FOMO-Driven AI Coding: Speed Illusion, Hidden Review Debt

GitHub's 2022 experiment showed Copilot users finishing a scoped JavaScript task 55% faster, while METR's 2025 study found experienced open-source maintainers working 19% slower with AI tools on their own mature repositories—despite predicting a speedup themselves. Both results are valid; the difference lies in who bears the verification cost. Greenfield tasks with clear specs make checking cheap, but in codebases full of implicit invariants, verifying plausible-but-wrong AI output becomes the real job.

DORA and GitLab data confirm this at the system level: most organizations report their bottleneck has shifted from writing code to reviewing it, and treat AI-generated code as a new category of technical debt. As generation got cheaper, volume rose while review capacity stayed flat, simply moving the queue from the IDE to the pull request stage—an outcome explained by basic queuing theory rather than model quality.

Benchmarks offer little stability either: SWE-bench scores went stale within months, and a hundred-line agent later outperformed last year's flagship product. FOMO-driven adoption tends to fund visible costs like seats and plugins while deferring invisible ones like test coverage and review staffing—debt that resurfaces as provenance issues, new attack patterns like slopsquatting, and public stumbles such as Microsoft's Recall rollback or Builder.ai's collapse.

The lesson isn't to avoid AI tools but to sequence adoption correctly: identify the real organizational constraint first—review latency, change failure rate—and measure whether generation speed actually affects it, using system-level metrics rather than IDE acceptance rates.