85% say the bottleneck is reviewing AI code, not writing it. @gitlab asked 1,528

AlphaSignal argues that AI has moved the slow step in software work from writing code to checking it. Its post cites a GitLab survey of 1,528 developers and buyers in which 85% named reviewing AI code as the bottleneck rather than writing it. It also cites a Sonar survey of more than 1,100 people: 42% of committed code is AI-generated, 96% do not fully trust that it is functionally correct, and only 48% always check it before they commit. AlphaSignal relays these figures secondhand. The surveys themselves were not reviewed here, and they record what respondents said rather than measuring review queues.

The proposed answer is to let automated checks, instead of human reading, decide whether a change is acceptable. AlphaSignal points to Uncle Bob Martin, who it says ships code written by agents without reading it, because a set of checks can reject that code without him: unit tests he never reads, acceptance tests he spot-checks, mutation testing on the test suite, and quality metrics on the output. Mutation testing plants small deliberate bugs in the code to see whether the tests catch them, which shows whether the tests themselves can be trusted.

The general rule the post draws is that a change earns autonomy when something cheap can fail it, and that reading every line only makes the reviewer the new queue. It leaves teams with three questions: when a change can ship unread, who is accountable if it ships wrong, and what fails if nobody reads the diff. This is an argument built on one practitioner's setup. The post offers no data on whether code shipped unread under such checks has more or fewer defects, and neither the full breakdown it links to nor its attached image was reviewed.