AI Slop Is Changing How Engineers Review Code
IEEE Spectrum Aaron Mok ● Covered by 3 sources
AI can spit out thousands of lines fast, but engineers now have to sift through the slop. The surprise: the time saved on writing code can vanish in review, security checks, and cleanup.
Based on reporting by IEEE Spectrum, Aaron Mok — 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 coding tools have made it easy to generate a mountain of code in minutes. The hard part hasn’t gone away; it’s just moved. Teams still have to inspect what the models wrote, and that code can hide bad assumptions, security holes, or errors that only show up after deployment. If those problems take too long to catch, the productivity boost starts to look thin fast.
That is pushing code review into a new shape. Some teams are writing detailed plans before an agent types a single line. Others are handing the first pass to specialized AI reviewers, then sending only the risky stuff to humans. And some are making developers explain, in plain language, why the code their agents produced should be trusted at all.
The pressure is real. In a Sonar survey of more than 1,100 developers, respondents said AI made up 42 percent of the code they added to shared codebases. But trust lagged badly: 96 percent said they did not fully trust the output to work correctly. Sonar also found that 38 percent of developers felt AI-generated code took more effort to review than code written by colleagues, and 61 percent said it often looked correct while still being unreliable.
Synthesia has lived that shift in a very visible way. The AI video platform’s 118 engineers went all-in on Claude Code in November 2025, and by August its pull requests were up 120 percent year over year. Ninety-five percent of those requests contained AI-generated code. CTO Peter Hill says the team has seen repeated duplication, with AI writing several versions of the same function because it lacks context. In one case, the company found as many as 10 versions of the same function, which meant engineers had to clean up the mess and then retrain the agent.
Amazon is tackling the problem earlier in the process. McLaren Stanley, a senior principal engineer at Amazon Stores, says his team now spends more time specifying what code should do before generation begins. That matters because one missing instruction once sent an agent off to produce 25,000 lines in the wrong Swift version, leading to 600 errors. After the specification was fixed, the code regenerated correctly in 15 minutes. AWS, meanwhile, uses agents to test code, compare it with the original plan, and check for security flaws before a human sees it.
The broader lesson is blunt: AI is shifting engineering work from writing code to judging code. That sounds efficient until review becomes a rubber stamp for output nobody really reads. Companies now know it, which is why they’re building more gates, more checks, and more ways to make the humans own the result.
My take — AI-written commentary, not fact-checked reporting
This is what AI coding was always going to do: move the mess downstream and call it productivity. The clever teams aren’t pretending the model is trustworthy; they’re building rituals to keep it on a leash. That’s the real story, not the demos or the line counts.
Read more about this at: IEEE Spectrum