AI in Healthcare Statistics 2026: Adoption, FDA, Accuracy

AI in healthcare statistics for 2026 show clinical AI spreading faster than the evidence for it. The FDA's list of AI-enabled medical devices held 1,614 entries with decision dates through 29 June 2026 (FDA, Jun 2026), yet a review of 519 studies of large language models in health care found that only 5% used real patient care data (JAMA, Bedi et al., Jan 2025).

The page covers hospital and clinician use, patients, device authorisations, AI scribes and burnout, diagnostic accuracy, drug discovery, investment, safety and rules. Each one links to the organisation that produced it: government agencies, regulators, peer-reviewed journals, and survey firms and vendors that are labelled as such. For the wider picture, see our AI statistics, generative AI statistics and AI usage statistics pages.

Key Takeaways

71% of US hospitals used predictive AI built into their health record in 2024, up from 66% in 2023 (ASTP/ONC, Sep 2025).
Only 9% of doctors and 18% of nurses say they never use AI at work, against 43% and 34% a year earlier (Wolters Kluwer, Mar 2026).
32% of US adults used AI chatbots for health information or advice in the past year (KFF, Mar 2026).
14% of recent AI health users skipped a provider visit after an AI answer, about 14 million US adults (Gallup and West Health, Apr 2026).
72% of US adults say it is extremely or very important that their provider tells them when AI is used in their care (Pew Research Center, Aug 2026).
The FDA's list shows 335 AI-enabled devices with 2025 decision dates, up from 235 in 2024 and 18 in 2016 (FDA, Jun 2026).
Radiology accounts for 1,230 of the 1,614 devices on the FDA's list, or 76.2% (FDA, Jun 2026).
In a randomised trial of 238 UCLA outpatient physicians, one AI scribe cut time in notes by 9.5% and the other produced no significant change (NEJM AI, Lukac et al., Dec 2025).
AI-supported mammography screening found cancers at 6.4 per 1,000 women against 5.0 with standard double reading, and cut screen-reading workload by 44.2% (Lancet Digital Health, Hernström et al., Feb 2025).
Language models alone identified relevant conditions in 94.9% of medical scenarios, but people using them identified fewer than 34.5% (Nature Medicine, Bean et al., Feb 2026).
ChatGPT Health undertriaged 52% of gold-standard emergencies in a stress test of 960 responses (Nature Medicine, Ramaswamy et al., Feb 2026).
The adenoma detection rate in colonoscopies without AI fell from 28.4% to 22.4% after endoscopists had worked with AI (Lancet Gastroenterology and Hepatology, Budzyń et al., Aug 2025).
AI companies took 46% of all healthcare investment in 2025, more than $18 billion (Silicon Valley Bank, Jan 2026).
Only 8% of the 50 responding WHO European countries have published a health-specific AI strategy (WHO Europe, Nov 2025).

How fast is healthcare AI adoption growing in hospitals and clinics?

Hospitals adopted first and are still working out how to govern it. Most run AI inside their health record software, and the gap between large systems and small independent hospitals is widening.

AI in healthcare statistics: share of US hospitals using each type of predictive AI, 2023 and 2024 — Figure 1: Billing and scheduling were the fastest-growing uses of hospital predictive AI in a single year, while inpatient risk scoring was already near-universal among hospitals that used any predictive AI. Source: ASTP/ONC, Hospital Trends in the Use, Evaluation, and Governance of Predictive AI, 2023-2024, Sep 2025. Chart published by the source, captured 14 Oct 2026.

Which hospitals use predictive AI, and for what?

System-affiliated hospitals used predictive AI at 86% in 2024, against 37% for independent hospitals (ASTP/ONC, Sep 2025).

96% of large hospitals (400 or more beds) used predictive AI in 2024, against 59% of small ones (100 beds or fewer) (ASTP/ONC, Sep 2025).
81% of urban hospitals used it, against 56% of rural hospitals and 50% of critical access hospitals (ASTP/ONC, Sep 2025).
90% of hospitals on the market-leading health record vendor used predictive AI, against 50% of hospitals on other vendors (ASTP/ONC, Sep 2025).
Use of predictive AI to simplify billing rose from 36% to 61% in a year, and use for scheduling rose from 51% to 67% (ASTP/ONC, Sep 2025).
80% of hospitals used predictive AI from their health record developer, 52% used third-party tools and 50% used models they built themselves (ASTP/ONC, Sep 2025).
82% of hospitals evaluated their predictive models for accuracy, 74% for bias and 79% monitored them after deployment (ASTP/ONC, Sep 2025).
74% of hospitals said several entities were accountable for evaluating predictive AI, 66% named a dedicated committee and 41% named IT staff (ASTP/ONC, Sep 2025).

