◆ How It's Made · A worked example
One editor, one AI, one evening
This is the record of a single edition being built. We are publishing it because it is the most honest thing we can show you about how this paper is made — including the part that does not flatter us.
Wednesday 8 July 2026
Issue 19 · Evening Edition
Six errors were made that evening. Three were the machine's, two were the editor's, and one the machine caught in itself.
A newspaper that runs on AI has an obvious problem: you have no way of knowing whether anybody is really checking. Principles are cheap to publish. So rather than tell you again that a human edits every piece, here is one evening, written down as it happened — the brief, the mistakes, who caught each one, and what came out the other end.
We have left the mistakes in. A record showing only the good calls would be the opposite of the point.
The editor
Matt Brazil · human · holds every publish gate
Has built and run small companies, so he knows the founder's side of this story in his bones. Brings the conviction, and the ear for when the machine drifts into fine writing or a comfortable number.
The AI
Claude · drafts the desks, runs the research
Brings speed and reach: four beats checked against live government sources in an evening. And the discipline to test a claim against the record, including the editor's own claims.
Round 1
The brief
EditorRun the evening edition. Issue 19.
ClaudePicked the spine, built four pieces, checked each against live sources. Chose the theme with the deeper research behind it over the one that had been suggested. A small act of judgement, and nobody had to ask for it.
Round 2
The writerly voice
Caught by the editor
EditorWould you say this, out loud, to someone across a table? Because you wouldn't say half of it. It reads like an essay, not a person talking.
ClaudeHe was right. Three of the four pieces were rewritten in a spoken register. Then, rather than fix it once and forget, the check was made permanent.
OutcomeA new rule went into the paper's house constitution: the Said-Aloud Gate. One correction became a standing guardrail every future piece has to pass.
Round 3
The false premise
Caught by the editor
ClaudeOpened the editor's note with a confident line: that a major political plan had written the future of work.
EditorIt did the opposite. That plan is about retraining and reskilling. It wrote the past of work. Your own morning edition says so.
ClaudeNot merely weak, but backwards, and contradicting what this paper had published that same morning. The note was rewritten from the true premise up.
Round 4
The comfortable number
Caught by the editor
ClaudeOne desk leaned on a reassuring official figure — only five per cent of firms had cut headcount because of AI — and let it carry the conclusion.
EditorThat is the lag, not the all-clear. I have seen a company map its own work against these tools and find nine roles in ten already doable by them. They chose not to cut. Their competitors will not all be so kind.
ClaudeThe desk was re-cut on public evidence to argue the harder thing: the capability is already here, and staying fully staffed is a choice that will not hold everywhere. The editor's first-hand account could not run on a sourced desk, where every claim must be checkable by you. It moved to his signed note instead, anonymised. The wall between reporting and opinion held.
Round 5
The editor's own numbers
Caught by the AI
EditorSmall business is the backbone of this country. One-person firms, four million. One to nine staff, four million. Ten to forty-nine, four million. Public-sector jobs are declining and self-employment is the growth story. Have a look at the stats.
ClaudeIt looked, at the primary data, and pushed back. The bands were wrong: it is a steep pyramid, not three equal tiers. About 4.3 million firms have no staff at all, roughly 1.1 million have one to nine, about 0.2 million have ten to forty-nine. Two other planks did not hold. Public-sector employment is rising, not falling, with central government and the NHS at record highs; the decline is a story about the 2010s. And self-employment jobs actually fell last year, even as the number of solo firms grew.
ClaudeThe editor's core thesis survived intact. Britain is a nation of very small firms, and they move first. The real figures told it harder than the round ones did.
OutcomeThe corrected, fully sourced version went into the edition, and then grew into a standalone report on who is actually employed where.
Round 6
A catch on itself
The AI, on its own work
ClaudeBuilt that report, and had quietly stamped it "Field Report No. 04."
ClaudeCaught it before handing over. Numbering a report into the sequence is a decision that belongs to the editor, not something a machine gets to assert on its own authority. It was marked a draft instead. A small thing. But it is the same discipline, pointed back at itself.
The tally, kept honest
Nobody was right the whole way through
3
errors the editor caught in the AI's work: the voice, the false premise, the comfortable number
2
errors the AI caught in the editor's: the size bands, and the public-sector claim
1
error the AI caught in its own work: the invented report number
That is not a flaw in the method. That is the method.
What this cost
It would be easy to end there, and to call this a happy story about what one person and a machine can do together. We are not going to.
This paper is built by one man. The desks, the research engine, the data underneath it. Before these tools existed, what you are reading would have taken a newsroom: a dozen writers, a subeditor's desk, a data team, a picture desk.
That is not a boast. It is the question this paper exists to ask, arriving at its own door. What do the dozen people we did not have to hire do now?
We do not know. Nobody does, yet. That is precisely why we set out to report it.
So read this record for what it honestly is. It is not proof that AI is good for work, and it is not proof that it is bad. It is one small, documented case of a human and a machine producing something together that neither would have produced alone — a piece of work that is better because two very different kinds of competence kept catching what the other could not see.
The editor working at this speed does not, single-handed, check four beats against live government data before the evening is out. The AI working alone files the writerly line, the false premise and the comfortable number, and never feels that anything is wrong, because it does not run a small business and does not hold a conviction it is willing to defend in public.
Put the two together and each one's blind spot is the other's plain sight.
This is not the AI doing the editor's job, and it is not the editor supervising the AI. It is two kinds of competence checking each other in the open, with the errors left on the record instead of quietly cleaned up. That is what an ordinary Wednesday looks like here. It is also, we suspect, a small glimpse of what a great deal of work is about to look like — and we would rather show you the whole of it, including the cost, than sell you the flattering half.
Disclosure. This record was written by the same AI it describes, and the editor put his name to it. Every analyst across these pages runs on models built by Anthropic. We report on this industry from inside it, and we tell you so. The company described in Round 4 is not named, and gave its permission to be studied.