AI Didn't Make You Faster. It Moved Where You Get Stuck.

A two-year study followed 800 developers and found this: people using AI delete and undo their own code 14 times more often than people who don't. In other words: AI didn't make you faster. It just moved where you get stuck.

Why "faster" and "slower" are both true

Both findings describe the same mechanism from two angles. AI genuinely collapses the time to produce a first draft, that part of the speed feeling is real, I am not going to pretend otherwise. What it does not collapse is the time to verify that draft is correct, and every AI-written line still needs the same review, the same tests, the same understanding a human-written line needs, arguably more, because you never built the mental model of it while typing. The bottleneck did not disappear. It moved from typing to reviewing and undoing.

What this actually predicts about code quality

• Industry data from 2026 shows AI-generated code carrying meaningfully more vulnerabilities per line than human-written code in the same codebases
• Independent evaluation of AI code review tools found wide variance in what they actually catch, some missing over 40% of real vulnerabilities in benchmark testing
• The large majority of developers surveyed in 2026 say they don't fully trust the functional accuracy of AI-generated code, even as adoption keeps climbing

Adoption and trust moved in opposite directions at the same time, and that is not a contradiction, that is an accurate read of the tool. It is useful enough to keep using and unreliable enough to keep double-checking.

How to actually work with this

Budget review time as part of the task, not as overhead you eliminated by using AI, because you did not eliminate it, you just pushed it later in the process. Treat an AI-generated diff exactly like a pull request from a junior who types very fast: read every line, run every test, do not merge on vibes. The teams getting a real, measured speed-up are the ones who shortened the loop between "AI drafts it" and "a human or a real test catches what's wrong," not the ones who skipped that loop and called it velocity. Where did your bottleneck just move to?