Science & Tech · AI · 12 hrs ago

81% of software teams say they merged AI code they did not fully understand

81% of software teams say they merged AI code they did not fully understand

A survey found that 81% of software teams merged AI-generated code they did not fully understand.

The findings cover software development teams and point to a gap between the growing use of AI coding tools and checks on their output.

Only 41% used automated checks before merging AI-written code.

Other teams relied on manual review, used dedicated quality assurance teams, or asked another AI model to judge the code.

Teams using an AI judge were more likely to ship code they did not understand: 29%, compared with 15% among teams that did not.

Just 13% had an incident response procedure for AI-generated errors in live projects.

The survey said verification has not kept pace with AI code output, and teams with independent checks saw fewer defects reach production; such checks are expected to become standard.

1 source

Related

What people are saying

Asad KhanTestMu AI co-founder and CEO
Verification, however, has not kept pace with that shift.
via TechRadar
We believe any output AI produces should be validated by a system independent of the model that produced it, before it merges and every time it changes.
via TechRadar
The teams in our survey that already follow this discipline see fewer defects reach production, and we expect it to become the standard for engineering organizations shipping with AI.
via TechRadar