Nearly all software teams admit to shipping AI-generated code as study finds verification 'has not kept pace with that shift'

Computer programming code. Programming code abstract technology background of software developer and Computer script.
(Image credit: Shutterstock/BEST-BACKGROUNDS)

  • The majority of software development teams have merged code with AI contributions that they didn’t understand
  • Some teams (55%) manually review code, while others use one AI model to review another, which has mixed results
  • TestMu AI’s State of AI in Quality Engineering 2026 report has also found that only 13% of teams have an incident response provision for publishing AI-generated errors

Generative AI coding tool shave been wholeheartedly embraced by software development teams, to such an extent that 81% have admitted to merging AI generated code that wasn’t fully understood, new research has claimed.

TestMu AI’s State of AI in Quality Engineering 2026 report also found a mixed approach to reviewing AI code and very few teams employing an incident response procedure in the event of AI generated errors occurring on live projects.

These results raise some questions on how development teams are using AI and what steps are being employed to verify output. While some teams use manual review, others rely on a second LLM to review and judge the code. However, this appears to have increased the likelihood of code no-one understands being published.

Latest Videos FromTechRadar

LLM-as-a-judge

Revealed in the report is the notion of the verification gap, a challenge for teams relying on AI-generated code. Just 41% of teams use automated checks of AI-written code before merging it with existing, often live code. This gap is the beginning of the problem – there is simply too much code to efficiently check.

With 55% of development teams using manual review, or passing to a dedicated QA team (51%), it’s clear that with some teams, important measures are in place.

Employing a second LLM to grade the generated code might seem like a smart move. In terms of fixing issues, it probably is. But the result of this strategy is that teams using LLM-as-a-judge are shipping code they don’t understand at a greater rate – 29%, compared with 15% who don’t use AL grading.

Verify your output

Asad Khan, the Co-founder and CEO of TestMu AI said the key point that verification of AI-generated code is not keeping pace with the volume of output.

“Software teams now rely on AI to write their code, generate their tests, and increasingly determine whether both are ready to ship,” he noted.

“Verification, however, has not kept pace with that shift. 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.” The development teams that have put in place the procedures to do this are doing better, it seems.

“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,” Khan added.


Google logo on a black background next to text reading 'Click to follow TechRadar'

Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.


TOPICS
Christian Cawley

Christian Cawley has extensive experience as a writer and editor in consumer electronics, IT and entertainment media. He has contributed to TechRadar since 2017 and has been published in Computer Weekly, Linux Format, ComputerActive, and other publications.

He currently heads up the team at smart home website Matter Alpha, and writes about retro gaming at Gaming Retro.

You must confirm your public display name before commenting

Please logout and then login again, you will then be prompted to enter your display name.