The Budget credits AI with faster lung cancer results for four million patients. Its source counts X-rays.

The Budget published on 28 October makes a specific claim for the government’s use of AI in the NHS. In its section on reforming the state, it says: “The government’s AI-supported diagnostic tools have already helped over four million patients receive faster lung cancer diagnoses or results, with the AI Diagnostic Fund supporting technologies focused on high-demand services for lung cancer, such as chest X-ray and CT diagnostics.” Footnote 66 gives the source: an NHS England spreadsheet titled “Chest X-ray AI – Studies Successfully Processed by NHS Trust, by month.”

The title says what it counts. The spreadsheet adds up 4,303,064 chest X-ray studies that AI software processed, compiled from the suppliers’ processing records. It has no column for reporting time, diagnosis or result, and it counts studies, so a patient X-rayed twice appears twice. The Budget presents a count of scans the software handled as a count of patients it helped, and says they got results faster. We rate that sentence Misleading (1/5).

Whether anyone got a result faster is a question about a comparison, and two controlled studies have made one. The independent evaluation of the AI Diagnostic Fund, commissioned by the National Institute for Health and Care Research and published in March, compared trusts using AI with trusts that were not. The share of X-rays reported within 24 hours “did not change significantly after accounting for changes in non-AI sites,” and “AI deployment did not have a significant association with the time between x-ray and CT scan.” The LungIMPACT randomised trial, published in Nature Medicine the same month, found that moving the X-rays the AI flagged to the top of the reporting list cut the median time to a report from 47.0 to 34.1 hours, and found no significant difference in time to CT, diagnosis or treatment.

“Or results” is broader than diagnosis, and the trial’s faster reports show that some patients can get a result sooner when the AI reorders the list. That is a long way from four million. None of this shows the software is useless. In the six trusts where it was measured, the evaluation found the tools classed 90% of X-rays that readers judged highly suspicious for cancer as high priority. The trial tested one use of the AI, prioritisation, with the AI switched on in both arms; it did not compare AI with no AI.

The programme is large. The software reads each chest X-ray, often the first test when lung cancer is a possibility, and marks findings such as nodules; at most of the sites studied it also moves the X-rays it judges abnormal up the radiologists’ reporting list. In January to March 2026 it supported 781,000 of the 2,052,110 chest X-rays performed in the NHS in England, 38.1%, according to a ministerial answer on 19 October. Much of the rollout was paid for by a £21 million AI Diagnostic Fund, and in June the government announced £20 million more to extend “this proven technology to every NHS trust in England by 2029.” NICE’s guidance on the software, for chest X-rays of adults referred from primary care with suspected lung cancer, still says it “should not be used for clinical decision making in the NHS until more evidence is available.” Centres already using it may continue only under an evaluation framework and alongside clinician review.

Where the claim appeared

The four-million figure first appeared in a Department of Health and Social Care press release in June. The Treasury added further figures in October, four days before the Budget. Asked in Parliament what evaluation of clinical outcomes had been done, the department pointed to its own June press release.

What the footnote counts

NHS England’s spreadsheet lists, for each trust and month from the trust’s go-live to 31 May 2026, the number of chest X-ray studies its AI software processed. The definition is printed above the table: “CXR studies successfully processed by Chest X-ray AI (counted once a successful AI result is returned).” The notes name where the numbers come from: Harrison.ai’s backend processing records for five imaging networks, reports from the Sectra Amplifier platform for two, Qure.ai’s backend processing records for north-west London, and, for Leeds, which uses the Newton’s Tree platform, figures “as supplied by the customer-success team.” The grand total is 4,303,064.

The total includes urgent and routine studies, abnormal and normal, first X-rays and repeats. It is a measure of how much the software was used. For the Budget’s sentence to follow from it, the government would need to count patients rather than studies, and to show that more than four million of them got a diagnosis or result sooner than they would have without the software. The spreadsheet records neither who the patients were nor when any result was issued.

The running count passed four million in April 2026, two months before the June press release. The spreadsheet itself is dated July. The earliest copy of the June release in the Internet Archive, taken 28 minutes after publication on 10 June, and another from 11 June show the four-million sentence with no link; copies from 20 July onward link it to the spreadsheet. The page carries no note of the change. The release called the four million “patients” throughout.

What would show “faster”

A faster result means faster than the same patient would have got without the software. Reporting times also change for other reasons, such as workload and changes to the pathway around the software, so a before-and-after figure from trusts that installed AI cannot by itself separate the software’s effect from everything else. The two controlled evaluations examined here approach it differently: the fund’s evaluation compared trusts using AI with trusts that were not, and LungIMPACT randomised one use of the AI.