Predictive AI means risk scores and forecasts, the older layer of clinical AI. The gaps line up with hospital size, ownership and software vendor.

How many hospitals use generative AI, and is it working?

31.5% of US hospitals (762) used generative AI integrated with their health record in 2024, and another 24.7% planned to within a year (JAMA Network Open, Everson et al., Dec 2025).

32% of hospitals did not use generative AI and had no plans to (JAMA Network Open, Everson et al., Dec 2025).
In a survey of 43 US health systems with a 64% response rate, every one had at least begun developing or piloting ambient notes, but only 14% had deployed them fully (JAMIA, Poon et al., Jul 2025).
90% of those systems had deployed imaging and radiology AI in at least some areas (JAMIA, Poon et al., Jul 2025).
53% reported a high degree of success with clinical documentation AI, 38% with clinical risk stratification and 23% with revenue cycle AI (JAMIA, Poon et al., Jul 2025).
77% cited immature AI tools as a barrier, 47% financial concerns and 40% regulatory uncertainty (JAMIA, Poon et al., Jul 2025).

Pilots are common and full deployment is rare. The buyers themselves say the tools are not mature.

How many doctors and nurses use AI at work?

54% of 3,151 US physicians said they were using AI in clinical practice, and the share rose from 47% in spring 2025 to 63% in the latest wave (Doximity, Jan 2026). Doximity runs a platform for physicians, so treat this as vendor research.

74% of doctors and 70% of nurses used AI at least weekly for professional purposes in March 2026, against 38% and 46% a year earlier (Wolters Kluwer, Mar 2026). The vendor survey covered 355 clinicians.
48% of 2,206 clinicians in 109 countries had used AI for work in spring 2025, up from 26% in 2024 (Elsevier, Jul 2025). This is also vendor research.
Of those who used AI for work, 97% had used a generalist tool such as ChatGPT and 76% a clinical-specific tool (Elsevier, Jul 2025).
21% of clinicians used AI to analyse medical images and 16% to help make clinical decisions, while another 48% wanted to use it for clinical decisions (Elsevier, Jul 2025).
Literature search was the leading use at 35% of physicians in January 2026, ahead of AI scribes at 29% (Doximity, Jan 2026).
Neurologists reported the highest use at 64%, followed by gastroenterologists at 61% and internal medicine physicians at 60% (Doximity, Jan 2026).
47% of physicians said their institution's AI decision process was still evolving, only 8% said it was clear, and 71% named accuracy and reliability as their top concern (Doximity, Jan 2026).
Only 30% of clinicians said their institution gave adequate AI training and 29% said it had effective AI governance (Elsevier, Jul 2025).

The three surveys disagree on the level because they define use differently, from any use at work to weekly professional use. They agree on direction, and on the lag in training and policy behind the tools.

How many patients use AI for health questions, and do they trust it?

Patients are ahead of their providers. More than a quarter of adults already ask a chatbot about their health, often because a visit is too slow, too costly or too dismissive.

Share of US adults who say a provider should tell them when AI is used for each healthcare task, 2026 — Figure 2: Eight in ten Americans want to be told when AI reads a scan, makes a diagnosis or explains lab results, and even scheduling draws majority support for disclosure. Source: Pew Research Center, Americans want transparency when AI is used in their healthcare, Jun 2026. Chart published by the source, captured 14 Oct 2026.

25% of US adults have used an AI tool or chatbot for health information or advice, according to a Gallup panel of 5,660 adults (Gallup and West Health, Apr 2026).

