More than seven million patients sit on the NHS waiting list in England, alongside around 100,000 staff vacancies. The queue is not a side effect of the system; in places it is the rationing mechanism.
BBC News, 2025
Corroborated3 in 4 patients surveyed say they have waited too long to see a specialist, with their longest wait averaging 56 days.
WSJ paid programme (sponsor survey), 2025
Physicians spend nearly two hours on electronic health records for every hour of direct patient care, much of it after the clinic closes.
Vaid et al. (Univ. of Alberta), 2026
About one in six adults (17%) already ask AI chatbots for health information and advice at least once a month.
Bhatt et al., 2024
Diagnosis, detection, decision support and drug discovery: the gains the research can put a number on.
Microsoft's MAI-DxO, paired with OpenAI's o3, reached 80% diagnostic accuracy on complex case records, four times the 20% average of generalist physicians on the same cases.
Nori et al. (Microsoft AI), 2025
Company-reportedCheXNet, a deep-learning model reading chest X-rays, scored an F1 of 0.435 for pneumonia against a 0.387 average for four practising radiologists on the same images.
Rajpurkar et al. (Stanford), 2017
Peer-reviewedIn a blinded text-chat consultation study, Google's AMIE outperformed primary-care physicians on 28 of 32 axes rated by specialists and 24 of 26 rated by patient actors.
Tu et al. (Google), 2024
Company-reportedAcross 39,849 patient visits in 15 Nairobi primary-care clinics, clinicians with an AI decision-support tool made 16% fewer diagnostic errors and 13% fewer treatment errors.
Korom et al. (Penda Health / OpenAI), 2025
Company-reportedUsing AlphaFold-predicted structures in an AI drug-discovery pipeline, researchers found a hit compound for a novel cancer target within 30 days, after synthesising only 7 compounds.
Ren et al. (Insilico Medicine), 2022
Company-reportedA randomised trial of 238 physicians across 72,000 encounters found ambient AI scribes cut documentation time by up to 9.5% and improved burnout scores.
Lukac et al. 2025, cited in Karbalaie et al., 2026
Peer-reviewedIncluding ethnicity in a diabetes-risk model, rather than excluding it, narrowed the net-benefit gap between the most and least benefited groups (white and Asian) by 48 (95% CI 32–66).
Benitez-Aurioles et al., 2024
Single-sourceAn interpretable model flagged radiological signs of acute heart failure on chest CT with a test AUC of 0.87, potentially before the formal radiology report is written.
Ørting et al. (Univ. of Copenhagen), 2025
Single-sourceNo verified figure exists for how many earlier diagnoses changed the treatment actually given. Systems record when a diagnosis was made, not what would have happened without it.
Nobody publishes the share of remote consultations that ended in an in-person visit anyway. The systems that hold the data do not release it.
Skewed data, gains at scale, open-source tools and the thin evidence on what any of it costs.
Across 8,594 tumour samples from 33 cancer types in TCGA, widely used to train cancer models, 82.0% came from White patients, 10.1% from Black and 7.5% from Asian patients.
Chen et al. (Harvard Medical School), 2021
Single-sourceApplied across Penda Health's clinics, the AI tool would avert diagnostic errors in about 22,000 visits and treatment errors in 29,000 visits a year.
Korom et al. (Penda Health / OpenAI), 2025
Company-reportedAn open-source AI scribe used by Alberta emergency physicians cost under $30 per physician per month, 70–95% less than commercial tools.
Vaid et al. (Univ. of Alberta), 2026
Single-sourceThe same diagnostic accuracy can cost very different amounts: orchestrated MAI-DxO reached 79.9% at $2,397 a case, where off-the-shelf o3 reached 78.6% at $7,850.
Nori et al. (Microsoft AI), 2025
Company-reportedA systematic review of AI's economic impact in health care could take only 6 of 66 publications forward, and none of those had a complete cost-impact analysis.
Wolff et al., J Med Internet Res, 2020
Peer-reviewedThe effect of pooled booking on clinician retention is unmeasured. Systems reporting fewer vacancies over the same period cannot separate booking policy from pay settlements.
No source puts a national figure on what a wait costs patients: time off work, childcare, travel. Small surveys exist; none are comparable.
Bias in triage, errors in the notes, misinformation on request and tools priced beyond reach.
Swapping only the patient's gender in otherwise identical emergency vignettes flipped a model's triage level up to 43.8% of the time; all five models tested exceeded the pre-registered 5% threshold.
Young et al. (EQUITRIAGE), 2026
Single-sourceIn feedback on 50,123 AI-scribed encounters, the largest category of safety issues, 18.5%, was medication names, doses or instructions recorded wrongly.
Dai et al. (UC Berkeley / UCSF), 2025
Single-sourceInstructed to give false health answers with fabricated references, four of five leading chatbots did so 100% of the time; only Claude refused more than half the time.
Reuters, on Modi et al. (Annals of Internal Medicine), 2025
Peer-reviewedCommercial AI scribes cost $99 to $600 per physician per month, out of reach for rural clinics, safety-net systems and lower-income countries.
Vaid et al. (Univ. of Alberta), 2026
Single-sourceApplied mechanically to France's 20.9 million annual emergency visits, a 1.1% gender differential in triage would mean on the order of 110,000 more lower-severity assignments for women a year (illustrative range 90,000–130,000).
Guerra-Adames et al., 2025
Single-sourceOne estimate puts the burden of hospital readmission at $17.9 billion a year, with 76% of readmissions avoidable: the strain risk-prediction tools target.
Huang et al. (ClinicalBERT), 2019
Single-sourceIn the same audit, DeepSeek-V3.1 undertriaged women more than twice as often as men (F/M 2.15:1).
Young et al. (EQUITRIAGE), 2026
Single-sourceConsumer symptom checkers show top-1 diagnostic accuracy from 5% to 50%, reaching 95% only when 20 possible conditions are counted.
Bhatt et al., 2024
Single-sourceFewer than 2% of FDA-cleared AI medical devices report having been updated by retraining on new data, while nearly a quarter report updates as new features.
Wu et al., 2024
Peer-reviewedWide cost ranges, models that comply with bad instructions and risks no one has yet sized.
Off-the-shelf models that reached 70–79% accuracy did so by ordering extensive tests, at $4,000 to $7,900 a case. Physicians on the same benchmark averaged $2,963 a case at 20% accuracy.
Nori et al. (Microsoft AI), 2025
Company-reportedFive leading models were set up to answer health questions falsely, with invented references to real journals. Four complied every time; Claude 3.5 Sonnet refused more than half the time.
Reuters, on Modi et al. (Annals of Internal Medicine), 2025
Peer-reviewedFeedback on 50,123 encounters from 470 physicians points to real safety risks in AI scribes, especially around medication, yet the authors say the absolute degree of risk still needs study.
Dai et al. (UC Berkeley / UCSF), 2025
Single-sourceAll five models audited, across 374,275 evaluations on 18,714 vignettes, changed triage levels by gender more often than the pre-registered 5% limit.
Young et al. (EQUITRIAGE), 2026
Single-sourceNearly a quarter of FDA-cleared AI devices report updates in the form of new features or claims, but almost none report retraining, so how they change after approval is largely unseen.
Wu et al., 2024
Peer-reviewedBoth sides of this page are measured to the same standard and given the same weight. Every figure carries a source and a confidence label; where none existed, the page says so instead of estimating. The trade-off is real; what it should cost a patient is not a number this research can supply.
Lukac et al.'s scribe trial is cited via Karbalaie et al., and the chatbot study via Reuters' report of it.
