Confidence in AI-generated code
has outpaced the practices that verify it.
We surveyed engineers, engineering leaders, and business leaders on what happened to software reliability after their teams adopted AI tooling. This is what they told us.
of teams are confident shipping AI-generated code to production.
That confidence holds even though the picture underneath it doesn't look good. Reliability has stalled or slipped for most teams since adopting AI, and 84% have already had an AI-related incident, or admit it's only a matter of time. Most aren't even tracking AI-generated work separately from human work, so when something breaks, there's often no way to trace it back.
The confidence isn't unfounded, exactly. It's just resting on a foundation that hasn't been inspected yet.
“A significant portion of the code was generated with AI. When developers revisit the code later, they may not fully understand the reasoning behind certain implementation decisions. They can explain what the code does, but not why a particular approach was chosen.”
You can't manage what you don't measure.