That equals over 66 million Americans (West Health, Apr 2026).
29% of adults used AI in the past year for information about their physical health and 16% for their mental health (KFF, Mar 2026).
28% of adults aged 18 to 29 used AI for mental health advice, against 8% of those aged 50 and over (KFF, Mar 2026).
27% of adults used AI to look up symptoms or general health information, 19% to understand test results and 16% to decide whether to see a doctor (KFF, Mar 2026).
65% of AI health users named quick, immediate advice as a major reason, 19% could not afford a provider and 18% had no regular provider or could not get an appointment (KFF, Mar 2026).
59% of recent users researched with AI before a doctor visit and 56% after one (Gallup and West Health, Apr 2026).
21% of recent users said they used AI because a provider had dismissed or ignored them in the past (Gallup and West Health, Apr 2026).
61% of recent users used general conversational AI such as ChatGPT or Copilot, and 55% used AI summaries inside web search (Gallup and West Health, Apr 2026).
More than 40 million people ask ChatGPT health questions every day, and one in four of its 800 million regular users submits a healthcare prompt each week (OpenAI, Jan 2026). This is the vendor's own usage data.

Patients now ask an assistant before they call a clinic, so which pages an assistant cites about a practice has become a measurable question.

Do patients trust AI health answers?

Only partly. 4% of recent AI health users strongly trust the accuracy of what they get (Gallup and West Health, Apr 2026).

33% of recent users trust AI health information, 33% neither trust nor distrust it and 34% distrust it (Gallup and West Health, Apr 2026).
11% of recent users said AI recommended health advice they believed was unsafe (Gallup and West Health, Apr 2026).
69% of adults who used AI for physical health advice trust it a great deal or a fair amount, against 18% of adults who have not used it (KFF, Mar 2026).
58% of those who used AI for physical health advice later followed up with a doctor, and 42% of mental health users followed up with a professional (KFF, Mar 2026).
77% of adults are concerned about the privacy of medical information given to AI tools, yet 41% of AI health users have uploaded test results or doctor's notes (KFF, Mar 2026).
In a 2022 survey of 11,004 adults, 60% said they would be uncomfortable if their own provider relied on AI to diagnose disease and recommend treatments (Pew Research Center, Feb 2023).
57% in that survey expected AI to make the patient-provider relationship worse, against 13% who expected it to improve (Pew Research Center, Feb 2023).

Patients are comfortable using a chatbot on their own and less comfortable when it sits between them and a clinician.

Do patients want to be told when AI is used in their care?

Yes. 81% want a provider to tell them when AI analyses scans or makes a diagnosis, and 56% want to be told even when it schedules an appointment (Pew Research Center, Aug 2026).

80% want to be told when AI explains lab results and 64% when it orders prescription refills (Pew Research Center, Aug 2026).
53% of adults say they have little or no say over whether a provider uses AI in their care (Pew Research Center, Aug 2026).
63% want more say over AI use in their care, against 21% who are comfortable with the say they have (Pew Research Center, Aug 2026).
Only 16% say a provider has used AI in their care, 37% say none has and 46% are not sure (Pew Research Center, Aug 2026).
Among patients who know AI was used, 22% understand well how it was used and 32% understand it not too well or not at all (Pew Research Center, Aug 2026).

The survey was fielded in June 2026, when hospitals were already using AI for notes and scans.

How many AI medical devices has the FDA authorised, and what evidence backs them?

The device list grows every year, and almost all of it is imaging software cleared by comparison with earlier devices rather than tested in new trials.

US FDA AI medical devices authorised per year, 1995 to 2025, with the cumulative total — Figure 3: Annual FDA authorisations of AI medical devices stayed in single digits until 2016 and passed 250 in 2025, with the cumulative total above 1,300. The FDA list has since gained more entries for 2025, which is why counts in the text are higher. Source: Stanford HAI, AI Index Report 2026, Medicine chapter, Apr 2026, using FDA data. Chart published by the source, captured 14 Oct 2026.

How many AI-enabled devices are authorised each year?

The cumulative total crossed 1,000 devices in 2024, and by December 2025 the FDA had authorised 1,357 AI-enabled devices from 693 companies across 17 clinical specialties (Stanford HAI, Apr 2026).

The list held 863 entries with decision dates through 2023, 1,098 through 2024 and 1,433 through 2025, counting rows in its CSV download (FDA, Jun 2026).
It holds 181 entries dated in the first half of 2026 against 176 in the first half of 2025 (FDA, Jun 2026).
The AI Index counted 258 authorisations through September 2025, already more than any earlier full year (Stanford HAI, Apr 2026).
98 new companies entered the market in 2025, after 109 in 2024 and 103 in 2023 (Stanford HAI, Apr 2026).
Siemens Medical Solutions USA has 56 entries, Canon Medical Systems 42 and Aidoc Medical 30 (FDA, Jun 2026).
The number of 510(k)-cleared AI imaging-related devices reached 246 in 2025, from 16 in 2016 (Stanford HAI, Apr 2026).

