The work of digital government is a political endeavour. At it’s best, it is a slow grind that seeks to use the power and network of the internet to help balance power between electorate and government.
For every fraudulent contract, there is a SpendNetwork (and Contracts Finder) to try and make that visible. For every bigoted speech there is Hansard and TheyWorkForYou so they have to sign their name to it. For gerrymandering and administrative burden on elections, we have local government’s democratic services officers working with Democracy Club to make it easy to vote without knowing how voting works.
All of these are deeply insufficient for the times we live in, but they are the start of infrastructure that with help, co-operation and willpower, gives people recourse at the very least to information to structure your politics on. Entities like FullFact would not exist without a strong data layer of civic technology, and while FullFact are also deeply insufficient, they demonstrate what could be possible. They all demonstrate the ecosystem that can work, but that it relies on paying your work forward. The next person in the chain can take your work and use it, while still relying on the work in the dependencies.
But, it is political. The movement to make this data open in government has sometimes been presented as a bureaucratic change, rather than something with deeper meaning and implications. To do it, you have to believe in the idea that it provides a benefit to the country to have an open government and an informed citizenry.
I read this blogpost from Paul Maltby, ex of DLUHC and now working for Faculty (an AI shop).
https://twitter.com/maltbyps/status/1668160513245626369
I think it’s an interesting read, but leaning on what I said above, I’d like to put forward why I think this prioritisation isn’t right.
I think when you’re prioritising work for something like public sector information, you have to actually prioritise it based on needs. There isn’t infinite money to move pixels around, and some things have more impact than others.
Obviously things have changed a great deal since the UK’s digital government journey started. Paul quite rightly picks up on Martha Lane Fox’s report and how little of it beyond “fix publishing has ever come to pass beyond limited, essential examples. But: I can’t help but think about the past few years and their political fallout. At times, guidance on what was legal and illegal changed overnight, with full guidance being written on the hoof. A defence barrister friend of mine keeps talking about how procedurally weird some of the Covid prosecutions are, how they are based around things that were illegal for a short amount of time, and reconstituted in various ways throughout the different lockdowns and how neither the defendant, police, prosecutors etc. were used to the offences in question.
Now, imagine if that information had been disseminated though a generative AI chatbot. Chatbot answers that involve generative AI are great for parlour tricks, but if it’s about information that could see you taken to court, that feels like unnecessary danger for unspecified benefit. I think it is subtly different to formatting your information to be consumed by a Smart Answer in Google or to be read out by a voice assistant as the information might be ingested into the corpus and regurgitated long after it has ceased to be true. It might be combined with other data or information to produce a different kind of false answer. And if it doesn’t do those things, how is it different to an older style chatbot other than having a more believable tone of voice?
So, in this line, I would argue that there is no direct user need for this as it risks false information being put out with the stamp of authenticity, with little recourse. But, user needs aren’t the only reason we do things. There is also The Economy. Except, here too we see a problem. There are several large players in the UK AI space of which Faculty is one (others are listed here: https://arstechnica.com/information-technology/2023/06/uks-plans-for-first-global-summit-on-ai-safety-draw-criticism/). Their service offering seems to be that they will ingest data (usually, but not always, openly licensed data) and then provide innovative applications for that data.
This sounds a lot like enclosure. Taking a public good and converting it into private benefit. These aren’t open source models, these aren’t even open sourced code applications. This is a way of taking data infrastructure and making its productive use gate-kept by a private firm. For the government to prioritise this over making data clearly and generally available and assuming a human-first prioritisation on written content risks closing a market by acceding to almost universally closed and hand-waving data practices across the AI sector.
We are in the middle of a climate crisis. The carbon costs of large language models is deeply concerning. While I don’t doubt we’ll see compelling use cases for AI in the future, there is no way for us to have a say in not barrelling on with planet damaging practices. I’m not sure we can stop AI creating an enormous tonnage contribution of CO2 at this rate, but it needs to be said out loud each time we talk about these companies that they are creating an extremely dirty industry and it doesn’t even provide a tangible benefit like driving you somewhere.
People do not have individual time to participate in micro-decisions about the way they are governed, so politicians and public servants have to act in the best interests of people as a whole. After all, they are mostly accountable to them, are removable by election (plus civil servants being removed by politicians). Third part suppliers are not, their roadmaps are not based on value to the citizen, but value to their owners.
The political project of digital government is to make government transparent and legible to the citizen because government should have nothing (much) to hide. It contains a range of small, petty tools for citizens to prod the machine. I worry that these suggested directions to make data more easily ingestible for LLMs take an ecosystem of civic and public technology and seek to place their outputs in a walled garden.