GOV.UK Chat launches in app with 90% accuracy claim and day-one tax errors

TL;DR:

• GOV.UK Chat launched in the GOV.UK app on Friday 15 May, allowing users to ask questions in plain language and draw answers from the 80,000-page GOV.UK corpus, behind GOV.UK One Login.
• Technology Secretary Liz Kendall framed it as a way to cut helpline load on departments that field roughly 100,000 calls per day, with research suggesting up to half of caller questions could be handled by the tool.
• Within hours of the official launch, tax expert Dan Neidle published evidence to The Times that the chatbot misled on the £100,000 tax-free childcare cliff edge and incorrectly suggested a personal MacBook sale could attract capital gains tax.

GOV.UK Chat is the most ambitious public-facing AI deployment from a UK department to date. The Government Digital Service (GDS) has built it on the GOV.UK content base — tens of thousands of pages of guidance — and exposed it inside the GOV.UK App to anyone with a One Login account. Ian Murray, minister for digital government and data, told The Times the system has been “tested to essentially destruction” and is “much more accurate than the human context”.

The 90% accuracy threshold

In testing, the chatbot was said to be “90 per cent accurate” — meaning one in ten responses did not match official guidance in every detail. That is the headline figure ministers are leaning on. The flip side is the failure modes that 10% mask. Neidle, founder of Tax Policy Associates, asked the system whether a £1,000 pay increase from £99,000 would affect tax-free childcare eligibility; the bot replied that there is no upper income limit at that level, omitting the well-known £100,000 cliff edge where eligibility collapses. He then asked whether selling two old MacBooks on eBay for £1,300 needed declaring; the bot framed the answer around capital gains tax, which Neidle bluntly described as “stupid” — personal-use chattels like a MacBook are not subject to CGT.

What the rollout signals for UK public-sector AI

The launch lands alongside Warwickshire’s £2.4m council-wide AI procurement (covered on Resultsense earlier this week) and the Anthropic-PwC enterprise expansion — three datapoints from the same week pointing to AI moving from pilot to production across UK service delivery. The consolidation matters: GOV.UK Chat is the standard against which every council and department chatbot will now be benchmarked.

Looking forward

The next test is how the GDS team triages the 10% wrong-answer category. If errors cluster in low-stakes general queries, the consumer impact is limited. If they cluster — as Neidle’s spot-testing suggests — around tax, benefits, and means-tested entitlements, the cost shifts from chatbot embarrassment to actual citizen detriment. UK SMEs and individual users should treat day-one chatbot answers on tax, childcare, and HICBC as starting points, not advice — and HMRC’s own helplines remain available for the borderline cases.