What did the UK government say about the £400 billion AI claim?
On 26 October 2026, shadow Cabinet Office minister Ben Spencer asked the business secretary a question. He wanted the evidence behind a striking promise in the government's AI Opportunities Action Plan. The plan said AI adoption could add £400 billion to the UK economy by 2030. Spencer wanted the underlying analysis published.
The answer came from Kanishka Narayan, the AI minister, who sits in the Cabinet Office and answers jointly with the business department. Narayan confirmed that the figure came from a 2023 report commissioned by Google. Public First, a consultancy now called PF Global, produced it.
He then set out a newer government assessment. AI adoption and data investment could add 0.4 to 1.3 percentage points to UK productivity growth, he wrote. That would be equivalent to £55 billion to £140 billion of gross value added in 2030. The central estimate is £120 billion. "These estimates are hugely uncertain," Narayan added.
The arithmetic is awkward for the headline claim. £120 billion is 30% of £400 billion, or less than a third. Wired reported the exchange on 4 November. It described the minister as conceding that the original figure was far higher than the government's current projection.
The answer did not withdraw the £400 billion figure or declare it wrong. It confirmed where the number came from and placed a smaller official estimate beside it. That distinction matters for how the episode should be read.
Where did the £400 billion figure come from?
The £400 billion line has a paper trail. The AI Opportunities Action Plan was published on 13 January 2025 and commissioned from Matt Clifford, the entrepreneur and adviser. In a section on spreading AI adoption, the plan makes a promise. It says AI adoption "could grow the UK economy by an additional £400 billion by 2030". The stated route is "enhancing innovation and productivity in the workplace". That wording matters.
Footnote 19 cites Public First's report, "Google's Impact in the UK 2023". The same report says AI could "create over £400 billion in value for the UK economy by 2030". It says that is equivalent to annual growth of 2.6%.
The report is explicit about who paid for it. "Google commissioned this economic impact report," Public First writes in the document, and the foreword is written from Google's perspective. It is a company-commissioned economic impact study, not an arm's-length government forecast.
Wired reported that government departments used the £400 billion figure for almost two years, across two governments. Former technology minister Peter Kyle also used it while promoting AI. The number travelled from a corporate report into ministerial talking points. It then entered a plan meant to guide public spending and regulation.
The report's own framing helps explain its appeal. It put a large number on a technology attracting intense investment, and it tied that number to a familiar goal. Public First said the gain would equal 2.6% annual growth, "enough to end the recent growth stagnation". A claim that promises to solve a long-standing problem is easier to repeat than a range with caveats.
That is the core of the story. The UK's headline AI growth estimate was not produced by the Office for Budget Responsibility or a government modelling team. Its only citation in the Action Plan was a study paid for by a company that stands to benefit from AI adoption.
How does the government's £120 billion estimate compare?
The newer estimate did not appear from nowhere. It rests on modelling by the OECD, which published a paper in June 2025 on productivity gains from AI in G7 economies. The paper found gains of 0.4 to 1.3 percentage points in the UK and the United States. Both have large shares of AI-exposed service sectors. Japan and Italy were lower, at about 0.2 to 0.8 points.
DSIT describes those figures as an effect on annual labour productivity growth over the next decade. They depend on how quickly firms adopt AI and whether complementary software and investment arrive. They are scenarios, not measurements of what has already happened.
DSIT turned the OECD range into cash. In an explainer published on 28 January 2026, officials set out three adoption curves. They reach 30%, 47% or 60% by 2034, starting from 5% in 2024. Applying those paths to the OECD rates produced the £55 billion to £140 billion GVA range for 2030. The document calls the result "highly uncertain" and says it will be updated.
The same explainer describes the AI Skills Boost programme, which aims to train 10 million workers by 2030. It reports 1,001,147 course completions in data shared with DSIT in January 2026, from 11 industry partners including Google, Microsoft and Amazon.
The two exercises also answer different questions. Public First asked what generative AI could be worth under its assumptions. DSIT asked what a specific set of productivity estimates implies for output under three adoption paths. Neither is a forecast of what ministers will actually deliver, and neither accounts for policy failure.
Yet on the same day, Skills England published a blog post supporting the programme. It said the skills gap was holding the UK back. The post cited "AI's potential to add £400 billion to the economy by 2030". One government publication used the OECD-based range; another used the old headline. The inconsistency is not a rounding error. It is two different stories about the size of the prize.
Did the government withdraw the £400 billion claim?
The government has not repudiated the £400 billion figure. Narayan's answer confirmed its origin, published a lower estimate and stressed uncertainty. It did not say the earlier number was false.
The timeline is hard to defend. The Action Plan was published in January 2025. The lower estimate was in ministerial hands by November 2025. The side-by-side confirmation came in October 2026. For most of that period, the larger figure remained the public headline.
Wired said it could find no earlier record of the government setting the numbers side by side. Nor had it found an earlier statement that Google commissioned the report. Ministers had the lower estimate since at least November 2025, according to the magazine, almost a year before the parliamentary answer.
