133 AI in Healthcare Statistics October 2026

AI in healthcare statistics from 2026 show physicians use AI mostly for documentation, not diagnosis. Among US physicians, 30% now use AI to create discharge instructions, care plans or progress notes, 28% to document billing codes, charts or visit notes and 28% to generate chart summaries, up from 20%, 21% and 12% in 2024 (AMA, 2026 Physician Survey on Augmented Intelligence, Mar 2026). Assistive diagnosis, the use that gets the headlines, sits lower at 17%, behind drafting replies to patient portal messages at 19% and translation at 18% (AMA, 2026 Physician Survey on Augmented Intelligence, Mar 2026).

Spending follows the same pattern. Ambient clinical documentation is the largest category of healthcare AI spending, an estimated $600M in 2025, ahead of coding and billing automation at $450M (Menlo Ventures, 2025 State of AI in Healthcare, Oct 2025). Diagnostic AI is regulated as medical devices: the FDA's list is mostly radiology software, and the strongest trial evidence comes from breast screening. Most patients want to be told when AI is involved in their care, and most report low trust that their health system will use it responsibly.

The figures below cover physician and hospital adoption, ambient AI scribes, FDA AI-enabled devices, accuracy and safety, patient attitudes, and money. Non-US figures are labeled.

How many doctors and hospitals use AI in 2026?

Physician AI use is past the early-adopter stage. Most US doctors now use it, and they mostly use it to read and write rather than to diagnose. Hospitals got there first, through predictive models built into the EHR.

How common is physician AI use in 2026?

Most US physicians now use AI at work, and each one uses it for more tasks than in 2023. Adoption outside the US is lower and uneven, with UK GPs well behind.

• The average physician now uses 2.3 AI use cases in practice, up from 1.1 in 2023 (AMA, 2026 Physician Survey on Augmented Intelligence, Mar 2026).
• 76% of physicians say AI gives them an advantage in caring for patients, up from 65% in 2023, while 40% remain equally excited and concerned about it (AMA, 2026 Physician Survey on Augmented Intelligence, Mar 2026).
• 85% of physicians want to be consulted on or responsible for bringing AI into their practice, and 92% want more AI education and training (AMA, 2026 Physician Survey on Augmented Intelligence, Mar 2026).
• Worldwide, 49% of clinicians and 57% of doctors used an AI tool for work in a survey of 2,757 clinicians in 118 countries, fielded December 2025 to February 2026 (Elsevier, Clinician of the Future 2026, May 2026).
• Among clinicians who use AI at work, 34% now use clinical-specific AI tools frequently or always, up from 22% in 2025, with North America highest at 41% (Elsevier, Clinician of the Future 2026, May 2026).
• In the UK, 25% of GPs used generative AI tools in clinical practice in January 2025, up from 20% a year earlier (Blease et al., Digital Health, Nov 2025).
• 95% of UK GPs said they had no professional training in using generative AI at work, and 85% said their employer had not encouraged them to use it (Blease et al., Digital Health, Nov 2025).

A single-hospital survey shows how much of this use is informal.

• At one large US academic hospital, 66.7% of 54 surveyed hospitalists used AI in clinical practice without any institutional rollout, and OpenEvidence (51.9%) was used far more than ChatGPT (7.4%) (Bagla et al., JMIR, Mar 2026).

Which tasks do physicians use AI for?

Physicians mostly use AI to read and write: summarizing research, drafting notes and coding visits. Diagnosis ranks low in the national surveys, although hospitalists who already use AI lean on it for differential diagnoses.

• Summaries of medical research and standards of care are now the most common AI use case, used by 39% of US physicians in 2026 compared with 13% in 2024 (AMA, 2026 Physician Survey on Augmented Intelligence, Mar 2026).
• Doctors who use general-purpose AI tools such as ChatGPT or Copilot mostly use them to query medical literature (61%), look up drug information (47%) and for professional education (47%), and far less for patient communications (31%) or medical image analysis (18%) (Elsevier, Clinician of the Future 2026, May 2026).
• Among UK GPs who use generative AI, 35% use it to write documentation after appointments and 27% to suggest a differential diagnosis (Blease et al., Digital Health, Nov 2025).
• 71% of UK GPs who use generative AI for clinical tasks said the tools reduced their work burden (Blease et al., Digital Health, Nov 2025).
• Among the hospitalists at one US academic hospital who use AI, 88.9% use it to answer clinical questions and 86.1% to generate differential diagnoses, but only 44.4% to make patient education materials (Bagla et al., JMIR, Mar 2026).

