AI broke code review. What about knowledge sharing?
The New Stack Ankit Jain ● Covered by 3 sources
AI is producing so much code that human review is breaking. The real risk isn’t just bugs — teams may stop understanding what they built.
Based on reporting by The New Stack, Ankit Jain — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AI-generated code has pushed code review past the point where a person can realistically read every line. Engineers are now staring at huge diffs they didn’t write, made by models they don’t fully control, and the old habit of treating review as the place where knowledge spreads through a team is under strain.
That knowledge-sharing role has always mattered. Review is where junior engineers see how senior ones think, where architectural choices get challenged, and where shared ownership starts to form. If AI moves the work earlier, the learning has to move earlier too. Otherwise the team keeps the speed and loses the understanding.
Vanitha Kumar at ThoughtWorks stumbled into that idea while building an agent to catch anti-patterns and explain them to a developer outside her team’s usual archetype. She set it up around a reference document and had it flag deviations and explain why they mattered. What started as a teaching tool slowly turned into a review tool, which is probably the right order of operations anyway.
The deeper problem is cognitive debt: technical debt lives in the code, but cognitive debt lives in people. Margaret-Anne Storey saw it in students building with AI. They were moving quickly, until they realized they couldn’t keep changing the product because they’d lost track of what they were building, why they were building it, and who actually understood it. In one case, only the person supervising the AI had a grip on the code, and even that person didn’t fully understand what had been generated.
The answer is to review intent, not just output. Instead of waiting for a 600-line diff, teams should capture scope, constraints, acceptance criteria, and what’s explicitly out of scope before the change is submitted. That gives reviewers something better to do than squint at code: they can judge whether the team is solving the right problem under the right conditions. And it keeps the author from handing over their own understanding to the machine.
That’s the bit many teams still miss. AI can crank out code, but it doesn’t automatically preserve the mental model that makes a team durable. If companies want the productivity without turning the codebase into a black box with nice comments, they need to move knowledge-sharing upstream, not pretend review can survive as a reading exercise.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI coding everyone likes to skip past: speed is easy, shared understanding is hard. A lot of teams are clearly happy to let one person become the translator between the model and the rest of the humans, which is a lovely way to rediscover old-fashioned bus factor problems with newer branding. Better to force intent into the process now than wait for the first mystery feature nobody can safely touch.
Read more about this at: The New Stack