The FDA says its list is not comprehensive and identifies devices mainly by AI-related terms in their summaries (FDA, Jun 2026). Counts differ by date because decision summaries are added later.

Which specialties and approval routes dominate?

Almost all AI devices go through the 510(k) route: 1,553 of the 1,614 entries (96.2%) carry 510(k) numbers, against 40 De Novo and 21 PMA entries (FDA, Jun 2026).

Cardiovascular is second with 154 entries (9.5%) and neurology third with 73 (4.5%) (FDA, Jun 2026).
Anaesthesiology has 30 entries, gastroenterology-urology 27, hematology 22, and pathology and ophthalmic 10 each (FDA, Jun 2026).
81 non-radiology entries have 2025 decision dates, against 7 in 2016 (FDA, Jun 2026).
Predetermined change control plans, which allow pre-agreed updates after authorisation, were used in about 10% of 2025 clearances (Stanford HAI, Apr 2026).
A review of 1,016 authorisations found more than 100 devices that use AI to generate data, and none that involved a large language model (npj Digital Medicine, Singh et al., Jul 2025).
The FDA says it will explore ways to tag devices that use foundation models, from large language models to multimodal systems (FDA, Jun 2026).

The 510(k) route asks whether a device is substantially equivalent to an earlier one. It does not require a new clinical trial.

How strong is the evidence, and how often are devices recalled?

Recalls are uncommon but cluster early: 60 of 950 AI-enabled devices (6.3%) were associated with 182 recall events, and 43.4% of those events came within 12 months of clearance (JAMA Health Forum, Lee et al., Aug 2025).

That early-recall rate is about double the rate reported for all 510(k) devices, and 59.3% of recalls were unresolved when the study ended (JAMA Health Forum, Lee et al., Aug 2025).
Recall-free survival was 96.6% at one year, 93.5% at three years and 91.8% at five years (JAMA Health Forum, Lee et al., Aug 2025).
Publicly traded companies made 53.2% of the devices but accounted for 91.8% of recalls (JAMA Health Forum, Lee et al., Aug 2025).
Lack of clinical validation was independently associated with recall, with an odds ratio of 2.8 (JAMA Health Forum, Lee et al., Aug 2025).
In a separate study of 903 devices authorised through August 2024, 43 (4.8%) were recalled after a median of 458 days (JAMA Network Open, Ren et al., Jun 2026).
Devices that lacked information on supporting clinical studies had a higher hazard of recall, with a hazard ratio of 1.39 (JAMA Network Open, Ren et al., Jun 2026).

Both studies point the same way: devices with less clinical testing fail more often after release.

What do AI scribes and automation change for clinicians, and who pays?

Scribes are the first clinical AI to reach routine use. The trial evidence is real but smaller than the before-and-after surveys suggest.

Effect of ambient AI on professional fulfilment and burnout in a randomised trial of 66 practitioners — Figure 4: In a stepped-wedge randomised trial, ambient AI reduced burnout by 0.44 points while the gain in professional fulfilment stayed small, so the benefit is relief from exhaustion more than extra satisfaction. Source: NEJM AI, Afshar et al., A Pragmatic Randomized Controlled Trial of Ambient Artificial Intelligence, Dec 2025. Chart published by the source, captured 14 Oct 2026.

How much documentation time do AI scribes save?

At The Permanente Medical Group, 7,260 physicians used an AI scribe in 2,576,627 patient encounters between October 2023 and December 2024, and an analysis found it saved nearly 16,000 hours of documentation time (Kaiser Permanente Division of Research, Apr 2025).