How many hospitals and health systems have adopted AI?

Most US hospitals already run predictive AI inside their EHR, and large health systems have deployed imaging AI almost everywhere. Small, rural, critical access and independent hospitals trail by a wide margin.

• Billing automation was the fastest-growing hospital AI use, rising from 36% to 61% of hospitals using predictive AI between 2023 and 2024, while scheduling rose from 51% to 67% (ASTP/ONC Data Brief 80, Sep 2025).
• In 2024, 82% of hospitals using predictive AI evaluated it for accuracy, 74% for bias and 79% monitored it after implementation (ASTP/ONC Data Brief 80, Sep 2025).
• Only 37% of independent hospitals used predictive AI in 2024, against 86% of multi-hospital system members; the gap was similar for critical access hospitals (50% vs 80% for other hospitals), small hospitals (59% vs 96% for large) and rural hospitals (56% vs 81% urban) (ASTP/ONC Data Brief 80, Sep 2025).
• Imaging and radiology is the most widely deployed clinical AI in large US health systems: 90% of 43 systems surveyed in fall 2024 had it deployed in at least some areas (Poon et al., JAMIA, May 2025).
• Only 38% of health systems using AI for clinical risk stratification, such as early sepsis detection, reported a high degree of success with it (Poon et al., JAMIA, May 2025).
• 77% of health systems named immature AI tools as their biggest or second-biggest barrier, ahead of financial concerns (47%) and regulatory uncertainty (40%) (Poon et al., JAMIA, May 2025).

Census data on health care and social assistance businesses shows AI use rising through 2026.

• 25.5% of US health care and social assistance businesses used AI in a business function in the two weeks before the 7 to 20 Sep 2026 survey, up from 21.0% in the wave collected 29 Dec 2025 to 11 Jan 2026 (U.S. Census Bureau, Business Trends and Outlook Survey, Sep 2026). The rate across all US businesses in the same wave was 23.8% (U.S. Census Bureau, Business Trends and Outlook Survey, national data, Sep 2026).

How fast is ambient AI scribe adoption growing?

By the reading of the JAMIA survey authors, ambient scribes moved past early adopters faster than imaging AI did. The first rigorous studies say the time they save is real but small, and largest in primary care and for clinicians who use the scribe on most visits.

How many clinicians use ambient AI scribes?

Few large US health systems have switched ambient notes on everywhere, and clinicians who have the tool use it on a minority of visits.

• Only 14% of 43 large US health systems had deployed ambient notes fully in fall 2024, while 47% had deployed it in limited areas and 40% were still developing or piloting it (Poon et al., JAMIA, May 2025).
• Microsoft said its healthcare AI automated 28 million patient encounters in the April to June 2026 quarter, double the year before, and was on pace for more than 100 million in calendar 2026 (Microsoft, FY26 Q4 earnings call, Jul 2026).
• In the UK, 14% of 1,003 GPs surveyed in August 2025 used an ambient AI scribe, 39% planned to adopt one soon and 46% had no plans to (Blease et al., BMJ Health & Care Informatics, Jul 2026).

Two academic health systems show how uptake builds after launch.

• By 31 March 2025, about 11 weeks after Vanderbilt University Medical Center made ambient scribing available to more than 2,400 clinicians in a single day, 1,223 clinicians had used it and 20.1% of visit notes included it (Wright et al., JAMIA, Oct 2025).
• At UCSF Health, 698 of 1,565 attending physicians (44.6%) had adopted an AI scribe by April 2025, yet encounters with adopters made up only 15.2% of 1.2 million ambulatory visits (Holmgren et al., JAMA Network Open, Jan 2026).