The government’s own evaluation

In October 2023 the National Institute for Health and Care Research commissioned its Rapid Service Evaluation Team, a collaboration of University College London, the Nuffield Trust and the University of Cambridge, to evaluate the implementation and outcomes of the AI Diagnostic Fund independently, across 66 NHS hospital organisations. Its second-phase report was published on 16 March 2026, with a full slide set of results. It is observational, combining detailed data from a few sites with national data, and its findings have not yet been peer reviewed.

Inside the trusts that adopted AI, reporting changed. Overall turnaround times fell after deployment. In six trusts of one imaging network that used the AI to prioritise abnormal images, X-rays that readers judged suspicious for cancer were turned around within 24 hours significantly more often than those judged normal. The evaluators then used national data to compare the fund’s trusts with trusts that had not deployed AI. Their summary of that comparison: “scans reported within 24 hours did not change significantly after accounting for changes in non-AI sites: other factors may have contributed.”

On the step that matters for a patient with a suspicious X-ray, the evaluators wrote: “Overall, AI deployment did not have a significant association with the time between x-ray and CT scan across the AIDF sites.” Their survival model, covering 45 trusts and CT scans within 28 days of the X-ray, gives a hazard ratio for AI deployment of 0.83 (0.67 to 1.02, p = 0.07); with the CT scan as the event, a ratio below 1 leans towards slower scans, and the result is not significant. One site saw a significant fall in the share of CT scans done within 72 hours, which the evaluators suggest may reflect CT capacity.

Their reading: “AI tools may support faster diagnosis of e.g. lung cancer. However, the time gained may be clinically negligible. Key improvements are likely to come from any AI contribution to decisions to follow-up cancers which may otherwise have been missed.” Their overall assessment: “Varied findings on impact and cost-effectiveness suggest value in further deployment, but more research is needed on effectiveness (with stronger baseline data and use of controls) and impact on immediate and longer-term patient outcomes (e.g. staging at diagnosis, mortality).”

This evaluation was published almost three months before the June release called the technology proven, and seven months before the Budget.

The randomised trial

LungIMPACT ran from July 2023 to December 2024 in five NHS trusts in England, with Leicester and Nottingham contributing the largest shares of X-rays. It covered 93,326 chest X-rays requested by GPs, from 86,945 patients. A single product, Qure.ai’s qXR, read every X-ray in both arms, and the people writing the reports could see its marks. What was randomised, by day, was prioritisation: on some days the X-rays the AI judged abnormal went to the top of the reporting list, and on others they did not. The trial was funded by SBRI Healthcare, and the authors thank Qure.ai’s team for implementing the AI. One author received a Qure.ai travel grant to attend a radiology congress in 2023, and another has held SBRI funding as a co-applicant with Qure.ai. The paper describes the analysis as independent, “with no influence from either the AI developers or the investigators.”

Prioritisation did what it is designed to do at the first step. Median reporting times were 47.0 hours on other days and 34.1 on prioritisation days, and the trial’s comparison of reporting times was statistically significant. None of the later steps showed a significant difference. That is not proof of no effect: for time to diagnosis the ratio of 0.98 has an interval from 0.83 to 1.16, so the trial cannot rule out a difference of about a sixth in either direction.

The authors’ explanation is that the report got faster but not fast enough to change what happened next, given “the limited capacity within Trusts to both report immediately and organize downstream tests and appointments.” Their conclusion: “AI prioritization of CXR requested by UK primary care has no significant impact on the lung cancer pathway. Therefore, CXR AI deployments should not include worklist prioritization in this context.” Of the ten sites the government’s evaluation studied, seven used their AI for prioritisation.

The trial has limits that matter here, and its own peer review aired them. One reviewer wrote that it was “more of an analysis of the prioritization strategy rather than the utility of an AI tool.” Another argued that with a 13-hour difference “AI was set up to fail,” and that a fair test would have held patients with a flagged X-ray for same-day review and CT. The authors replied that such a design “would have been a pathway change” whose effect could not be attributed to the AI. In the paper they write that the primary outcomes “evaluated only the impact of AI prioritization, not the presence or absence of AI in the clinical workflow.” Their cost-effectiveness preprint, below, compares AI with no AI using a control group from an earlier trial.

The October figures

The Treasury’s release four days before the Budget gave three more specific numbers: an 11.5% increase in suspected lung cancer cases identified within 24 hours, “around 400 more patients” receiving “faster clarity each month,” and an average wait for follow-up CT down from 6.3 to 5.6 days. It gave no source. We could not locate the analysis behind them in the Budget documents, in NHS England’s published data, in written parliamentary answers, or in the government’s January progress report on its AI action plan. That report matches the release’s “around a third,” saying “one-third of NHS chest X-rays – or 2.4 million scans – are now AI-assisted”; the other three figures are not in it. The June release has its own unsourced statistic: “Early data shows the technology helps radiologists analyse scans in an average of just 4 days, compared to 8 days for the most complex cases previously.”

