The Economist says humans need to get better at editing. I think we need to go further.

The Economist published a leader on 30 July with a headline that caught my attention immediately: “AI is getting better at writing. Humans must get better at editing”.

I agree with the direction of the argument.

I think it stops one step too early.

I should be honest about my position here, because I have made a version of this case on this site four times since June — that judgement is becoming the competitive advantage, that the words were never the whole value. What is different this week is not the argument. It is that somebody outside my own head, with 1.2m words of analysis behind them, has arrived at half of it independently. That is worth engaging with properly rather than claiming as validation.

Because the half they stop at is the interesting part.

As AI gets better at writing, the scarce human skill isn’t simply editing.

It’s judgement.

The deeper shift is this:

AI gets better at writing
    ↓
writing becomes abundant
    ↓
production becomes cheap
    ↓
judgement becomes scarce

That matters because most conversations about AI and writing are still focused on the output. Does it sound robotic? Did it use too many em dashes? Is it too verbose? Does it have that strangely enthusiastic ChatGPT tone?

Those are real problems. They are also, on the evidence, temporary ones.

The tells are already going

The Economist’s own analysis makes this point better than I could. It compared its journalism against versions written by ChatGPT, Claude, Gemini and Grok — 55,940 sentences and 1.2m words in total — and checked the results against work from CNN, the New York Times and the Washington Post to be sure it was detecting machine habits rather than its own house style.

The em dash finding is the one that should worry anybody confident they can spot AI writing, because it is the tell everybody thinks they know. It no longer works. Of the models tested, only Claude uses em dashes more often than human writers. ChatGPT, the model that taught everyone to be suspicious of them, now uses them less than any other writer in the study. The reliable signals turn out to be quieter and much harder to spot in a single paragraph: fewer commas, fewer semicolons, almost no parentheses, long sentences, a heavy reliance on “and”.

So the detection heuristic most people are using is roughly two model generations out of date, and the ones replacing it are the kind of thing you need a corpus to see.

That seems like the shape of things from here. Models will get better. Prompts will get better. AI-generated writing will become harder to distinguish from competent human writing.

At which point “make this sound less like AI” stops being an interesting problem.

The bigger question becomes: was this worth writing in the first place?

Writing is becoming abundant

For most of the history of professional publishing, producing the words was the expensive part. A decent 1,000-word article required time, research and a reasonably skilled writer.

That has changed, and the book market shows it at a scale that is hard to argue with. Work by the economists Imke Reimers and Joel Waldfogel found that new ebook releases on Amazon ran at roughly 100,000 a month before ChatGPT and passed 300,000 a month by late 2025. When they ran AI detection across a sample of more than 50,000 titles, the detected share went from close to zero in 2022 to 30% in 2023, 45% in 2024, and past 60% by the end of 2025.

That is not AI “playing a role” in publishing. That is most of a market, in three years.

We are heading towards a world where the supply of content is effectively unlimited. Human attention isn’t.

Which changes where the value sits. If anyone can generate 1,000 plausible words, generating 1,000 plausible words is no longer much of a competitive advantage.

The advantage moves upstream.

What do you know that somebody else doesn’t? Which evidence matters? What is the argument? What is merely plausible and what is actually true? What should be challenged? What should be left out? What is distinctive enough to deserve somebody’s attention?

Those aren’t primarily writing questions. They are editorial judgement questions.

Editing the prose isn’t enough

The Economist offers sensible advice for working with AI. Commission clearly. Check the details. Know the audience. Edit aggressively.

All good editorial practice. But organisations need to interpret “editing” much more broadly than polishing whatever an AI has produced — because by the time you are editing the prose, several more important decisions have already been made.

Someone has decided what the subject is. Someone has selected the evidence. Someone has chosen the argument. Someone has decided which customer quote matters and which does not. Someone has interpreted the research. Someone has decided that this thing deserves to exist at all.

If those decisions are weak, beautiful sentences won’t rescue the piece.

AI can make a bad idea sound remarkably professional.

That is not a neutral capability. A weak argument that reads badly gets caught. A weak argument that reads well gets published, circulated and cited. Improving the prose without improving the thinking doesn’t fix the problem — it removes the warning label.

The real bottleneck moves upstream

This is one of the ideas behind what I’ve been calling Editorial Intelligence. The workflow I’m increasingly interested in looks something like this:

evidence
    ↓
judgement
    ↓
argument
    ↓
AI-assisted production
    ↓
editorial judgement
    ↓
publication

Writing sits in the middle of that chain. It isn’t the whole system. Increasingly, it may not even be the most difficult part.

Imagine a company holding 20 customer interviews, hundreds of sales calls, research reports, product data, executive expertise and years of internal knowledge.

The AI writing problem is easy. Ask a model to generate an article and it will.

The difficult questions are different. What patterns are appearing across those conversations? Is there a contradiction between what customers say and what the company believes? Is there an emerging theme that nobody has articulated yet? Does the evidence support the argument the company wants to make? Is there anything genuinely interesting here?

Those require interpretation. They require context. And they require the ability to say: no, that isn’t the story. This is.

That is editorial judgement, and it is the work the Evidence Engine exists to make repeatable rather than accidental.

Better AI makes judgement more valuable

There is a strange tendency to assume that as AI gets better, human involvement inevitably becomes less important.

In some parts of the production process, that is probably true. I already use AI to do work that would once have taken me hours.

But making production easier doesn’t necessarily reduce the value of the human contribution. It can move it somewhere else.

Photography didn’t eliminate the need for deciding what was worth photographing. Digital publishing didn’t eliminate the need for editors. Cheap website builders didn’t eliminate the need for good product decisions.

AI writing won’t eliminate the need for editorial thinking. It may increase it.

Because when producing something becomes almost frictionless, restraint becomes a competitive advantage. Knowing what not to produce matters. Knowing when the evidence is weak matters. Knowing when an argument is derivative matters. Knowing when there isn’t actually a story matters.

That last one applies to this piece as much as to anything else. The test I applied before publishing was not whether I could make the argument again. It was whether there was new evidence that changed what I could legitimately claim. There was: a 1.2m-word corpus showing the stylistic tells fading, and a book market that went majority-machine in three years.

The AI writing problem is really a judgement problem

So I think The Economist is right. Humans do need to get better at editing.

But I’d push the argument further. We need to become better editors of more than sentences.

Edit evidence. Edit ideas. Edit organisational knowledge. Edit arguments. Edit what deserves attention.

The important skill in an AI-saturated information environment isn’t turning mediocre machine prose into good prose. It is deciding what is true, distinctive, useful and worth saying before the machine starts writing.

AI is rapidly making writing abundant.

Judgement is what becomes scarce.