How much documentation time do AI scribes save?

The savings are real but small. The largest multisite study and two randomized trials find modest gains, biggest in primary care and among heavy users, and after-hours EHR time barely moves.

• Primary care clinicians who adopted an AI scribe spent 26.9 fewer minutes on documentation and 25.0 fewer minutes in the EHR per 8 scheduled patient hours than primary care nonadopters (Rotenstein et al., JAMA, Apr 2026).
• Clinicians who used the scribe in 50% or more of visits spent 27.3 fewer minutes on documentation per 8 scheduled patient hours and delivered 1.0 more visit per week than nonadopters (Rotenstein et al., JAMA, Apr 2026).
• After-hours EHR time did not change significantly with AI scribe adoption (3.1 fewer minutes per 8 scheduled patient hours, 95% CI 6.8 fewer to 0.5 more), even as weekly visits rose by 0.49 (Rotenstein et al., JAMA, Apr 2026).
• In a 24-week stepped-wedge randomized trial of 66 practitioners at University of Wisconsin-led clinics in two states, ambient AI cut time on notes by 0.36 hours a day; a 0.50-hour drop in work outside work lost significance once the top 3% of days were excluded (Afshar et al., NEJM AI, Nov 2025).
• UCSF physicians who adopted an AI scribe billed 1.81 more relative value units (RVUs) per week, worth about $3,044 a year at 2025 Medicare rates, and saw 0.80 more patients a week, with no change in claim denials (Holmgren et al., JAMA Network Open, Jan 2026).
• Across 14 emergency departments, clinicians spent no less active editing time on AI-assisted notes than on conventional ones (median difference 0.17 minutes more), though they typed 722 fewer characters per note (Kashiouris et al., Applied Clinical Informatics, Sep 2026).
• 46% of US medical practice leaders said new AI tools had made their providers more productive over the past two years, 27% saw no gain, and those reporting gains named ambient AI scribes as the main driver (MGMA Stat, May 2026).

Dictating the note after the visit is the older route to a shorter note. Speaking and typing speeds are compared in the typing speed statistics hub, and clinical dictation software covers that workflow.

Do AI scribes ease burnout and note quality worries?

Trial and survey scores improve with ambient scribes, but the burnout evidence is still rated low certainty and comes mostly from volunteer early adopters. Clinicians still report errors in scribe notes, and patients grow less willing to consent the more they are told.

• Physicians given either AI scribe in the UCLA trial gained 2.83 and 2.69 points on the 10-to-50 Mini-Z well-being scale, and DAX users cut task load by 39.9 points on a 0-to-400 scale (Nabla's 31.7-point drop was not statistically significant), secondary results the authors say need confirmation (Lukac et al., NEJM AI, Nov 2025).
• Ambient AI lowered work exhaustion and interpersonal disengagement by 0.44 points on a five-point scale in a randomized trial, but the 0.14-point rise in professional fulfillment was not significant (Afshar et al., NEJM AI, Nov 2025).
• Among 263 ambulatory clinicians at 6 US health systems, the share reporting burnout fell from 51.9% before to 38.8% after 30 days with an ambient AI scribe, in a pre-post study without a control group (Olson et al., JAMA Network Open, Oct 2025).
• A meta-analysis of 21 studies found ambient AI documentation was associated with lower odds of burnout (odds ratio 0.47, from 3 studies), but rated the certainty of that burnout evidence as low (Gong et al., JMIR, Aug 2026).

How common burnout is in the first place is covered in the physician burnout statistics hub. On note quality, the scribes are not error-free.

• Physicians in the UCLA trial rated clinically significant inaccuracies in scribe notes as occurring occasionally (2.7 and 2.8 on a five-point scale), and one mild adverse event was reported (Lukac et al., NEJM AI, Nov 2025).
• 32% of UK GPs who use ambient scribes said errors happen often or always, and 14% reported errors with significant-to-critical implications (Blease et al., BMJ Health & Care Informatics, Jul 2026).
• 81.6% of patients at one US academic health center consented to ambient documentation when given basic information, but only 55.3% did once AI features, data storage and corporate involvement were explained (Lawrence et al., JAMA Network Open, Jul 2025).

