AI code-volume stats are just lines of code with better PR

The software industry spent two decades establishing that lines of code is a terrible way to measure developer value — then quietly resurrected it. Google’s ‘75% of new code is AI-generated,’ Anthropic’s and OpenAI’s ~80% figures, and Cursor’s ‘100M+ lines per day’ are all volume claims dressed up as progress. Unlike GitHub’s earlier Copilot pitch (tasks completed 55% faster), which was a falsifiable outcome claim, these adoption percentages can only go up — they say nothing about delivery speed, incident rates, or customer value. Unsurprisingly, every company making them is an AI vendor with adoption numbers to pump.

The shift happened because the outcome evidence got messy. Cui et al. found a 26% task-completion boost across ~5,000 developers, but GitClear documented rising churn and collapsing refactoring, and METR famously found experienced devs were 19% slower with AI while believing they were faster — a result METR largely walked back in February 2026, conceding it can no longer cleanly measure the effect because developers refuse to work without AI. At the firm level, an NBER survey of ~6,000 executives found roughly nine in ten companies report no measurable productivity impact, with cross-study consensus around 10% gains. Anthropic itself embodies the contradiction: its marketing touts ‘8x more code shipped’ while its research arm published an RCT showing AI-assisted developers understood their own code 17% worse with no significant productivity gain.

The stakes are concrete: these vanity metrics are underwriting layoffs. Block cut over 40% of its workforce with AI as the explicit thesis — while reporting strong, growing gross profit — and Atlassian followed with a 10% cut. The author argues that if AI genuinely freed up capacity, companies with endless roadmaps would ship more, not shrink; choosing layoffs suggests the productivity claim is post-hoc cover for decisions made on other grounds. His prescription isn’t AI skepticism — he advocates daily AI use and thinks resisting it is career suicide — but a return to battle-tested outcome measures like DORA metrics, reliability, and revenue instead of adoption ladders whose top rung is ‘buy more of our product.’