The Myopia of Learning

In the AI Story

The three mechanisms operate in parallel. Temporal myopia discounts the future: an exploitation strategy producing measurable returns this quarter will outcompete an exploration strategy that might produce larger returns in three years, because the learning system updates beliefs based on observed outcomes, and exploitation's outcomes are observed first. Spatial myopia favors the local: an exploitation strategy improving performance in the organization's current market outcompetes exploration that might open new markets, because local improvements are attributable while distant possibilities are speculative. Failure-aversion underweights negative outcomes from unexplored alternatives and overweights positive outcomes from current strategies, producing a ratchet where exploitation successes encourage more exploitation while exploration failures discourage further exploration.

AI intensifies each mechanism with specific precision. The temporal mechanism is compressed to its extreme: when a developer working with Claude Code describes a feature, receives working code, and ships it within a single session, the learning system does not wait for a quarterly report. The outcome is visible before the developer has finished her coffee. The immediacy is intoxicating — Edo Segal describes it as productive vertigo — and the intoxication is itself the diagnostic. A learning system receiving exceptionally strong positive signals updates its beliefs rapidly and forcefully: this works, do more of this.

The spatial mechanism is equally intensified. AI tools are general-purpose in theory but local in practice. Organizations adopt AI to improve current processes — code generation, documentation, analysis — and the improvements consume attention entirely. The Berkeley study documented this narrowing empirically: workers who adopted AI expanded into adjacent domains, but the expansion was horizontal. Designers wrote code; engineers wrote documentation. No one started doing work that had not previously existed in the organization at all. The AI intensified exploitation across a wider domain but did not catalyze exploration of genuinely new territories.

The failure-aversion mechanism operates at the level of organizational culture and is therefore hardest to observe. When AI makes exploitation reliable, the organization's tolerance for the unreliability of exploration drops. Why fund an uncertain experiment when exploitation returns are so high? Why tolerate the messiness of genuine inquiry when the AI can generate clean, confident output on demand? The colonization of pauses by AI-assisted productivity that the Berkeley researchers documented is not a symptom of bad management. It is a symptom of a learning system operating exactly as the myopia framework predicts — favoring the near, the certain, and the measurable with an efficiency that leaves no room for the distant, the uncertain, and the unmeasurable.

Origin

March and Levinthal developed the framework in their 1993 paper in Strategic Management Journal, extending March's 1991 work on exploration and exploitation into explicit diagnosis of the mechanisms through which adaptive systems produce their characteristic pathologies. The paper's title — The Myopia of Learning — was deliberately clinical. The authors wrote as diagnosticians identifying the pathways of a chronic disease, not as moralists condemning organizational failure.

The framework's genealogy extends through Herbert Simon's work on bounded rationality and through the entire Carnegie School tradition of treating organizations as information-processing systems with specific structural limitations. March's contribution was to demonstrate that the limitations were not merely cognitive but institutional: the myopia operates through reward systems, performance metrics, budget allocation processes, and cultural norms that individually appear rational and collectively produce systematic bias toward the near.

Key Ideas

Three mechanisms. Temporal, spatial, and failure-averse myopia operate in parallel to produce the drift toward exploitation.

Rational at the point. Each individual decision favoring exploitation is well-supported by available evidence; the error is emergent at the system level.

Chronic, not acute. The myopia is a structural feature of all learning systems, not a defect of particular organizations or particular leaders.

AI intensification. Every mechanism is intensified by AI's compression of feedback loops and amplification of exploitation returns.

Invisibility of the cost. The returns to exploration are systematically invisible to a system that learns from what has already happened.

Debates & Critiques

Critics argue that the framework understates the capacity of sophisticated organizations to overcome myopia through deliberate structural design — innovation labs, corporate venture arms, skunkworks programs. Defenders respond that these structures are themselves subject to the myopia they were designed to counter: they are evaluated on exploitation-friendly metrics, staffed by exploitation-trained managers, and systematically starved during budget constraints. The empirical record of innovation programs — most of which fail or are eventually absorbed into the exploitation machinery — tends to support the defenders.