How many FDA AI-enabled devices are authorized?

The FDA list grows faster every year, and it is mostly a list of software that reads medical images. Nearly all of those devices reached the market through 510(k) clearance, which rests on comparison with an existing device rather than new clinical trials, and the public evidence behind most of them is thin.

How many AI-enabled medical devices has the FDA authorized?

2025 set a record for FDA authorizations of AI-enabled devices, and the first half of 2026 is already ahead of the same months of 2025. Almost every device came through the 510(k) pathway.

• The FDA authorized 335 AI-enabled devices in 2025, up 43% from 235 in 2024 and 226 in 2023 (FDA, AI-Enabled Medical Device List, Sep 2026).
• The list already holds 181 devices with decisions from January to June 2026, against 176 for the same months of 2025, and the 2026 figure will rise as FDA adds devices whose summaries are still unpublished (FDA, AI-Enabled Medical Device List, Sep 2026).
• 96.2% of AI-enabled devices on the list, 1,553 in all, reached the market through 510(k) clearance; 40 came through De Novo and 21 through premarket approval (PMA) (FDA, AI-Enabled Medical Device List, Sep 2026).
• The FDA list does not yet flag which devices use large language models; the agency says it will tag LLM-based functionality in a future update (FDA, AI-Enabled Medical Device List, Sep 2026).

The yearly counts come from FDA's own downloadable file, counted by decision date. FDA adds devices in later updates, so recent years tend to grow after the fact.

Which medical specialties have the most AI devices?

Radiology, by a wide margin. Every other review panel is in single digits, and even in dentistry nearly half of AI devices are for oral radiology.

• Cardiovascular is a distant second on the FDA list with 154 devices (9.5%) and neurology third with 73 (4.5%), while cardiovascular authorizations doubled from 15 in 2024 to 30 in 2025 (FDA, AI-Enabled Medical Device List, Sep 2026).
• Only 30 of 1,358 FDA-cleared AI devices (2.2%) were specific to urology as of September 2025, and 26.7% of those were cleared in 2025 alone (Qian et al., BJU International, Jun 2026).
• Researchers found 52 AI or machine-learning dental devices in FDA's 510(k) database, and 48% of them were for oral radiology (Naved et al., BMC Oral Health, Nov 2025).
• 60% of FDA-cleared AI dental devices did not disclose what type of AI algorithm they use, and 50% had no public record of clinical deployment (Naved et al., BMC Oral Health, Nov 2025).
• All 9 AI breast cancer screening products cleared or approved by the FDA from 2017 to 2021 were authorized on previously collected, retrospective data, and none reported interval cancer detection or cancer stage at detection (Potnis et al., JAMA Internal Medicine, Dec 2022).

How are AI medical devices tested and monitored after clearance?

Most public FDA summaries for AI devices give no demographic data, few report prospective or randomized evidence, and devices without clinical validation are recalled most often.

• 95.5% of FDA-cleared AI and machine-learning devices did not report demographic information in their public summaries, and 46.7% did not report their study design (Lin et al., JAMA Health Forum, Sep 2025).
• Only 6 of 691 FDA-cleared AI devices reported data from randomized clinical trials, and 53 (7.7%) reported data from prospective studies (Lin et al., JAMA Health Forum, Sep 2025).
• 489 adverse events were reported for 36 FDA-cleared AI devices, including 458 malfunctions, 30 injuries and 1 death (Lin et al., JAMA Health Forum, Sep 2025).
• Only 3.6% of FDA approvals for AI devices from 1995 to 2023 reported the race or ethnicity of study subjects, and 81.6% did not report their age (Muralidharan et al., npj Digital Medicine, Oct 2024).
• FDA summaries for 1,012 AI devices scored an average of 3.3 out of 17 on a transparency scale for model development and performance, improving by only 0.88 points after FDA's 2021 machine-learning principles (Mehta et al., npj Digital Medicine, Nov 2025).

