How to review AI code without reading the code

The argument over whether engineers should stop reading AI-generated code has been resolved, as demonstrated by Faros AI's report, which shows a significant increase in merged pull requests without review. AI review is now used for 50% to 80% of PRs to find bugs, but it falls short in capturing the crucial discussions about decisions made during development.

Teams often fall into two modes: excessive line-by-line review or complete rubber-stamping, neither of which provides a balanced approach. AI review excels at the mechanical aspects of coding but lacks the ability to understand the team's decision-making process and the rationale behind those decisions. To address this gap, a new review framework has been proposed that involves five layers: Challenge, Capture, Codify, Debate, and Own.

The first three layers are handled by machines, while the latter two require human intervention. By implementing Debate and Own, teams can foster discussions around the decisions made by AI, ensuring that they are building the right system and not just following the code. This approach eliminates the argument over whether engineers should read AI-generated code, as humans are better suited to evaluate the correctness of decisions, assess trade-offs, and take ownership of the final outcome.