GOV.UK Chat rolls out in beta as GDS plans agentic AI deployments

Much of the negative attention directed toward large language models (LLM) and other generative AI tech focuses on so-called AI slop: imagery generated by an algorithmic model that has trained on huge amounts of data, but still often looks fake, garish or creepy. Not to be outdone, the UK government is expanding on what seems like the top contender for “AI tech people hate the most” – the institutional chatbot.

“GOV.UK Chat is an AI-powered chat system we’ve built at the Government Digital Service (GDS),” says a blog post from the agency. “It’s a new way to interact with government, designed to make it easier for people to find what they need on GOV.UK.” The tool, based on Anthropic’s Claude LLM, will draw on information from about 100,000 GOV.UK pages, and try to mimic the conversational tone of a real person.

Anyone who has experienced mounting frustration as they try and get an automated call answering service to connect them to a real human will surely wonder at the limits of this approach. It is somewhat remarkable that GDS uses, as an example of a question you might want to ask GOV.UK Chat, “I’ve just had a baby, do you know what help I can get?” GDS says the “answer will bring together content from across government: in this example joining up the support offer from HMRC, DWP and the Department for Education, into a single, simple conversational answer.”

The example is perhaps the worst one possible, echoing OpenAI chief Sam Altman’s much-derided statement on late night TV that he can’t imagine raising a child without ChapGPT. It demonstrates both a comical confidence in AI’s ability to help in an extremely complex situation, and an institutional memory that apparently does not extend back past 2022, when everyone on Earth had babies without ChatGPT.

Chatbots to assume active role in government operations

Nonetheless, the UK government chabot is being pitched as a “core part of the work GDS is doing to create a new generation of personalized, joined-up public services,” highlighted as “one of 5 ‘kickstarters’ in the blueprint for modern digital government.” Rollout, which follows two scaled pilot deployments in 2025, will begin with a release in the GOV.UK app early next year, and eventually be widely available across the GOV.UK website.

“As with all our products, we’re taking a user-centred, agile approach,” GDS says, “rolling out in stages and making improvements as we go. We’ll continue to invest heavily in measures designed to minimise risks and protect users’ privacy and security.”

The government is confident enough in the chatbot’s potential to have laid out an extensive expansion plan that involves agentic AI, and “the idea that GOV.UK Chat can evolve from providing answers to performing actions.”

“We can envisage use cases where it makes sense to pass GOV.UK Chat users on to a department’s AI agent or service when needed, and we’re carrying out experimental work to build the technology to do this.”

Ukraine provides model for successful government digitization

This latter idea is “inspired by the pioneering work of Ukraine’s Diia.AI,” the app that began as a document sharing platform but has evolved into a “superapp” with more than 21 million unique users, offering access to 24 different documents like identity cards and passports, and more than 30 services.

Ukraine has established Dii.AI as a core piece of digital public infrastructure, and also managed to find a way to monetize it by charging financial institutions to use the document sharing feature. According to GovInsider, banks can pay a one-time fee of 21,617 Ukrainian Hryvnia (about $500) to connect to the service, and a small fee for each successful share. “With a revenue generation model in place, there are plans to legally separate the platform into an independent state-owned enterprise, with the aim of preparing it for an Initial Public Offering (IPO) in the future,” says the report.

The app also has a beta version of an AI assistant modelled on ChatGPT, which responds to natural language prompts in an attempt to perform various tasks for users. In a post on the blog of the government’s digital agency, Vice Prime Minister and Minister of Digital Transformation of Ukraine Mykhailo Fedorov calls it “the world’s first national AI agent that goes beyond answering questions – it actually provides government services directly within a chat.”

“Our focus is to transform Diia from a digital services platform into a fully functional AI agent that operates 24/7, without the need to manually fill out forms or fields,” Fedorov says. “The AI agent doesn’t simply act as a chatbot that responds to queries – it takes action based on the user’s request. For example, you write to the assistant in the chat, ‘I need an income certificate’, and receive it directly in your personal account on the Diia portal, with an email notification once it’s ready.”

