Ben Hollebon

The rulebook assumes competence

The government's digital standard went from one mention of AI to more than 60. The expectation on thousands of practitioners has outrun the help to meet it.

Red file folders are neatly arranged on a shelf

On 1 October 2026 the Government Functional Standard for Digital, GovS 005, was reissued at version 2.2. It is the biggest change since the standard came into effect in 2023, and the headline is artificial intelligence: references to it went from a single mention to more than 60, worked through the whole document. The Government Digital Service (GDS), which owns the standard, calls the additions “proportionate references”. I think they are the right additions. I also think they lift what is expected of thousands of people well ahead of anything that would let those people meet it.

The standard now requires organisations to set out how AI will be used “safely, lawfully, ethically and effectively, where relevant to organisational objectives”, and to ensure “accuracy, reliability, bias, transparency and human oversight”. Read it as a list of things to aim for and none of it is arguable. Every word is something you would want to be true. The difficulty is that each of those words names a judgement. Ensuring accuracy, noticing bias, deciding what counts as adequate human oversight: these are not boxes to tick. Each one is worked out by a person, case by case, and only if that person knows enough to do it.

The standard binds a long chain of people: permanent secretaries, directors general, the chief executives of arm’s length bodies, senior officers and every practitioner who manages digital activity. The last group is the one that worries me. Accountability can be written so that it runs up to a permanent secretary, and that is right. But the accuracy check, the bias check and the oversight happen a long way down the chain, in the hands of people who were not in the room when the standard was written, and who have been given a higher bar with not much to help them clear it.

That word, “proportionate”, is fair from where GDS sits. More than 60 references is not much for a document that governs every piece of digital work across central government, at a moment when AI reaches into most of it. The word carries an assumption, though. It measures the change against the scale of the centre and the size of the risk, and not against what it asks of the individual who has to turn each line into a decision about a real case. A change that is proportionate to the problem can still be heavy for the person at the far end of it.

The people who have to meet it

GovS 005 isn’t mine to apply, since I don’t work in central government. The problem isn’t peculiar to government, though. Anyone who runs digital work in a large, dispersed organisation, a university or a health system, will know it. A centre writes a standard. The standard is good. It then has to be met by hundreds or thousands of people spread across parts of the organisation the centre doesn’t directly run, each of whom reads the same sentence and has to work out what it means for the particular piece of work in front of them.

A centre can write a standard quickly, because writing it is within the centre’s gift. Equipping people to meet it is slower and more expensive, and it has to happen out in the parts, where central resource is hardest to direct. So the expectation lands everywhere on the day the document is published, while the means to meet it spreads out slowly, unevenly and in some places not at all.

The usual answer to a gap like this is training, and training helps, but it lags the expectation by design. You cannot build a judgement in people faster than they can build it, and the standard does not wait for the programme to catch up. In between, people are being asked to exercise a judgement the organisation is only starting to help them develop, and some of them know it. That is an uncomfortable place to put capable, willing people, and it is worth saying so rather than smoothing it over.

There is a second-order risk as well. When a standard assumes a competence that isn’t yet there, the safe course for a cautious person is to record that they considered accuracy, bias and oversight, whether or not they were in a position to judge any of them well. The requirement then gets met on paper while the thing it was there to protect does not reliably happen. I have watched that with other kinds of policy, where a rule is satisfied in form long before anyone can satisfy it in substance, and I would expect AI to be no different, perhaps worse, because the judgement it asks for is newer and easier to fake to yourself.

The practical point is a small one. When a centre issues a standard that assumes people can already do something, it should say in the same breath how it is going to help them get there, and who is paying for that. Published on its own, a standard raises the bar for everyone at once and leaves the hardest part, the part that happens out where the work is done, to sort itself out.

I’d like to see a standard that arrived with the training budget already attached. If you know of one, tell me.

Ben

Photograph by Zulfugar Karimov.