The unit nobody has placed left one assessment behind. It is the most careful thing the British state has written about AI and jobs.

The document at the centre of this edition was published on 28 January. It is worth reading in full, and here is what it says, because it is more restrained than either side of this argument usually admits.

Start with exposure, where around 70 per cent of UK workers are in occupations containing tasks that AI could potentially perform or enhance, on IMF estimates. That is higher than the United States at around 60 per cent, and higher than the advanced economy average. The reason is not that Britain is behind. It is that Britain sells services.

Break the seventy down, because the halves point in opposite directions. Thirty-five per cent of UK workers are high exposure and high complementarity, meaning the job has a social or physical context that makes unsupervised machine work unlikely, so AI is more likely to make them faster. Thirty-two per cent are high exposure and low complementarity, where the machine may do tasks a person currently does. Thirty-three per cent are low exposure. The American split is 30, 30 and 41.

So roughly a third of British workers sit in the column the report describes as facing transition challenges that will need to be managed. That is the number I would put on a wall.

The report is equally clear about the prize. It cites OECD work putting potential UK labour productivity growth from AI at 0.4 to 1.2 percentage points a year over the next decade, second only to the United States among the G7, against a productivity level around 20 per cent below America's.

Then the hiring evidence, where the care is the point.

Analysis of UK job postings found that a one standard deviation increase in AI exposure was associated with a 3.9 per cent fall in posting volume, becoming statistically significant around seven months after ChatGPT was released. Postings returned to their original levels after roughly twenty months. The fall concentrated in high-salary occupations, with no significant change in low-salary roles. Separately, McKinsey found UK job adverts fell 38 per cent for high-exposure occupations between 2022 and 2025, against 21 per cent for low-exposure ones. And in 2024, in the first fall in UK digital sector employment in a decade, the number of 16 to 24 year olds in computer programming dropped 44 per cent in a single year.

Now the discipline, and this is why the document deserves respect.

It states plainly that these patterns do not establish AI as the cause, and gives two reasons. Exposure is not adoption: the indices measure whether an occupation's tasks are suitable for a machine, not whether any firm has deployed one. And exposure correlates with interest rate sensitivity, so a tightening cycle beginning in early 2022, before generative AI was publicly available, would produce a pattern that looks much the same.

It then does something official documents rarely do and puts the contrary evidence in. The Yale Budget Lab, examining 33 months of US data, found the occupational mix is not changing faster than in previous technological transitions. Danish micro-data found AI adoption had no measurable effect on worker earnings or hours, including among daily users at early-adopting workplaces. And hospitality, a sector with low AI exposure, accounted for 53 per cent of UK job losses between October 2024 and August 2025.

Adoption itself is modest. Around one firm in five uses or plans to use AI. Within firms that have adopted, fewer than a third of employees use it. Micro businesses are at 14 per cent.

That is the state of the evidence, and it is honest about being thin.

Which brings me to the last page, and the sentence this edition turns on. The report sets out six indicator areas for real-time monitoring: capability advancement, proliferation across business and society, the reshaping of jobs, distributional impacts, the macroeconomic picture including the tax base, and new job creation. Then comes the condition attached to all of it: subject to agreement across government on the approach to monitoring and publication, the Unit intends to publish updates on those indicators.

One further detail, and it is the sharpest thing in the file. The assessment is published jointly by DSIT and the AI Security Institute. The AI Security Institute's new address is known and published: it has moved to the Office for the Prime Minister and the Cabinet. Its co-author's has not been stated anywhere we can find. One half of the authorship of Britain's assessment of AI and the labour market has been placed. The other half has not.

Our own disclosure belongs in this piece rather than in a footer. The report's second footnote cites Anthropic research on the impact of AI on software development. The desks of this paper run on models built by Anthropic. We are reporting on a government assessment that draws in part on our own supplier's research, and we would rather tell you than leave you to find it.