Recalls follow the evidence gap. Two separate research teams found the same pattern in different snapshots of the list.

• 60 of 950 FDA-cleared AI devices (6.3%) were tied to 182 recall events; devices with no reported clinical validation had 2.8 times the odds of a recall, and publicly traded companies made 53.2% of the devices but accounted for 91.8% of recalls (Lee et al., JAMA Health Forum, Aug 2025).
• Of 903 FDA-authorized AI devices tracked to August 2024, 43 (4.8%) were recalled, after a median of 458 days on the market (Ren et al., JAMA Network Open, Jun 2026).
• 7.8% of AI devices without a clinical performance study were recalled, against 2.6% of devices with one (Ren et al., JAMA Network Open, Jun 2026).

How accurate and safe is AI in medicine?

The strongest evidence for AI in medicine comes from breast screening, where a Swedish randomized trial found AI-supported reading caught more cancers with far fewer radiologist reads. Chatbots have less trial evidence and less regulation than imaging AI.

Does AI match doctors at diagnosis?

In breast screening, the Swedish MASAI trial found AI-supported reading non-inferior to two radiologists on interval cancers. Chatbot results are mixed, with GPT-4 failing to improve doctors' diagnoses in one trial and lifting their management scores in another.

• In MASAI, AI-supported mammography screening had 1.55 interval cancers per 1,000 women vs 1.76 with standard double reading, meeting the trial's non-inferiority goal (Gommers et al., Lancet, Jan 2026).
• AI-supported screening detected 6.4 cancers per 1,000 Swedish women vs 5.0 with standard double reading, a 29% higher detection rate, without a significant rise in false positives (Hernström et al., Lancet Digital Health, Feb 2025).
• Radiologists in the AI arm of MASAI did 61,248 screen readings vs 109,692 in the standard arm, a 44.2% lower reading workload (Hernström et al., Lancet Digital Health, Feb 2025).
• At MASAI's first safety analysis of 80,033 Swedish women, 28.3% of recalls in the AI-supported arm turned out to be cancer vs 24.8% with standard reading (Lång et al., Lancet Oncology, Aug 2023).
• Across 14 studies and 1,214,885 screening exams, standalone AI mammography reached a pooled accuracy (AUC) of 0.890, but heterogeneity was so high that the 95% prediction interval for a new program ran from 0.731 to 0.960, and the authors say local validation before deployment remains necessary (Ciurescu et al., Diagnostics, Sep 2026).

Chatbot trials measure something different: whether a general-purpose model helps a physician reason through a case.

• In a randomized trial of 50 US physicians, those given GPT-4 scored a median 76% on diagnostic reasoning vs 74% for those using conventional resources, a difference that was not statistically significant (Goh et al., JAMA Network Open, Oct 2024).
• In the same trial, GPT-4 working alone scored 16 percentage points higher than physicians using conventional resources (Goh et al., JAMA Network Open, Oct 2024).
• On management decisions such as treatment and testing, 92 physicians using GPT-4 scored 6.5 percentage points higher than those using conventional resources, but spent about 119 seconds longer per case (Goh et al., Nature Medicine, Feb 2025).

How often do AI tools make errors or show bias?

Errors are common enough that ECRI now asks providers to report them. Physician reviewers keep finding omissions and invented details in AI drafts, and a small share of drafts could harm a patient if sent unedited.

• In an ECRI survey of 124 mostly quality, safety and risk leaders, 31% said they had seen an AI output they believed was incorrect or misleading in the past year, and 9% said an AI error had reached a patient or affected a care decision (ECRI, AI error survey, Aug 2026).
• In 450 LLM-generated clinical notes, 1.47% of 12,999 note sentences contained hallucinated information, and 3.45% of 49,590 consultation transcript sentences were left out of the notes (Asgari et al., npj Digital Medicine, May 2025).
• When physicians reviewed GPT-4 drafts of replies to patient messages, they judged 7.1% of drafts could cause severe harm if sent unedited, and 1 of 156 (0.6%) could cause death (Chen et al., Lancet Digital Health, Apr 2024).
• Physicians who reviewed 100 AI-drafted hospital course summaries found omissions in 25%, inaccuracies in 20% and hallucinations in 2%, and rated 88% as having no potential for harm if left unedited (Grolleau et al., JAMA Network Open, May 2026).
• Tested alone, LLMs named the right condition in 94.9% of medical scenarios, but UK adults using the same LLMs named it in fewer than 34.5% of cases, no better than people using other sources (Bean et al., Nature Medicine, Feb 2026).

