‘Lines of code got a better publicist’: a critique of AI coding metrics

The AI industry's headline numbers this year are volume claims, not outcome claims: “percent of code written by AI” is lines of code with a better publicist, argues engineer David Curlewis in a post published on 10 June.

The numbers being quoted

He lists the 2026 claims: Google says 75% of new code is AI-generated; Anthropic says about 80% of merged production code comes from Claude and that engineers ship “8x more code per quarter”; OpenAI reports roughly 80%; Cursor advertises “100M+ lines of enterprise code written per day”. All are AI vendors, for whom rising adoption matters commercially.

The contrast with earlier claims is the point: GitHub's flagship Copilot claim was that developers completed tasks 55% faster — an outcome that could be tested and shown wrong. A volume number can only disappoint if adoption stalls.

What the evidence shows

The research record is mixed: the strongest pro-adoption result, Cui et al., found nearly 5,000 developers completing 26% more tasks, with the biggest gains for junior developers. But GitClear showed code churn rising and refactoring collapsing as Copilot adoption deepened, and METR found experienced open-source developers 19% slower with AI while believing they were 20% faster — a finding it walked back in February 2026.

An NBER survey of about 6,000 executives found 69% of firms using AI, with roughly nine in ten reporting no measurable productivity impact; the cross-study consensus sits near 10% organisational gains. Anthropic itself published both the “8x more code shipped” claim and a trial in which AI-assisted developers scored 17% lower on comprehension of code they had just shipped. He is equally critical of the AI maturity ladders sold by vendors and consultancies.

Why the numbers matter

Because they move budgets and headcount. In February, Jack Dorsey cut more than 40% of Block's workforce, over 4,000 people, citing AI as the core thesis: a smaller team using the tools the company builds can do more and do it better — while the business was strong and gross profit growing. Weeks later Atlassian cut about 1,600 people.

Curlewis asks for evidence that a share of a workforce is genuinely idle because fewer people can do the work, and notes he has never seen a product company without an endless roadmap: extra capacity should show up in usage, conversion and revenue.

The post is not anti-AI: Curlewis says every engineer should use AI daily. But adoption, he insists, is the starting line, not the scoreboard — delivery is measured by DORA metrics, reliability, rate of meaningful change and, ultimately, revenue. “Be AI-first in how you work, but battle-tested in how you measure it,” he writes.