3,447 of those physicians used the scribe in at least 100 encounters each (Kaiser Permanente Division of Research, Apr 2025).
88% of 102 surveyed physicians said the scribe had a positive effect on their visit interactions (Kaiser Permanente Division of Research, Apr 2025).
In the UCLA trial, the two scribes were used in 33.5% and 29.5% of visits, and physicians rated clinically significant inaccuracies as occurring only occasionally (NEJM AI, Lukac et al., Dec 2025).
In a Wisconsin trial of 66 practitioners, ambient AI generated 38% of 71,487 notes and cut time on notes by 0.36 hours a day (NEJM AI, Afshar et al., Dec 2025).
An NHS England-sponsored study found a 23.5% increase in direct patient interaction time and an 8.2% reduction in overall appointment length when AI scribes were used (NHS England, Jul 2026).
NHS England in the Midlands procured one ambient voice supplier for all 1,239 GP practices and over 70,000 clinicians across 15 trusts (NHS England, Jul 2026).
In an evaluation of 12,019 enrolled Canadian primary care clinicians, survey respondents reported saving an average of 1.11 hours a week during work hours after six months (Canada Health Infoway, Jun 2026).
Same-day note completion rose from 55.9% to nearly 75% in that programme (Canada Health Infoway, Jun 2026).
Where scribes are widely available, organisations told the Peterson Health Technology Institute that adoption is typically 20% to 50% (PHTI, Mar 2025).

Savings are real, modest and uneven. Heavy users gain most, and one of the two UCLA products showed no measurable time saving at all.

Do AI scribes reduce burnout, and how big is the staffing gap?

Burnout fell in every study, but randomised trials show smaller effects than before-and-after surveys. In the Wisconsin trial, ambient AI reduced work exhaustion by 0.44 points on a five-point scale, and the authors judged the rise in professional fulfilment not significant (NEJM AI, Afshar et al., Dec 2025).

In a six-system study of 263 clinicians, burnout fell from 51.9% to 38.8% after 30 days with an AI scribe, in a study with authors from the scribe vendor Abridge (JAMA Network Open, Olson et al., Oct 2025).
At Mass General Brigham, burnout among surveyed clinicians fell from 50.6% to 29.4% at 42 days, though only 30.4% answered the survey (JAMA Network Open, You et al., Aug 2025).
45.2% of US physicians reported at least one burnout symptom in 2023, down from 62.8% in 2021, and physicians were at higher risk than other US workers with an odds ratio of 1.82 (Mayo Clinic Proceedings, Shanafelt et al., Jul 2025).
The AAMC projects a shortage of up to 86,000 US physicians by 2036 (AAMC, Mar 2024).
WHO projects a global shortage of 11.1 million health workers by 2030, and nearly one in four doctors is older than 55 (WHO, Sep 2026).
31% of clinicians were considering leaving their role in the next couple of years, and 69% were seeing more patients than two years earlier (Elsevier, Jul 2025).
88% of physicians surveyed believe AI can help reduce burnout and improve job satisfaction (Doximity, Jan 2026).

Whether they fix burnout rather than ease it is still being tested.

Does AI pay for itself, and how much money is flowing in?

At UCSF, physicians who adopted ambient AI scribes generated 1.81 more relative value units a week than non-adopters, worth about $3,044 a year per physician at 2025 Medicare rates, with no change in claim denials (JAMA Network Open, Holmgren et al., Jan 2026).

Adopters also completed 0.80 more encounters a week, and 698 of the 1,565 physicians (44.6%) were adopters (JAMA Network Open, Holmgren et al., Jan 2026).
Healthcare AI spending reached $1.4 billion in 2025, nearly triple 2024, and ambient scribes generated $600 million (Menlo Ventures, Oct 2025). The report comes from a venture firm, so treat it as vendor research.
85% of generative AI spend in healthcare goes to startups rather than incumbents (Menlo Ventures, Oct 2025).
Large health systems expect scribe adoption to grow from 35% to only 40% over three years, and 67% of outpatient providers say they are likely to switch vendors (Menlo Ventures, Oct 2025).
US and European healthtech raised about $7.2 billion in the first half of 2026, its best half since 2022, and the top six deals made up about 50% of it (Silicon Valley Bank, Aug 2026).
Grand View Research puts the 2026 market at $50.7 billion, rising to $505.6 billion by 2033 (Grand View Research, Oct 2026).
Precedence Research, in a report by analyst Deepa Pandey, puts the 2026 market at $51.20 billion, rising to $744.34 billion by 2035 (Precedence Research, Sep 2026).
MarketsandMarkets puts the 2026 market at $36.67 billion, rising to $194.79 billion by 2031 (MarketsandMarkets, Oct 2026).

