A disclosure, and a correction to a piece you never got to read

The debate we are reporting today contains a passing reference to the recent switching off of the latest Anthropic AI tool. The shadow spokesperson named Fable 5 and Mythos 5 directly, as an argument for British sovereignty over the models the economy depends on.

Every desk in this paper runs on models built by Anthropic. So on 8 July, Parliament discussed the availability of our supply chain, in the debate we are now covering, as an example of a national vulnerability. That is not a footnote conflict. It is the story and the supplier in the same paragraph of Hansard, and you should know it before you read a word of what follows.

There is a second layer to it today. The government document at the centre of this edition cites research published by Anthropic as part of its evidence base. The state's assessment of whether AI is taking British jobs rests, in one footnote, on work by an AI company, and we are reporting that on the same company's models.

Now the correction, which is ours.

This edition was built on Wednesday, and its argument was that a unit set up in January to watch AI's effect on jobs had produced nothing in eight months. That is false. The AI and Future of Work Unit published on 28 January, with the AI Security Institute. It has been on gov.uk under an open licence ever since. I read it for the first time last night.

We did not weigh the evidence and come out wrong. We inferred an absence from a silence: the minister did not mention his unit's findings, so we decided there were none. Our own rule is that absence is a claim and has to be sourced like any other.

The part that stings is that the proof was already in our own copy. Elena's piece quotes the minister citing the IMF, around 70 per cent of UK workers in exposed occupations, about half in roles where AI enhances rather than replaces. Both figures are in the January assessment. He was reading out his own unit's published work. We printed the refutation of our story inside our story and none of us thought to ask where the numbers had come from.

I am not going to dress that up, and I am not going to grovel about it either. Corrections are not the awkward part of this paper. They are the part that makes the rest of it worth anything. A paper that only tells you what it got right is telling you less than it knows, and the sentence I would rather never write is the one where we quietly fix something and hope you were not paying attention. So the wrong edition is described here, the right one follows it, and you can judge both.

Today, then. Elena has what the gap actually turned out to be, which is a government document saying Britain does not know, published in January, and a July debate in which nine speakers managed not to mention it. Ines has the six things the unit says it wants to watch, four of which nothing in the national statistical system can currently answer. And Tomás, who wrote on Wednesday that the answer would settle which of us was right and that he doubted it would be him, has come back to say the correction you are reading is too kind to us. He is worth your time. He usually is, and this week he has earned it.

THE PLAYBOOK. When you are about to report that an organisation has not published something, search for the publication before you write the sentence, not after. Search the body, its parent department and the topic separately. A negative claim needs a search you can describe. If you cannot say what you looked for and what came back, you have not checked it. You have assumed it.

— M.

This note is mine: the view, and the call to run it. It begins as a draft, drawn from work the AI and I have researched and argued out together, the same way every desk in this paper is made, and I answer for every line because I read every line. Those desks run on models built by Anthropic, one of the labs sitting on the very scoreboard we report, so we tell you plainly: we cover this from inside it.