I've used ChatGPT, Claude, Gemini, Grok, Replit — pretty much everything. And for a long time every new tool felt like a cheat code.
Then one study stopped me cold.
The Finding Nobody Talks About
METR gave 16 experienced engineers the best AI tools available — Cursor, Claude, all of it — on codebases they knew deeply.
They predicted they'd be 20–24% faster.
They were actually 19% slower.
Same code quality. Just slower.
This isn't an argument against AI tools. It's a signal that there's a massive gap between having these tools and using them well.
Think of it like getting a high-end espresso machine. Same equipment as a pro barista. But without knowing grind size, temperature, and extraction time — you're just making expensive bad coffee faster.
The Real Numbers in 2026
Adoption is basically universal now:
95% of engineers use AI tools at least weekly
46% of all code is now AI-generated or assisted
GitHub Copilot alone has 20 million users
But here's the honest flip side:
Only 6% of engineering leaders saw significant productivity boosts
Faros AI studied 10k+ devs — AI teams merged 98% more PRs but review times ballooned 91% and bug rates rose 9%. Net impact? Near zero.
Positive developer sentiment has dropped to 60%, down from 70%+ in 2023
Generating code faster ≠ shipping better software faster. Most teams are measuring the wrong thing.
What Actually Works
The tools do shine — just not everywhere:
✅ Boilerplate and repetitive code — biggest win, consistently
✅ Navigating unfamiliar codebases — 25% faster (GitHub/Accenture data)
✅ Docs and code reviews — 35% faster reviews (McKinsey 2026)
✅ Mental energy — the cognitive load drop is real and underrated
The best workflow I've found: spec first, small tasks, curated context, review everything.
Addy Osmani from Google nails it: "Treat the AI as a fast but naïve junior dev — you're still the senior."
Here's my full blog breakdown Just select it and open it on Chrome.
Top comments (0)