31.3% more PRs are now merging with no review at all. Not no human review. No review, period. Nobody decided that was fine. Reviewers just can’t keep up with the pace what AI is producing.
The AI Engineering Impact Report 2026 is two years of telemetry across 22,000 developers and 4,000+ teams. Real data. The throughput gains are real too: epics per developer up 66%, task throughput up 33.7%. But downstream, things are breaking.
Bugs per developer: up 54% (it was 9% in their 2025 report). Incidents-to-PR ratio: up 242.7%. Code churn: up 861%.
The review problem
AI-generated code looks right. Clean naming, idiomatic style, consistent with the surrounding codebase. The bugs sit underneath, in logic and edge cases the model didn’t think through. Finding them means reconstructing intent, not scanning for obvious errors. That’s slow and expensive work.
Median time in code review is up 441.5%. It falls on the engineers who know the system best, because junior reviewers can’t catch what’s wrong with code that looks this polished. Senior engineers are buried. And a growing slice of the code ships without anyone looking at it at all.
Surveys vs. telemetry
DORA’s 2025 report says strong engineering foundations protect you from AI’s downsides. That came from survey responses. Faros’s telemetry says high-DORA orgs are experiencing the same deterioration as everyone else.
Surveys capture how people feel. Right now developers feel productive, because at the individual level they are. The incidents show up downstream. Telemetry catches that. Surveys don’t.
If you’re cutting engineers because AI output numbers look good, you’re removing the people absorbing the quality gap AI is creating.