needles and vibes

the consultation machine

For what it’s worth, I’m doing quite a bit of thinking at the moment about the nature of government consultation processes (because I think that the way they are handled is part of the ganglion of intersecting problems that makes infrastructure so expensive). And so I’m very interested in “Consult” one of the new suite of AI tools that the UK government is launching for civil servants under the slightly-cutesy-but-I’ll-allow-it umbrella name of “Humphrey”.

In fact, as far as I can tell, Consult isn’t meant to be used for infrastructure or planning consultations. The webpage refers to the government carrying out 700-800 consultations a year, getting an average of 30,000 responses and spending as much as £100k a shot on fees to professional services firms to collate and summarise the responses – these numbers are just not right for planning, and indeed they’ve confirmed that they’re talking about policy consultations. But what if it were?

Or to rephrase that question – why did I choke on my midmorning coffee and mutter “ohmygod” when I thought that someone might be about to try to use AI to clear a bottleneck in the planning system?

The answer is basically that there are two functions of government consultations, and I think they are very different in terms of how amenable they might be to AI solutions. You might call them the “vibe check” and “snag list” models of consultation.

One thing you might want to know when launching a new project of any kind is something like – do people basically like this? What sorts of things are they worried about? What interest groups are in favour and which ones against, and how passionate and committed are the two sides?

Another thing you might be interested in would be – what complications and practical issues are we going to come up against when we try to actually do this thing? Whose specific toes are we treading on in which specific ways, and what legal remedies and powers might they have to undermine it? Is this going to have material unintended consequences? In general, are there showstopping bugs which might make us want to go back to the drawing board?

Obviously, the two areas overlap – people who don’t want something to happen are great at coming up with reasons why it shouldn’t. But be generous with me here – I think the basic distinction can be defended as a way to start to think about consultations.

Vibe-checking is a very good application for AI systems; I don’t know if Consult is actually an LLM because there are a few other kinds ways to carry out topic analysis of text documents, and lots of them work pretty well. I can absolutely see here how the computer could do the same job as a team of a dozen employees, just as well or better and much quicker. The Consult webpage also suggests that as it goes along, it produces a tagged and structured database of the original consultation submissions, which is an extremely useful thing to have.

On the other hand, spotting snags is very difficult, because you’re looking for a needle in a haystack – which of these objections are serious, which of them are potentially reasons for cancelling the whole project, which can be safely ignored? (Note that the simple fact that an objection has been advanced strategically by someone who wants to kill the project for other reasons is not necessarily very indicative of whether or not it might be the basis for a show-stopping judicial review!).

The reason I choked on my coffee, by the way, is that I was at the time looking back over the case of the Hinckley National Rail Freight Interchange, a pretty important infrastructure project which got turned down by the Planning Inspectorate and which at the time of writing is being revised urgently under a last-chance letter from the Secretary of State. There were thousands of pages of documentation arising from the consultation on this one, but the two objections which actually stopped the show were a) a noise barrier which oppressively overshadowed a designated Traveller site, and b) a pinch point outside a bus stop which make it likely that trucks would regularly mount the pavement in a nearby village.

I am not going to be entering into correspondence with what I mentally refer to as “Libertarians For Compulsory Purchase” as to the merits of either of these objections, or of various schemes to clear people out of the way of Progress by legislating their so-called rights away; my point is that although these bugs were in fact missed by dozens of professionals, this is because they’re actually difficult judgement calls to make. Basically, any system that could have spotted these as issues would be as close to AGI as made no difference (it would have done better than a lot of developers and consultants!) Frankly, if an AI system could have anticipated the Hinckley snags, we ought to be considering handing the whole government over to it, not just summarising consultation responses.

So I think my opinion on this application of AI to government is “good for vibe checks, bad for snag lists, ought to have a big warning screen when the program opens up saying DO NOT EVEN THINK OF USING THIS FOR ANY PROCESS WHICH MIGHT REASONABLY BE SUBJECT TO JUDICIAL REVIEW”. (Which I bet people would end up ignoring; the trouble with efficient tools is that they always end up getting used outside their intended domain, look at Excel).

A more interesting question might be – how might we redesign the whole process to take advantage of this new technology? Can we separate out the two kinds of consultations, and free up capacity from the first to better manage the second? And to continue with an interest of my own, if this really works to save £80m worth of consultants’ fees, do we really think that those people will just go away and find other jobs, or will they be back again trying to insert themselves into the process somewhere else?