Bias shows up in both older algorithms and LLMs.

• Across 1.7 million outputs from nine LLMs on the same emergency cases, some cases labeled LGBTQIA+ were sent for mental health assessments about six to seven times more often than clinically indicated (Omar et al., Nature Medicine, Apr 2025).
• A 2019 study of a widely used commercial care-management algorithm found that correcting its racial bias would raise the share of Black patients flagged for extra help from 17.7% to 46.5% (Obermeyer et al., Science, Oct 2019).

What rules govern AI in medicine?

The US relies on agency rules and state bills rather than one AI law, and the EU, which does have one, has pushed its medical-device deadline back a year.

• By our count of the FY2025 and FY2024 inventory files HHS publishes, HHS listed 446 AI use cases in fiscal 2025, up from 271 in fiscal 2024; 167 of the 446 were deployed and 88 were in pilots (HHS, AI Use Case Inventory, Jul 2026).
• ONC's HTI-1 final rule, published in January 2024, replaced the old clinical decision support certification criterion with a decision support intervention criterion that covers predictive AI, and only the new criterion counts toward the Base EHR definition from January 1, 2025 (ONC, HTI-1 Final Rule, Federal Register, Jan 2024).
• By our count of NCSL's bill table, state legislators introduced 152 AI bills tagged as health use in 36 states during the 2025 session, and 20 of them were enacted or adopted, in 16 states (NCSL, Artificial Intelligence 2025 Legislation, Jul 2025).
• The EU amended its AI Act in July 2026 so that high-risk rules for AI built into medical devices and in vitro diagnostics now apply from August 2, 2028, a year later than the August 2, 2027 date in the original 2024 text (EUR-Lex, AI Act consolidated text, Jul 2026).
• WHO's guidance on large multi-modal models, the generative AI behind health chatbots, sets out more than 40 recommendations for governments, technology companies and health care providers (WHO, Jan 2024).

What do patients think about AI in their care?

Patients want disclosure and a say in how AI is used in their care, and a growing share put health questions to chatbots on their own.

Are patients comfortable with doctors using AI?

In Pew's 2023 survey, most Americans were not. Comfort rises when a clinician stays in charge, and patients who have had an ambient AI scribe in the room mostly found it helpful.

• 60% of US adults said they would be uncomfortable if their own health care provider relied on AI to diagnose disease and recommend treatments, against 39% who would be comfortable, in Pew's latest reading on the question (Pew Research Center, Feb 2023).
• 75% of Americans said their bigger worry is that providers will adopt AI too fast, before the risks to patients are understood; 23% worried more that providers would move too slowly (Pew Research Center, Feb 2023).
• Among 13,806 hospital patients surveyed in 43 countries, 57.6% held a positive view of AI in health care, and patients in poorer health were markedly less positive (Busch et al., JAMA Network Open, Jun 2025).
• 72.9% of the same patients preferred physician-led decision-making and 70.2% preferred AI that can explain itself, even at the cost of slightly lower accuracy (Busch et al., JAMA Network Open, Jun 2025).
• In a conjoint experiment with 3,000 US adults, having a clinician present raised the probability of choosing an AI-assisted visit by 18.4 percentage points; AI performance above specialist level raised it by 32.5 points (Bracic et al., JAMA Network Open, Mar 2026).

Direct experience tends to soften the worry, though disclosure still matters.

