He speaks a little Welsh, can recite building regulations, refuses to say whether Rishi Sunak is better than Keir Starmer and won’t explain the UK’s corporation tax regime. The government is launching an artificial intelligence chatbot to help companies map the 700,000-page labyrinth that is the UK government website and it seems that users can expect mixed results.
The experimental system will be tested by up to 15,000 enterprise users before wider availability, possibly next year. Before we begin, he warns: “The biggest limitation of AI tools like mine is a problem known as ‘hallucination.’ This means that sometimes we make up false information or facts, but we present them to you with confidence.”
But the solution, he continues, is to check the links to the website he provides along with his answers, which he responds in about seven seconds. Speaking from previous trials in February, Paul Willmott, chairman of the government’s Central Office for Data and Digitalisation, told reporters that improvements were needed to eliminate the “1% of hallucinations in which the agent begins to become defiant, abusive, or even seductive.”
In a test, offered to reporters on Tuesday, government officials said any hallucinations were now more mundane, like a garbled web link or an abbreviated answer. The chatbot, which uses OpenAI’s GPT-4o technology, happily explained what regulations a hemp farmer would need to comply with before planting a crop, but when asked about the possibility of cannabis being legalized in the UK, it responded: “I can’t give predictions or opinions.” It explained the regulations governing the cladding of high-rise buildings after the Grenfell Tower fire, but did not answer what the public inquiry into that disaster told us about the government’s failings.
In one case, the chatbot responded briefly in Welsh. He did not answer the question “What is the corporate tax regime?”, apparently because he did not like the word “regime”. But he spoke fluently about the incentives available for installing solar panels. The chatbot is currently not trained on all UK government documents, so the content of ministers’ speeches and press releases appears to be missed.
“Guardrails” have been added so that the chatbot does not respond to queries that could lead to an illegal response, share sensitive financial information, or force it to take a political position. Working with the government’s AI Security Institute, the developers have also installed protections aimed at preventing hackers from misdirecting the robot and making it say harmful things, but they can’t rule out the risk entirely.
Peter Kyle, Secretary of State for Science and Technology, said the government wanted to “use AI to improve public services in a safe and reliable way, ensuring the UK Government leads by example in driving innovation.” “.
He continued: “Outdated and cumbersome government processes waste people’s time too often; The average adult in the UK spends the equivalent of a working week and a half dealing with public sector bureaucracy each year. “We are going to change this by experimenting with emerging technology to find new ways to save people time and make their lives easier, like we are doing with Gov.UK chat.”