NHS England also collected data of its own. The evaluators record that they did not report the fund’s “Benefit Metrics” because “publication was very delayed and it covered only about a third of all AIDF trusts.” Those metrics are one possible source of the Treasury’s figures, but we have not established whether the Treasury used them or what, if anything, its figures compare AI trusts with. Without the method, a reader cannot tell whether they account for the changes at trusts without AI that the evaluation found mattered.

We were not the first to put the October figures next to the trial. On 25 October, three days before the Budget, Kevin Yeandel of Tracked Changes noted that none of the Treasury’s figures carried a citation and set them against LungIMPACT’s results.

What NICE says

The National Institute for Health and Care Excellence assessed AI software that analyses chest X-rays for suspected lung cancer in adults referred from primary care in 2023, and last updated its early value assessment in February 2025. It lists 13 products, including Annalise CXR and qXR, both used in the NHS programme, and recommends more research. “Access to the technology should be through company, research or non-core NHS funding.” Centres already using it “may continue to do so but only under an appropriate evaluation framework and only alongside clinician review.” Its plain-language summary: “AI‑derived software should not be used for clinical decision making in the NHS until more evidence is available.” NICE’s committee named the then-ongoing LungIMPACT trial as one that “may provide useful evidence,” and judged that “ruling out normal X‑rays with high confidence is the most promising area for this software in clinical practice.” The guidance has not been revised since the trial reported.

A written question tabled on 9 October asked the government what NICE recommends on AI-assisted chest X-ray interpretation outside research settings. The answer, on 19 October, did not say. It said NICE’s 2023 assessment “specifically referenced the potential for the AI Diagnostic Fund to answer the evidence gaps set out in the guidance.” The fund’s commissioned evaluation is the one described above.

What the evidence does support

The software appears to do what it is built to do at the first step. In the six trusts where it was measured, the government’s evaluation found the tools broadly effective at prioritising X-rays suspicious for cancer. Its economic model, which counts quality-adjusted life years, found AI use cost-effective against care before deployment, less costly and more effective, with health gains it calls “negligible.” The LungIMPACT team’s own cost-effectiveness analysis, a preprint not yet peer reviewed, leans the other way on different measures. Its primary analysis counts the cost of each day cut from the time between X-ray and any diagnostic decision, a lung cancer diagnosis or discharge from the pathway, across all patients; on that measure prioritisation was not cost-effective against AI without it. Secondary analyses count days saved to a lung cancer diagnosis among those diagnosed, and compare AI with a no-AI group modelled largely on the standard-care arm of radioX, the group’s earlier trial; there, “no AI was always more cost-effective.” The uncertainty is wide. NICE sets no threshold for days saved, so the authors tested values from zero to “an arbitrary value of £25” a day, and at every one the probability that AI with prioritisation was the cost-effective choice, against either comparator, stayed below half. They caution that the no-AI estimates come “from a separate study population and not through a within-trial randomisation,” and the preprint discloses Qure.ai support for one co-author’s role on LungIMPACT. The two analyses use different outcomes and comparisons, and neither settles the other.

Both studies point to a possible benefit the Budget does not claim. In the trial, 53 of the 558 lung cancers were in patients whose X-ray the AI flagged as abnormal, for any finding, and the report called normal; their median time to diagnosis was 106 days, against 38 days when both the AI and the report flagged an abnormality. The comparison is observational, the reports were written with the AI’s marks available, and the authors found only a few cases of definite reporting error on review. The evaluators wrote that the main gains are likely to come from the AI’s contribution to following up cancers “which may otherwise have been missed.” Neither study has measured what acting on those flags would change.

Faster diagnosis has been shown in the NHS by changing the pathway. In an earlier randomised trial by the same group, radioX, immediate reporting by radiographers with direct triage to same-day CT cut the median time from X-ray to lung cancer diagnosis from 63 days to 32, without AI. Two published before-and-after studies of AI that did report faster times, at a single London trust (6.0 to 3.6 days from X-ray to CT report) and in a service evaluation in NHS Grampian, changed the pathway at the same time. They suggest an AI-supported service with the pathway redesigned around it can be faster than the old one; as the LungIMPACT authors note, they cannot separate the AI’s share. For what to try next, LungIMPACT’s authors suggest a design in which “any AI-flagged abnormality prompts an immediate human review and, if confirmed, the implementation of a downstream bundle of investigations.”

What would settle it

Four things would let anyone check the government’s figures:

• the method behind the 11.5%, the “around 400” and the 6.3 to 5.6 days, including which trusts, which months and what they were compared with, and NHS England’s Benefit Metrics;
• time from X-ray to CT at trusts using AI against matched trusts without it, which the evaluation measured, and time to diagnosis, which only the trial measured;
• counts of patients, separately from counts of studies processed, with how much sooner their reports and results were issued than at comparable trusts without AI;
• how many cancers the AI flagged that readers did not, and how quickly those patients were diagnosed.