Market forecasts for the same year differ by almost $15 billion, because each firm defines the market differently. Treat all three as estimates by named firms rather than facts.

How accurate is AI medical diagnosis, and does it speed up drug discovery?

Narrow imaging tools have the best evidence. General chatbots are strong when tested alone and weak in the hands of real users.

Accuracy of language models alone versus users with the models in identifying conditions and dispositions in medical scenarios — Figure 5: Language models identified the relevant condition in nearly every scenario when tested alone, yet participants using the same models did no better than a control group that used any source it liked. Source: Nature Medicine, Bean et al., Reliability of LLMs as medical assistants for the general public, Feb 2026. Chart published by the source, captured 14 Oct 2026.

Does AI improve cancer screening?

In the Swedish MASAI randomised trial of 105,934 women, sensitivity was 80.5% with AI-supported screening against 73.8% with standard double reading, and specificity was 98.5% in both groups (The Lancet, Gommers et al., Jan 2026).

Interval cancer rates were 1.55 per 1,000 with AI against 1.76 without, a non-inferior ratio of 0.88 (The Lancet, Gommers et al., Jan 2026).
The AI group had fewer invasive interval cancers, 75 against 89 (The Lancet, Gommers et al., Jan 2026).
The earlier screening analysis found 270 invasive cancers with AI against 217, and no significant rise in recalls or false positives (Lancet Digital Health, Hernström et al., Feb 2025).
In the German PRAIM study of 463,094 women and 119 radiologists, AI-supported screening detected 6.7 cancers per 1,000 against 5.7, a 17.6% higher rate (Nature Medicine, Eisemann et al., Jan 2025).
The recall rate was 37.4 per 1,000 with AI against 38.3 without, and the positive predictive value of recall was 17.9% against 14.9% (Nature Medicine, Eisemann et al., Jan 2025).
An AI mammography system evaluated on UK and US screening data cut false positives by an absolute 5.7% in US data and 1.2% in UK data, and cut false negatives by 9.4% and 2.7% (Nature, McKinney et al., Jan 2020).
In a meta-analysis of 14 head-to-head imaging studies, deep learning had a pooled sensitivity of 87.0% against 86.4% for health professionals (Lancet Digital Health, Liu et al., Sep 2019).
Only 25 of the 82 studies in that review were externally validated (Lancet Digital Health, Liu et al., Sep 2019).

Screening is where AI has moved from retrospective tests to randomised trials, and the results are consistent: more cancers found without more false alarms, plus less reading work.

Can chatbots and language models diagnose better than doctors?

Alone, sometimes; as a doctor's aid, not yet. In a randomised trial of 50 physicians, diagnostic reasoning scored 76% with a language model against 74% with conventional resources, a 2-point adjusted difference that was not significant, while the model alone scored 16 points higher than the conventional group (JAMA Network Open, Goh et al., Oct 2024).

In the Oxford study of 1,298 participants and ten scenarios, models alone chose the right disposition in 56.3% of cases on average, and participants using them chose it in fewer than 44.2% (Nature Medicine, Bean et al., Feb 2026).
ChatGPT Health was tested on 60 clinician-written vignettes across 21 clinical domains, and failures concentrated at the extremes: 35% for nonurgent cases and 48% for emergencies (Nature Medicine, Ramaswamy et al., Feb 2026).
When family or friends minimised symptoms in the vignettes, triage shifted toward less urgent care, with an odds ratio of 11.7 (Nature Medicine, Ramaswamy et al., Feb 2026).
Microsoft's MAI-DxO orchestrator reached 80% accuracy on 304 NEJM cases against 20% for generalist physicians, and 85.5% when configured for maximum accuracy (Microsoft, Nori et al., Jun 2025). This is a vendor preprint, and the physicians were asked not to use search engines.
Of 519 studies of large language models in health care, 44.5% tested medical knowledge such as licensing exam questions and 19.5% tested diagnosis (JAMA, Bedi et al., Jan 2025).

Benchmark scores run higher than results with real participants. The two largest tests with real participants show that access to a capable model does not by itself produce better decisions.

Is AI finding new drugs faster?