The case of Diia is noteworthy for a few key reasons. One is the emphasis Ukraine has put on user experience, bringing aesthetics into the discussion; the app succeeded in part because, in Fedorov’s phrasing, they made it sexy. Another is that it provides a clear use case for agentic AI, in a much more straightforward way than tends to come from industry voices. It’s one thing to say that AI agents will transform work, another to say, “it can request and download this specific form for you.”

The other factor is Ukraine’s overall digital strategy, which explicitly aims to see the country, by 2030, become one of the top three in the world in terms of AI development and integration in the public sector.

Digital leader Estonia eschews LLMs for gov’t chatbot

A solid bet on number one is Estonia, which has partnered with Ukraine on shared digitization goals. The northeastern European nation has long been associated with advanced digital transformation. Unsurprisingly, it has also deployed a chatbot – but with a critical difference: per the BBC, “Estonia’s chatbots are not based on Large Language Models like ChatGPT or Google’s Gemini. Instead they use Natural Language Processing (NLP), a technology which preceded the latest wave of AI. Estonia’s NLP algorithms break down a request into small segments, identify key words, and from that infer what a user wants.”

The limitations on NLP mean that, unlike LLMs, they’re not great at imitating or processing subtleties in human speech. But “they are unlikely to give wrong or misleading answers.”

Estonia’s system throws into relief a big potential problem with widespread implementation: the term “AI” itself. LLMs and NLPs are not the same, though both are classified under the same blanket term. The tradeoff would seem to be that LLMs are fun: humanesque confidants that are free with their compliments and encouragement; even if they sometimes tell you to kill yourself, in general they’re likely to provide a more delightful user experience. NLPs, on the other hand, are boring but more accurate. In this, the AI era finds itself facing a familiar fork in the path, with one way leading to practicality and common sense, and the other to a legendary treasure trove guarded by a psychotic dragon.

UK bot trending well in accuracy

Whether or not GOV.UK Chat has the same juice as Ukraine or Estonia is questionable. But the Government Digital Service says the chatbot is “beating industry standards” for the accuracy of its response to users.

According to Public Technology, government records published in October via the UK’s algorithmic transparency recording standard hub suggest that AI hallucinations associated with incorrect answers have been mostly weeded out as GOV.UK Chat has evolved from its first iteration (based on ChatGPT) to its current one. Tests continue across a range of topic areas using automated and manual evaluation processes, assessing accuracy of responses.

“GOV.UK Chat team have iterated to improve for accuracy, and have seen continual improvement with the current version of Chat beating industry standards,” say the documents. “The GOV.UK Chat team have also undertaken actions to ensure all users of Chat are aware that the answers may be inaccurate, and that they should check their answers. Links to pages used to generate an answer are always provided to users, alongside a reminder to check their answers.”

“We also perform red teaming, systematically probing the chatbot with adversarial and edge-case inputs to uncover vulnerabilities and safety risks. Together, these methods provide both a realistic view of end-user experience and a risk-aware perspective on system performance.”

Does AI belong in government? We’re about to find out

That may not be much of a comfort for anyone worried AI is moving too quickly, and doesn’t work well enough to have a place in government. The UK has insisted that it does not, as the famous phrase goes, move fast and break things. But the pace of technological change cannot always be dictated by government.

It’s hard to say whether the global push to adopt AI into public services has been prompted by Silicon Valley’s hype machine, or whether governments are genuinely coming to believe on their own in the benefits promised by LLMs and other generative AI technology. For now, the last word goes to Sven Nyholm, professor of the ethics of artificial intelligence at Munich’s Ludwig Maximilians University, in comments to the BBC.

“A chatbot is not interchangeable with a civil servant,” Nyholm says. “A human being can be accountable and morally responsible for their actions. AI chatbots cannot be accountable for what they do. Public administration requires accountability, and so therefore it requires human beings.”

Article Topics

chatbots  |  digital government  |  digital ID  |  Diia  |  Estonia  |  Gov.UK  |  government services  |  Ukraine