"60% of New Code Is AI-Generated." I Went Looking for the Source. It Isn't There.

There's a statistic doing laps through every 2026 trend piece, LinkedIn carousel, and conference deck: Gartner forecasts that 60% of new code will be AI-generated by the end of 2026. I write next to these tools all day, so the number didn't strike me as impossible — it struck me as suspiciously round. I went looking for the primary source. Short version: I can't find it, and I don't think it exists.

The chase

Gartner's newsroom is searchable, and its traceable predictions in this territory are specific: 75% of enterprise software engineers will use AI code assistants by 2028, up from less than 10% in early 2023. More than 80% of enterprises will have used generative AI APIs or deployed GenAI apps by 2026 — a prediction published in October 2023. And this May, a release describing the enterprise AI coding agent market as entering "a new phase of expansion and competitive realignment." All real, all linkable. None of them says 60% of code.

The 60% figure lives exclusively in secondary sources — trend roundups citing other trend roundups, with the Gartner attribution attached somewhere along the chain and no report name, date, or link. That's citation telephone: a plausible number acquires a brand-name source because the brand makes it credible, and then it's unkillable.

What the traceable numbers actually say

If you want defensible statistics about AI-generated code in 2026, these survive contact with their sources: aggregated estimates put around 41% of all code as AI-generated. 76% of professional developers are using or planning to use AI coding tools — 62% using, 14% planning. 82% in one survey report daily or weekly use. Developers self-report 30–60% time savings on coding, testing, and documentation. And the one I keep coming back to: about 75% say they still manually review every AI-generated snippet before merging.

Every one of those carries methodology caveats — self-reported, differently sampled, measuring different things. But they share a property the 60% stat lacks: you can find out who measured them, and how.

Why "% of code" is unmeasurable anyway

Here's the deeper problem: what counts as AI-generated? A tab-completion accepted mid-keystroke? An agent-written file I rewrote a third of? On my own repos, Claude Code writes most first drafts and I rewrite the load-bearing parts — whose line is line 42? Any single percentage laundering that ambiguity into a headline is marketing exhaust, whoever it's attributed to. The honest instruments are per-repo: commit trailers, agent-branch statistics, git blame sampling — and even those give you a floor and a ceiling, not a truth.

The stat that actually matters

The 75%-review-everything number is the real story of 2026: the bottleneck moved from writing code to trusting it. Which is exactly the argument I field-tested in my Matt Pocock post — when generation is free, fundamentals and review get more valuable, not less. It's also why 70% of developers now run two to four coding agents at once — generation is cheap enough to parallelize; verification isn't. If you're planning a team or a toolchain around one number this year, plan around review capacity, escaped-defect rate, and time-to-trust — not authorship percentage.

The verdict

Adoption is real, and the traceable numbers are impressive without embellishment. The 60% stat, as far as I can establish, is exhaust — a round number wearing a borrowed suit. The sentence I'd actually plan 2026 around costs nothing and survives every methodology fight: every line of code is getting cheaper to write and more expensive to trust.

Sources

Gartner's traceable predictions: the 75%-by-2028 code-assistant forecast, the 2023 enterprise GenAI prediction, and the May 2026 coding-agents release. The aggregated adoption and review statistics are collected in NetCorp's AI-generated code statistics roundup. And an example of the 60% claim in the wild: this dev.to trends piece. If anyone can point me at a primary Gartner document with the 60% figure, I'll happily append a correction — that's rather the point of the post.