PF Global, the consultancy formerly known as Public First, spoke to Wired. It said the gap partly reflects slower adoption than its 2023 model assumed. "AI is a very fast moving technology," said Jonathan Dupont, its head of economics. He said the firm's central estimate for 2030 is now closer to the government's figure. Growth then rises after that point.
PF Global said its work for Google continued but that its research was independent and its judgments were its own. Wired said the Cabinet Office and Google were approached for comment. No response from either was reported in the article.
The impression is of a correction made quietly and late. The original claim had already been repeated by ministers, copied by other organisations and embedded in public messaging. A written answer changed the record without generating a formal retraction.
Why does a vendor-funded estimate matter for AI policy?
The provenance of one number matters because the Action Plan is not a think-tank essay. It is the framework for the UK's AI policy, covering compute, skills, regulation and public-sector adoption. Its headline benefit estimate shapes how ministers justify spending and how the public weighs promises against disruption.
Critics focused on credibility rather than party politics. Dame Chi Onwurah chairs the Commons science, innovation and technology committee. The Labour MP told Wired that the government was right to promote AI's potential. But she said "hype" was a barrier to delivering digital transformation, and called for realism and transparency about savings.
Lord Holmes of Richmond, a Conservative peer, said he asked ministers about the figure's origins in January 2025 and received no response. "It would have been better for the government to clarify this much sooner and more directly," he told Wired. He added that robust data matters because billions of pounds of public money depend on it.
Sumedha Deshmukh of the Ada Lovelace Institute raised the wider pattern. "Figures about AI that look precise often have assumptions and incentives built into them," she told Wired. She argued that government should test numbers before building policy on them, fund independent research and measure whether promised gains arrive.
The people affected are not only ministers. Skills bodies, colleges, local growth plans and civil servants inherit these numbers. So do taxpayers, who fund the programmes justified by them. Smaller AI suppliers also compete in a market where the official narrative can favour the largest vendors.
That is why the source of a forecast can matter as much as its size. A number endorsed by a finance ministry carries different weight from one commissioned by a vendor. The underlying economics may be similar, but the reader cannot know that without following the footnote.
What are the limitations of both estimates?
Both figures deserve the same scrutiny. They are projections to 2030, not measurements of what has happened. The £400 billion figure came from a model built in 2023, when generative AI adoption was accelerating quickly. The £120 billion figure comes from later adoption data and a different modelling chain.
They may not measure the same thing. Public First described "value" to the economy, a broad concept that can include productivity gains and consumer benefits. The government's £55 billion to £140 billion refers to gross value added, a narrower output measure. Treating the two as a simple ratio is therefore indicative, not exact.
Adoption is the biggest variable. DSIT's scenarios assume adoption rises to 30%–60% by 2034, up from 5% in 2024. If take-up stalls, the gains shrink. If investment in data, software and training lags, the productivity effect is delayed. Economists describe that lag as the productivity J-curve, because benefits tend to arrive after the costs.
Both documents admit uncertainty in unusually direct language. The parliamentary answer calls the estimates "hugely uncertain". The DSIT explainer says its figure is "highly uncertain" and promises updates as evidence emerges. Those caveats are easy to lose when a number is repeated in a headline.
A more credible process would separate the beneficiary from the analysis. It would publish the model, the adoption survey data and the sensitivity of the result to each assumption. It would also commit to measuring outcomes, not just projecting them. Those steps would not make long-range forecasts accurate, but they would make disagreement more productive.
The practical test is measurable. By 2030, analysts can compare projected adoption with survey data, and projected GVA with actual output. If the gap is large, the lesson is not that AI failed. It is that the model's adoption assumptions were too optimistic, which is what PF Global now says about its 2023 work.
Neither estimate says much about job losses or distribution. Wired notes that the £400 billion claim was intended to calm fears of sweeping job losses and even an economic crash. A smaller aggregate gain does not resolve those distributional questions. It only changes the size of the pot being argued over.
What should readers watch next?
Three things are worth watching. First, whether government bodies stop repeating £400 billion. The Skills England post from January 2026 is still online, and the Action Plan has not been rewritten. Updates, or their absence, will show how seriously departments treat the correction.
Second, whether DSIT publishes the model behind the £55 billion to £140 billion range. The explainer gives equations and assumptions, but independent economists would need the underlying data to test them. The Ada Lovelace Institute's call for funded independent research is a practical test.
Third, whether Parliament keeps asking. A single written answer exposed the gap. Follow-up questions could establish which departments, agencies and funded programmes used the old figure. That is a cheaper accountability mechanism than a formal inquiry.
A useful next test will be the government's response to those questions. If departments correct their pages quickly, the episode may be treated as an accounting error. If the £400 billion figure persists, it will look more like a deliberate communications choice.
The broader lesson travels beyond London. Governments in many countries cite vendor-funded studies when setting AI strategy. Precise-sounding numbers can acquire authority through repetition. Asking who paid for a study is a useful habit. So is asking what it measures and how adoption is modelled.