We could also find no published specification for the £20 million extension saying whether it will configure the software to prioritise reporting lists, the use the LungIMPACT authors advise against for GP-referred X-rays.

How we did this

Sources. All documents were read at their published addresses on 31 October 2026, except NHS England’s spreadsheet, which its own site blocked from automated download. We used the copy on a public mirror of NHS England Digital’s site, at the same path as the official file. Its 25,773 bytes match the 25.8 kB the official page lists, and two downloads gave the same SHA-256 hash, 244414046f01ae65…ec442e4133; a matching size does not prove the files are identical. Earlier states of the June press release come from the Internet Archive; the page’s change history comes from GOV.UK’s content API.

The count. We summed the spreadsheet’s monthly national totals. They reach 4,280,041 by May 2026; the remaining 23,023 are one trust, The Christie, whose figure is given only as a cumulative total. The trust rows add up to 4,303,063, one fewer than the national row; the gap is in a single month, November 2024, where Cheshire and Merseyside’s network subtotal (cell U17) is 324 and its trusts (U5 to U16) sum to 323. The notes give “reconciled cumulative” totals for three Greater Manchester trusts that exceed their listed figures by 4,788 studies in all, which the notes say cannot be assigned to a month; the listed figures, and so the grand total, do not include them. As a check, January to March 2026 sums to 781,000, the figure in the ministerial answer to written question 30882. The month-by-month reconciliation, with the file’s full hash, is published as a data file, with the script that produces it from NHS England’s spreadsheet.

The studies. LungIMPACT figures are from the paper’s Table 2, its results and discussion, and the peer review file published with it. The evaluation’s figures are from the phase 2 slide set, including the regression and survival models in its appendix, and the report page. The evaluation’s national model of reports within 24 hours includes interaction terms, so we quote the evaluators’ summary of it rather than a single coefficient.

The October figures. We searched the Budget documents, NHS England’s published statistics, written parliamentary questions and answers from 2024 to October 2026, and the AI action plan progress report for the 11.5%, the 400 and the 6.3 to 5.6 days, and could not locate the analysis behind them.

The rating. Rubric version 1.0, applied to the Budget sentence. The rating rests on what the cited source measures. It does not depend on reading the non-significant results as proof that the software has no effect.

Questions readers ask

For the general version of this problem, how many people a tool reached and what it changed for them, see our field guide How to read a health-chatbot impact claim. Our earlier NHS investigation, The NHS cancer blood test was “99% accurate.” Which 99%?, traces another cancer-pathway figure to its source.

Sources

• HM Treasury, Budget 2026, 28 October 2026, section 3.3 and footnote 66.
• NHS England, “Chest X-ray AI – Studies Successfully Processed by NHS Trust, by Month”, July 2026.
• Department of Health and Social Care, “AI to speed up cancer diagnosis for millions of NHS patients”, 10 June 2026; Internet Archive copies of 10 June and 20 July.
• HM Treasury, “Chancellor unveils Britain’s ambitious plan for AI”, 24 October 2026.
• UK Parliament, written questions 30467 and 30882, tabled 9 October and answered 19 October 2026.
• NIHR Rapid Service Evaluation Team, “Artificial intelligence in chest diagnostics: mixed-methods evaluation (phase 2)”, 16 March 2026, and slide set, 13 March 2026.
• Woznitza N, Baldwin DR et al., “AI-based chest X-ray prioritization in the lung cancer diagnostic pathway: the LungIMPACT randomized controlled trial”, Nature Medicine 32, 1737–1744 (2026), with its peer review file; statistical code on GitHub.
• Baldwin D, Ndwandwe S, Smith L, Rawlinson J et al., “Cost-effectiveness of immediate prioritisation of chest X-rays using artificial intelligence in lung cancer in England: the LungIMPACT study”, Research Square preprint, July 2026.
• NICE, HTG696, “Artificial intelligence (AI)-derived software to analyse chest X-rays for suspected lung cancer in primary care referrals: early value assessment”, published 28 September 2023, last updated 10 February 2025, and its committee discussion.
• Woznitza N et al., “Impact of radiographer immediate reporting of X-rays of the chest from general practice on the lung cancer pathway (radioX): a randomised controlled trial”, Thorax 78, 890–894 (2023).
• Storey M et al., “Early clinical evaluation of AI triage of chest radiographs”, NEJM AI (2025); Healthcare Improvement Scotland, “Artificial intelligence-assisted clinician review of chest X-rays for suspected lung cancer” (February 2025).
• Department for Science, Innovation and Technology, “AI Opportunities Action Plan: one year on”, January 2026.
• Kevin Yeandel, Tracked Changes, 25 October 2026.