Early signs are good for getting drugs into humans, less clear for getting them approved. AI-discovered molecules had an 80% to 90% success rate in Phase I trials, against roughly 40% in Phase II, in an analysis of AI-native biotech pipelines by Boston Consulting Group authors (Drug Discovery Today, Jayatunga et al., Apr 2024).

In a phase 2a trial of 71 patients with idiopathic pulmonary fibrosis, the AI-discovered drug rentosertib raised forced vital capacity by 98.4 ml at the 60 mg dose, against a 20.3 ml fall on placebo (Nature Medicine, Xu et al., Jun 2025).
Treatment-emergent adverse events occurred in 83.3% of patients on the 60 mg dose and 70.6% on placebo (Nature Medicine, Xu et al., Jun 2025).
The FDA's drug centre has seen over 500 submissions with AI components from 2016 to 2023 (FDA CDER, Oct 2026).
Isomorphic Labs alone accounted for about 15% of all biopharma investment so far in 2026 (Silicon Valley Bank, Aug 2026).
AI-driven protein research publications grew about 71% between 2024 and 2025 (Stanford HAI, Apr 2026).

Phase I mostly tests safety, so a high pass rate shows the molecules are drug-like without showing that they work. One mid-stage result in a small trial is the only efficacy signal in this list.

How safe, fair and regulated is clinical AI?

Regulators are writing rules for tools that have rarely been tested on real patients.

Barriers to and enablers of AI adoption in health care reported by WHO European Region countries, 2024 to 2025 — Figure 6: Legal uncertainty, not technology, is the most cited barrier to adopting AI in European health systems, and nearly all countries say clear liability rules would help. Source: WHO Regional Office for Europe, Artificial intelligence is reshaping health systems: state of readiness across the WHO European Region, Nov 2025. Chart published by the source, captured 14 Oct 2026.

How often does clinical AI fail?

ECRI ranked the misuse of AI chatbots in healthcare first on its 2026 list of health technology hazards, noting that the tools are not regulated as medical devices or validated for healthcare (ECRI, Jan 2026).

Across 31 language models, severe harm occurred in up to 22.2% of cases, and 76.6% of errors were omissions such as failing to recommend a critical test (NOHARM benchmark, Wu et al., Dec 2025).
The best models outperformed generalist physicians on safety by a mean 9.7% in that benchmark (NOHARM benchmark, Wu et al., Dec 2025).
LLM-generated clinical notes had a 1.47% hallucination rate and a 3.45% omission rate across 12,999 clinician-annotated sentences (npj Digital Medicine, Asgari et al., May 2025).
The Epic Sepsis Model missed 67% of patients with sepsis while alerting on 18% of all hospitalised patients, with an area under the curve of 0.63 (JAMA Internal Medicine, Wong et al., Jun 2021).
The deskilling study covered 1,443 colonoscopies at four Polish centres, and exposure to AI carried an odds ratio of 0.69 for detecting adenomas without it (Lancet Gastroenterology and Hepatology, Budzyń et al., Aug 2025).
A sepsis alert system monitored 590,736 patients at five hospitals, and when clinicians confirmed an alert within three hours, in-hospital mortality was 3.3 percentage points lower, an 18.7% relative reduction (Nature Medicine, Adams et al., Jul 2022). This compares confirmed with unconfirmed alerts, so it is an association.

The same field produces randomised-trial wins and a widely used sepsis model that missed two in three cases. Safety depends on the tool and on how it is checked after launch.

Is clinical AI biased?

A widely used US risk algorithm assigned Black patients the same risk score as sicker White patients; fixing it would raise the share of Black patients receiving extra help from 17.7% to 46.5% (Science, Obermeyer et al., Oct 2019).

The bias arose because the algorithm predicted health care costs rather than illness, and less money is spent on Black patients with the same needs (Science, Obermeyer et al., Oct 2019).
Across 21 open skin cancer image datasets holding 106,950 images, ethnicity was recorded for 1.3% of images and Fitzpatrick skin type for 2.1% (Lancet Digital Health, Wen et al., Nov 2021).
Only 15.8% of 519 language model studies assessed fairness, bias or toxicity, and 1.2% assessed calibration and uncertainty (JAMA, Bedi et al., Jan 2025).
In the ChatGPT Health stress test, patient race, sex and barriers to care showed no significant effects, though the confidence intervals did not rule out clinically meaningful differences (Nature Medicine, Ramaswamy et al., Feb 2026).
Among Americans who see racial bias in health care, 51% thought more AI would reduce it and 15% thought it would make it worse (Pew Research Center, Feb 2023).

