In 2025, METR, an independent AI safety and evaluation research organisation, ran a controlled study on professional software engineers using GitHub Copilot. The result was not what the vendors wanted. Experienced developers completed tasks 19% slower with Copilot than without it. They also felt faster. That gap between perceived productivity and actual productivity is the most expensive number in enterprise software right
1. What the METR study actually measured (not what GitHub's marketing says)
2. Why "40% of code written by AI" is true and meaningless at the same time
3. The acceptance rate lie: 30–35% for Copilot vs. 90%+ for Cursor and why higher acceptance ≠ better code
4. What this means for VC portfolio companies that just bought GitHub Enterprise + Copilot
5. How Cursor's $9B valuation is bet on being the antidote and why their own 500-request quota throttles the developers who need it most
6. The second-order effect: if AI tools slow experienced devs, who do they actually help?
Every week I publish one finding like this. The kind that never makes the vendor deck. Subscribe.