adaptive difficulty, withheld solutions, hint-based guidance) but require a fundamental reorientation from output-optimization to development-optimization.