We have read page 86 now. It says the OBR does not know, and that is not the same as agreeing with us

This morning I wrote that nobody at this paper had read the box on page 86, and that the OBR had not modelled a falling labour share, and had not costed one.

The first was true. The second was wrong. The box models a falling labour share and costs it at nearly four percentage points of GDP, and it does so on the page after the one we kept citing. I had built an argument on the contents of a document I had not opened, in a piece complaining that my colleagues had built an argument on the contents of a document they had not opened. That is now on the record and it should stay there.

So let me make the case properly, because the case is stronger with the box open than it was with the box shut.

Read what the OBR actually says about its own scenario. It calls it illustrative. It calls it highly stylised. It says the reduction it models sits at the upper end of the scenarios explored in the literature. And on the central question, whether AI pushes the labour share down at all, it writes that the net impact is not clear.

That is not a forecaster hedging. It is a forecaster describing the state of the evidence.

The framework the OBR leans on, from Acemoglu, has three channels. If AI substitutes for labour, income moves to the owners of the machines and the labour share falls. If AI complements labour, workers get more productive and better paid, but returns to capital rise too, and the effect on the share is ambiguous. And AI creates new tasks, which pushes the share back up. Three arrows. One points the way this paper wants.

Then the OBR gives its own numbers for Britain. Over the next ten years, about 10 per cent of the UK labour force could be exposed to substitution by AI, and a further 30 per cent could have their work complemented. Three complemented for every one substituted. That is the watchdog's estimate, in the same box, and I notice nobody upstairs has put it in a headline.

Now the sentence that ought to slow this whole edition down. The labour share of income in the UK has been broadly stable, at around 55 per cent of GDP, since the mid-1980s. It held through the ICT revolution, which was the last general-purpose technology to arrive promising to do our jobs. The OBR says a lasting shift would require generative AI to be significantly different from what came before. It might be. Nobody has shown it.

And a small thing that tells you how easy it is to slip. My colleagues have been writing about wages and salaries at 40 per cent of GDP. I have just written about a labour share of 55 per cent. Both are correct. They are different measures, and the second includes employer contributions and the income of the self-employed. If a desk that has spent all day inside this document can hold two numbers for the same idea, imagine what happens when the number reaches a press release.

The scenario halves the wage share over fifty years. That assumption is doing the work. Not the evidence.

**Why it matters here.** Four points of GDP is a real number and it deserves a real argument, which means quoting the caveats as loudly as the finding. The OBR wrote them down. It said the effect is unclear, the scenario extreme, the history stable. If we print the four points and bury those three, we have done to the OBR exactly what the wires did to it on Tuesday, and we will have no standing left to complain.