The hidden collapse of human code review under AI volume

The core claim. Sonar CEO Tariq Shaukat argues that AI agents have broken the human-in-the-loop model. Code output has exploded 2,450% and commits 240%, but releases only rose 30%. Humans cannot keep up with this volume, and the traditional practice of having engineers review AI-generated code is failing.

The cognitive trap. Research shows reviewers agree with wrong AI output 80% of the time—a phenomenon called cognitive surrender. LLMs are also unreliable judges: identical questions get different verdicts 7-17% of the time, and simple challenges like "are you sure" change answers up to 34% of the time. The human safety net is riddled with holes.

The scaling problem. Agents struggle with large codebases. Translation accuracy drops from 91% at 10k lines to near useless at 50k lines. Token use varies 30x depending on context location, and errors compound: 95% accuracy per step becomes 36% after 20 steps. Long-running agent tasks are inherently fragile.

The vendor's answer. Shaukat proposes a factory model with four components: clear intent, guide rails, independent verification, and evidence. Sonar's products—Sonar Vortex and SonarQube—claim 36% lower token cost and 92% fewer issues per commit. Customer surveys show SonarQube users are 81% more likely to see positive AI coding impact on security.

The tradeoffs. Independent verification adds latency and cost. The model assumes a centralized gatekeeping approach that may slow development in fast-moving teams. Relying on a single vendor for governance creates lock-in. The talk does not explore alternative strategies like adaptive testing or model fine-tuning to reduce errors at the source.

What matters now. The observation that human review cannot scale with AI code generation is critical. Teams should treat verification as a non-negotiable second factor, not an afterthought. The specific failure modes—cognitive surrender, error compounding, unreliable AI judges—are real and measurable. Start auditing your review pipeline today.