As AI takes on more implementation work, we also need to look at what happens to code review. Producing more code means there is more to verify, understand, and take responsibility for. Several industry reports now give us a clearer picture of this development.
Let’s look at a few findings:
- More changes: GitHub reports 23% more monthly merged PRs in Octoverse 2025. Faros reports a 16.2% increase in PR merge rate per developer between organizations’ lowest and highest AI-adoption periods in its 2026 report.
- Larger changes: Greptile reports 79% growth in median PR size between June 2025 and June 2026, while DX’s Q3 2026 summary reports 122% year-over-year growth.
- Longer review waits: DX reports 105% longer waits for first review in organizations with 80%+ versus 20–40% AI code during Q3 2026. Faros reports a 156.6% increase in median time to first review between lowest and highest AI-adoption periods.
- More changes without review: Faros’s Q3 2026 analysis reports a 76.3% relative increase in PRs merged without review. This does not mean that 76.3% of all PRs go unreviewed.
Those reports examine different teams, periods, and measurements. We cannot conclude that AI alone caused these increases. But together, they point to growing pressure on the way we review changes.
How can review adapt?
I explore what this means for human oversight and the traditional peer-review model in Why the Traditional Code Review Model Is Under Strain.
This development is one of the reasons I am working on SCOPE: Staged Change Oversight with Proportionate Escalation. SCOPE is a new model for guiding code review in environments with increasing AI-generated changes.
The central idea is that not every change needs the same oversight, and that review starts before implementation. SCOPE combines review by agents, developers, and independent peers, with the depth and involvement adjusted to the change’s risk and alignment needs.
As changes become more numerous and larger, this gives teams a way to decide earlier what needs careful verification, where shared understanding matters, and what additional human review should contribute. The goal is to use our limited attention where it brings the most value.
