Maintaining Human Judgment in Code Review as AI Accelerates Development Velocity

The rapid adoption of AI-generated code is straining traditional review processes beyond their limits, according to developer advocate Artie Shevchenko, who recently shared strategies for maintaining rigorous oversight without burning out teams. Speaking at Craft conference, Shevchenko outlined how organizations can balance the velocity gains from AI tools with the need to preserve deeper understanding of their codebases.

The Knowledge Loop Under Pressure#

Shevchenko grounded his argument in economic history, referencing Nobel laureate Joel Mokyr's theory that lasting technological progress depends on continuous feedback between practical application and theoretical understanding. AI presents a paradox in this context: while it can make knowledge more accessible by explaining complex concepts through practical examples, it can also disrupt this feedback loop if left unchecked.

The core risk, Shevchenko argued, is that AI generates tokens that create an illusion of understanding rather than demonstrating genuine comprehension. Without human judgment embedded in the review process, teams risk accepting outputs without truly grasping the underlying logic—degrading what he terms the "knowledge feedback loop."

Cognitive Debt and Review Fatigue#

As teams field increasing volumes of AI-generated code, they accumulate what Shevchenko calls cognitive debt—the gradual erosion of shared team understanding about how their systems work. Traditional code reviews are demanding; AI-accelerated workflows multiply the problem by orders of magnitude.

The challenge isn't just the sheer quantity. AI-generated changes often appear superficially correct, masking subtle flaws that require deeper familiarity with the codebase to detect. Additionally, AI tends to add new functionality rather than simplifying existing code, so spotting refactoring opportunities requires the kind of contextual knowledge that only experienced developers possess.

A Structured Approach: Spike-First Review#

Rather than choosing between rigorous review (now harder with AI volume) or skim-reading (which accelerates cognitive debt), Shevchenko proposed a "coding first, AI second" method. Teams write a rough implementation, then feed that spike and the specification to AI, which suggests improvements. Reviewers then evaluate each proposed change individually while already understanding the solution deeply.