# You feel faster. You're probably slower. A careful study put experienced developers on real tasks with AI tools. They expected to be 24% faster. They were actually 19% slower — and still believed AI had sped them up. Meanwhile teams ship 98% more pull requests but review time jumps 91% and company-wide delivery doesn't move. The AI productivity story has a hole in it, and it's not that AI is useless. It's that we sped up the one part that was never the bottleneck, and confused the feeling of speed for the real thing. - **Author:** Fedor Chishchin - **Published:** 2026-06-08 - **Tags:** methodology, careers - **Source:** https://fedorthinks.com/blog/you-feel-faster-youre-slower --- Here's a result that should make every AI-coding enthusiast pause. In a careful study, experienced developers worked on real tasks from their own projects, with and without AI tools. They predicted AI would make them about 24% faster. The measured outcome: [they were 19% slower with AI](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) — and even after living through the slowdown, they still believed the AI had sped them up by 20%. Sit with that gap. Not "AI didn't help as much as hoped." AI made them _slower_, and they couldn't feel it. That's a roughly 40-point hole between perception and reality, measured on experienced people doing real work. And it lines up with what companies are seeing at scale: massive individual activity, flat results. The question worth thinking about isn't "is AI fake?" It clearly isn't. It's _why the feeling and the outcome point in opposite directions_ — because the answer changes how you should use it. ## AI removes the friction you can feel, not the work that's hard Here's the first half of the explanation. AI is brilliant at the parts of coding that _feel_ like work but aren't where the difficulty actually lives — boilerplate, syntax, the tedious typing. When that friction vanishes, you feel fast, because the annoying part got easy. But the boilerplate was never the hard part. The hard part is understanding the problem, holding the system in your head, and deciding whether a given change is right. AI doesn't remove that — it often _adds_ to it, because now you also have to read, check, and fix code you didn't write. As the study's reading goes, developers ended up [spending much of their time cleaning up AI output instead of shipping](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/). The work moved from "write it" to "review it," and reviewing someone else's plausible-looking code is slower and less satisfying than people expect. You feel fast because the visible friction is gone. You're slow because the invisible work grew. ## You can't merge your way to value The second half shows up at the team level, and it's even more important. Faros AI measured teams with heavy AI adoption and found a striking shape: [98% more pull requests merged, but PR review time up 91%, PR size up 154%, and bugs up 9% — while DORA delivery metrics stayed flat](https://www.faros.ai/blog/ai-software-engineering). More code, more PRs, more activity. Same actual delivery. This is just a bottleneck problem wearing new clothes. If you speed up one stage of a pipeline — code generation — but not the stage after it — human review — you don't get more throughput. You get a bigger pile waiting at the slow stage. AI made writing code cheap, which means there's now far more code to review, by the same number of humans, who are now the constraint. The value the AI created gets absorbed at the review queue and never reaches the business. That's why 89% of executives say AI made work faster while only [6% can point to real organization-wide returns](https://www.faros.ai/blog/ai-software-engineering). Everyone's busier. Nothing ships faster. ## What this should change about how you work The mistake isn't using AI. It's optimizing the wrong stage and measuring the wrong thing. A few corrections fall out: - **Measure outcomes, not output.** PRs merged, lines written, "I feel faster" — these are vanity metrics now, easy to inflate and disconnected from delivery. Did the thing reach users and work? That's the only number that survived. - **Treat review as the bottleneck it became.** If AI tripled the code, your review and testing capacity is what decides whether any of it turns into value. Invest there — smaller PRs, better evals, automated checks — instead of generating yet more. - **Distrust the feeling of speed.** The study's scariest finding is that smart people couldn't perceive their own slowdown. Assume your sense of "this is faster" is unreliable, and check it against something real. - **Use AI where it genuinely wins.** The same research shows big gains for juniors on simple, unfamiliar tasks and near-zero or negative ones for seniors on complex code they already know. Aim it at the first case, be skeptical in the second. ## The bottom line The AI productivity paradox isn't a paradox once you see it clearly. AI sped up generation — the part that felt like work — and left untouched the parts that are actually slow: understanding, reviewing, judging. So individuals feel a rush of speed while the real bottleneck quietly gets worse, and the company's numbers don't move no matter how many PRs fly by. This is the same lesson from a different angle as [the demo never being the hard part](/blog/the-demo-was-never-the-hard-part): the generation got cheap, the judgment didn't. If you want AI to actually make you faster — not just feel faster — point it at the work that's genuinely tedious, then spend the time you saved on the part it can't do: deciding whether what came out is any good. The feeling of speed is free now. Real speed still has to be earned at the slow stage, and that stage is you.