Last week at Cloud Next, Sundar Pichai said seventy-five percent of new code at Google is now written by AI. Eighteen months ago the number was twenty-five. Most of that climb happened in the last eleven weeks.
At Anthropic and OpenAI, senior engineers are saying it is one hundred percent. Boris Cherny, who runs Claude Code, says pretty much one hundred percent of Anthropic's code is AI-generated. Roon at OpenAI says he does not write code anymore. Meta is targeting fifty percent within a year; some internal orgs are already between fifty and eighty. Microsoft is at twenty to thirty percent, with some repos past fifty. Amazon is shipping seventy-nine percent of auto-generated code without edits.
These are not marketing lines. They are numbers from the CEOs of the companies that employ the most engineers on earth, and from the technical leads of the labs that train the models those companies use. The methodology varies across companies. The direction does not.
The shift was not the model. It was the agent.
Something specific happened between February and April of 2026. Until late 2025, AI wrote code as autocomplete. A sentence here, a function there. Useful, but still a human at the keyboard doing most of the typing.
Then coding agents went mainstream. Claude Code. Cursor. OpenAI Codex. Copilot Workspace. The contract changed. Write a ticket, get a pull request. The human stopped writing lines and started writing tasks.
Cursor hit two billion dollars in annualised revenue by February. GitHub Copilot crossed twenty million users. These are not niche adoption curves. These are the tools the best engineers use every day. That is how fifty percent becomes seventy-five percent in a single quarter. Not a better model. A different mode.
The flip
The old contract was write and review. An engineer produced the first draft. Another engineer checked it. Both were authors. Both left their fingerprints.
The new contract is direct and review. The model produces the first draft. The human accepts, rejects, or redirects. The model is the author. The human is the editor.
The labour moved one step up the stack. The writer's chair became the editor's desk.
This is not a new move. When the camera arrived, illustrators did not disappear. They moved up the stack, to interpretation, abstraction, editorial framing. When the model arrived, engineers moved up the stack, to specification, review, architecture. Same pattern, different tool. Every time the drafting got mechanised, humans moved to judgement.
So what do five thousand Anthropic engineers do all day?
This is the fair question. If the model writes one hundred percent of the code, what is the job?
Spec the system. Review the output. Architect the trade-offs. Test what matters. Debug what slips. Decide what to ship next. None of that scales linearly with output. More code means more reviewing, more directing, more judgement calls.
Anthropic employs around five thousand people to do this. The human work did not get cheaper. It got denser. A team of editors producing the output of an army of writers.
This is not just about code
Every creative role has a first-draft task. The strategist writes the deck. The brand director writes the campaign. The copywriter writes the page. The designer drafts the layout. The lawyer drafts the clause. The analyst drafts the model.
In every one of those roles, the first draft is a cost, not a product. It is what has to exist before the judgement can begin. Whatever can produce that draft faster will. Whatever can produce it for free will win.
Adobe Firefly generated twenty-two billion creative assets in its first two years. I have shipped twenty plus AI-native products in the last eighteen months. I did not write the first draft of any of them. The design curve is already well up. The strategy curve is right behind.
The ceiling varies. Novel research still hits limits. Boilerplate does not. Taste-dependent work lands somewhere in between. But every job has a first-draft task. And every first-draft task is on the curve.
The skill is recognition, not specification
Someone asked me what I actually do when I work with Claude. Not what the model does, but what I do. The answer surprised me when I said it out loud.
I do not always know exactly what I want up front. But I know when the output is not it. And that is enough.
Reject. Redirect. Refine. Over and over. Not because I can describe the right answer in advance, but because I recognise the wrong one the moment it arrives.
That is what editorial work has always been. A newspaper editor does not sit down and describe the perfect headline. A film director does not pre-visualise the final cut. A creative director does not specify the exact layout. They recognise when something is not right and iterate until it is. The skill is taste applied under time pressure.
Taste does not feel like a skill until the volume of options goes up. Then it becomes the only skill that matters.
"Edited by Humans" is the new quality stamp
Here is the opportunity most people are missing. When the machine can write the first draft of everything, human judgement becomes the premium. Not the volume. The filter.
Grass-fed. Hand-stitched. Small-batch. Every premium label started as an ordinary way of working, until mass production made the ordinary way feel like the exception. The same is about to happen to creative work. The stamp will not go on the thing the machine made. It will go on the thing the human decided to ship.
Edited by Humans becomes a quality signal. A badge on a landing page. A footer on a case study. A line in a CV. The work that went through a judgement layer, made by someone with the taste to reject ten drafts and approve one, is the work worth paying for.
For strategists, brand directors, CMOs, and heads of anything creative, this is the most flattering thing AI has ever done to your job. You were already editorial. You already reviewed work and directed drafts and spotted what was wrong. The only thing that changed is the volume coming through.
For the people who learn to edit early, the curve is a promotion. For the people waiting for first-draft work to come back, the curve is something else. That split starts now.
Worth stamping.
