Individual developers are writing code faster than ever with AI tools. Their organizations might actually be moving slower because of it. That's the central argument of a new research paper published on Zenodo that introduces what the authors call "The Verification Paradox."
The logic is straightforward once you see it. AI coding assistants like Cursor, GitHub Copilot, and Claude Code let a single developer produce far more code per hour than they could manually. But every line of that code still needs to be reviewed, tested, and verified by humans. The bottleneck hasn't moved - it's just gotten worse.
More Output, Same Review Capacity
Think about a typical pull request workflow. A developer writes code, submits it for review, a colleague reads through it, flags issues, and approves or requests changes. AI tools dramatically increase the "writes code" step but do nothing for the review step. If one developer can now produce three times the code, that means three times the review burden on the rest of the team.
This isn't a theoretical concern. Teams adopting AI coding tools have reported pull request queues growing faster than they can be processed. Senior engineers - the people most qualified to review complex code - become the bottleneck, because their review capacity hasn't scaled alongside their team's output capacity.
The Trust Problem
The paper also raises a subtler issue. AI-generated code often looks correct and passes basic tests, but can contain logic errors that are harder to catch in review than hand-written mistakes. A human writing code from scratch leaves a trail of intent - variable names, comment patterns, structural choices that reflect their reasoning. AI-generated code doesn't carry that same signal, which makes review slower per line, not just per volume.
This creates a compounding problem: more code to review, and each unit of code takes longer to review confidently.
What Actually Helps
The research suggests organizations need to invest in verification infrastructure alongside AI coding tools. That means better automated testing, stronger CI/CD pipelines (automated build-and-deploy systems), and clearer ownership boundaries so code reviews don't pile up on a few people.
The practical takeaway for teams using AI coding tools: your limiting factor probably isn't how fast code gets written anymore. It's how fast you can confirm that code is correct. Buying everyone Cursor seats doesn't help if your review process was already a bottleneck. Fix the review pipeline first, then let AI tools increase throughput through a system that can actually handle it.