Judgment Under Velocity

In The You On AI Encyclopedia

Gawande's research on time-pressured medical decision-making identified attentional narrowing as the predictable cognitive response to sustained urgency. Under pressure, practitioners focus on the most salient features of the situation — does it compile? does it pass tests? does it produce the expected output? — while peripheral information that might signal deeper problems goes unexamined. The narrowing is not a character flaw. It is the cognitive system doing what it was designed to do when full deliberation is structurally unavailable: default to pattern-matching on the most informative features and proceed.

The practical consequence for AI-assisted building is that builders operating at AI velocity are systematically predisposed to accept output that matches expectations and overlook output that departs from expectations in subtle ways. The AI-generated implementation that delivers the requested feature, compiles without error, and produces the expected behavior activates the pattern-match for "correct." The subtle architectural flaw, the edge case omission, the security gap — these are peripheral signals that narrowed attention is predisposed to miss. The fluency of the output compounds the effect by suppressing the ambiguity cues that would trigger broader evaluation.

Gawande's medical profession addressed the degradation of judgment under velocity not by slowing the work — emergency medicine cannot be slowed — but by building triage heuristics: simplified decision rules that sacrifice the accuracy of full deliberation for the reliability of systematic, repeatable verification under pressure. The trauma team runs the ABCDE protocol — Airway, Breathing, Circulation, Disability, Exposure — addressing immediate threats in sequence. The protocol is less thorough than comprehensive assessment. It is more reliable than unstructured assessment under the conditions that actually obtain.

The application to AI-assisted building is direct: build verification heuristics calibrated to the categories of failure most likely to produce consequential downstream damage. First-pass checks of AI-generated external references against the actual codebase and documentation (fabricated library calls, deprecated API signatures). Second-pass checks of architectural assumptions against project-specific constraints (defaulted patterns from the training distribution that mismatch the system's actual load profile or maintenance requirements). Third-pass checks of edge case handling (null inputs, concurrent access, timezone boundaries, overflow). The heuristics are not substitutes for comprehensive review — they are triage protocols designed to catch the most dangerous failures when the workflow's pace does not permit full deliberation.

Origin

Gary Klein

The concept draws on Gawande's treatment of surgical emergencies in Complications (2002) and on the broader literature on naturalistic decision making associated with Gary Klein's recognition-primed decision model. The specific translation to AI-assisted building is the analytical move of Chapter 6 of the Gawande companion volume, placing AI-era workflow pressures in the same cognitive category as time-pressured surgical judgment.

The adjacent literature on ironies of automation — Lisanne Bainbridge's 1983 insight that automation transforms the human's role into monitoring, which humans do badly — provides the complementary frame.

Key Ideas

AI velocity creates surgical-emergency cognition. The structural conditions of AI-assisted workflow produce the same attentional narrowing documented in time-pressured medical decision-making.

Narrowing favors expected outcomes. Under pressure, practitioners pattern-match on the most informative features and miss peripheral signals — exactly the signals AI fluent fabrications occupy.

Do not exhort, restructure. Individual vigilance is unreliable under sustained velocity; the remedy is verification heuristics embedded in the workflow.

Triage protocols over comprehensive review. Prioritized checklists calibrated to the most consequential failure modes are more reliable than aspirational thoroughness.

Context-dependent judgment remains irreplaceable. The builder brings knowledge of users, team, trajectory, and constraints that no training corpus encodes — judgment is the binding constraint, not the bottleneck.