Here's a statistic worth sitting with longer than it usually gets: in a controlled study, experienced developers using AI assistance on their own familiar codebases took 19% longer to complete real tasks than developers working unaided. Afterward, the same AI-assisted developers estimated they'd been about 20% faster. A roughly 39-point gap between how fast it felt and how fast it actually was.
That's not a study of unfamiliar tools on unfamiliar code — these were experienced open-source developers working in repositories they'd known for years on average. If the gap shows up there, it's not explained away by inexperience with the tool.
Why the feeling and the reality point in opposite directions
Using AI assistance changes the texture of the work, not just the speed. Typing less, waiting for generations, reviewing rather than producing from scratch — all of that feels different from heads-down manual coding, and "feels different" gets misread as "feels faster" surprisingly easily. The friction that used to signal "this is taking a while" (staring at a blank function, working through a bug by hand) mostly disappears, even when the actual clock time working through review cycles, correcting agent mistakes, and re-prompting ends up comparable or worse.
Why this matters more than the raw speed number
If the 19% slower number stood alone, it would just be a caution about a specific study's conditions. What makes it genuinely important is pairing it with the perception gap: developers are not just occasionally wrong about whether AI made them faster — the research suggests they're systematically wrong in one direction, overestimating gains fairly consistently. That has a direct implication most people miss: if you can't trust your own sense of whether AI use made you faster, you probably also can't fully trust your own sense of whether it's quietly eroding a skill in the background. Both are judgments about your own performance, and this is specifically the kind of judgment the research says is unreliable.
What to do with that, practically
Track something concrete instead of relying on how a task felt: actual time from start to done on comparable tasks, with and without heavy AI assistance, logged rather than estimated afterward from memory. It's unglamorous, but it's the only way to route around a bias the research says most people share.
Get an external read instead of trusting your own sense of it →
Source: METR (Model Evaluation & Threat Research), randomized controlled trial of experienced open-source developers, February-June 2025 — also referenced in Is AI Making Developers Worse? on this site.
Related reading: Is AI Making Developers Worse? · MIT Study: Cognitive Debt