Bias usually enters through the training data or the proxy chosen, so it rarely shows up unless someone measures it by group.

What rules govern clinical AI?

Few countries have written AI-specific health rules yet. Of 50 WHO European countries, 8% have developed liability standards for AI in health and 6% have legal requirements specific to generative AI in health (WHO Europe, Nov 2025).

86% named legal uncertainty as the top barrier to adoption, and 92% said clear liability rules would be the key enabler (WHO Europe, Nov 2025).
64% of those countries used AI-assisted diagnostics and 50% used chatbots for patient support, yet only 24% offered in-service AI training for health workers (WHO Europe, Nov 2025).
WHO's January 2024 guidance on large multi-modal models sets out over 40 recommendations for governments, developers and providers (WHO, Jan 2024).
US rule 45 CFR 92.210 requires covered entities to make reasonable efforts to identify and mitigate discrimination risk from patient care decision support tools (eCFR, HHS, Oct 2026).
The FDA's final guidance on predetermined change control plans for AI-enabled device software functions is dated August 2025 (FDA, Aug 2025).
The FDA issued updated clinical decision support software guidance in January 2026 (FDA, Jan 2026).
The FDA's Digital Health Advisory Committee met on 6 November 2025 to discuss generative AI-enabled digital mental health devices (FDA, Nov 2025).
California's AB 3030, chaptered on 28 September 2024, requires generative AI patient messages to carry a disclaimer and instructions for reaching a human, unless a licensed provider reviewed them (California Legislature, Sep 2024).
Illinois signed a law in August 2025 that prohibits using AI to provide therapy or make therapeutic decisions, while allowing administrative and supporting uses (Illinois IDFPR, Aug 2025).
The EU's MDCG 2025-6 guidance says a medical device that contains a high-risk AI system must satisfy the Medical Devices Regulation and the AI Act at the same time (European Commission, MDCG, Jun 2025).
Joint Commission and the Coalition for Health AI released guidance on responsible AI use in September 2025, and plan a voluntary AI certification open to more than 22,000 accredited organisations (Joint Commission, Sep 2025).

Rules are arriving in three layers: device law from regulators, anti-discrimination and disclosure law from states and civil rights agencies, and voluntary governance from accreditors.

Methodology and Sources

We included a number only after tracing it to the organisation that produced it and opening that page, or the repository copy of the same paper, to confirm the figure and its date. Roundups were followed to the original study or dropped. Government and regulator data, peer-reviewed papers and statistical releases make up most of the page. Surveys and reports from vendors and commercial publishers (Doximity, Elsevier, Wolters Kluwer, OpenAI, Menlo Ventures, Microsoft and the market-research firms) are named and labelled so you can weigh them. FDA device counts come from the rows of the FDA's downloadable list as of 29 June 2026 and are marked as counted. The page is refreshed monthly.

Sources

Glossary

AI scribe (ambient AI, ambient voice technology): software that records a visit and drafts the clinical note for the clinician to review.
510(k), De Novo, PMA: the three FDA routes to market for devices. 510(k) shows equivalence to an earlier device, De Novo covers new low-to-moderate-risk devices and PMA covers high-risk devices.
PCCP: predetermined change control plan, an FDA-agreed plan for updating an AI device after authorisation.
EHR: electronic health record, the software hospitals use to store patient records.
Predictive AI: models that produce risk scores or forecasts, such as readmission or sepsis risk.
Generative AI / LLM: systems that write text, summaries or images, and large language models that power chatbots and scribes.
ASTP/ONC: the US Assistant Secretary for Technology Policy and Office of the National Coordinator for Health IT.
RVU: relative value unit, a measure of physician work used in Medicare payment.
Interval cancer: a breast cancer found between two screening rounds that screening did not detect.
Sensitivity and specificity: the share of real cases a test catches, and the share of healthy people it correctly clears.
Undertriage: sending a patient to a lower level of care than their condition needs.
Deskilling: loss of a clinician's own skill after routine reliance on a tool.
MDCG: the EU Medical Device Coordination Group.

158 AI in Healthcare Statistics October 2026