• In a survey of 1,455 Duke Health patients, respondents rated AI-drafted portal replies higher than clinician-written ones (0.30 points on a 5-point satisfaction scale), but satisfaction was 0.13 points lower when the reply was labeled as AI-written than when it was labeled as written by a clinician (Cavalier et al., JAMA Network Open, Mar 2025).
• Of 2,202 Stanford Health Care patients whose visit used an ambient AI scribe, 70.1% found it helpful and 73.6% wanted it used at future visits (Shah et al., JAMIA Open, Jun 2026).

Do patients trust health systems to use AI responsibly?

Low trust is the majority view in the main US survey on the question, and many patients cannot say whether AI has been used in their care at all.

• 65.8% of US adults surveyed in mid-2023 reported low trust in their health care system to use AI responsibly (Nong and Platt, JAMA Network Open, Feb 2025).
• 57.7% reported low trust that their health system would make sure an AI tool would not harm them (Nong and Platt, JAMA Network Open, Feb 2025).
• Adults who had experienced discrimination while seeking care had lower odds of trusting their system to use AI responsibly (odds ratio 0.66) and to protect them from AI harm (odds ratio 0.57); health literacy and AI knowledge showed no association with trust (Nong and Platt, JAMA Network Open, Feb 2025).
• Only 41.8% of hospital patients in the 43-country survey trusted AI to give accurate information on how they would respond to treatment, the lowest of the trust items asked (Busch et al., JAMA Network Open, Jun 2025).
• 63% of US adults want more input on whether AI is used in their health care, three times the 21% who are comfortable with how much input they have (Pew Research Center, Aug 2026).
• 46% of US adults are not sure whether AI has been used in their health care, 37% say it has not and 16% say it has (Pew Research Center, Aug 2026).
• 77% of US adults are very or somewhat concerned about the privacy of personal medical information given to AI tools (KFF Tracking Poll on Health Information and Trust, Mar 2026).

How many people ask AI chatbots health questions?

Use is rising in both surveys that track it over time, Rock Health's and KFF's, and physicians say few patients tell them.

• 34% of US adults use AI chatbots for at least one of eight health reasons, and 25% use them to figure out what is causing their symptoms (Pew Research Center, Aug 2026).
• 32% of US adults have ever turned to an AI chatbot for health information, double the 16% of a year earlier (Rock Health, 2025 Consumer Adoption Survey, Mar 2026).
• 64% of those health AI users ask a chatbot health questions weekly or more often (Rock Health, 2025 Consumer Adoption Survey, Mar 2026).
• 41% of adults who used AI for health in the past year have uploaded personal medical information such as test results or doctor's notes, equal to 13% of all US adults (KFF Tracking Poll on Health Information and Trust, Mar 2026).
• More than 300 million people worldwide put health-related questions to ChatGPT every week, by OpenAI's own count of its users (OpenAI, Jul 2026).
• Only 8% of US physicians say most of their patients disclose using AI, while 30% believe most of their patients are probably using it (AMA, 2026 Physician Survey on Augmented Intelligence, Mar 2026).

How much money is flowing into healthcare AI?

Health systems supply most of the money spent on healthcare AI, and documentation takes the largest share. Market-size forecasts differ by about 40% even for 2026, so read each one as a single firm's estimate.

How much venture funding goes to healthcare AI startups?

Rounds got bigger and fewer in 2025, a handful of mega deals carried the year, and funding kept rising in the first half of 2026.

• US digital health startups raised $14.2B in 2025, 35% more than the $10.5B raised in 2024 and the highest annual total since 2022 (Rock Health, 2025 Year-End Digital Health Funding Overview, Jan 2026).
• AI-enabled digital health startups raised rounds about 19% larger on average than startups that do not center AI in their products, and the gap reached 61% at Series C (Rock Health, 2025 Year-End Digital Health Funding Overview, Jan 2026).
• Fewer startups got more money in 2025: deal count fell 5% to 482 while the average deal grew to $29.3M from $20.7M (Rock Health, 2025 Year-End Digital Health Funding Overview, Jan 2026).
• Mega deals of $100M or more took 42% of all 2025 digital health funding, nearly double the previous year's share, and removing the top nine companies by dollars raised would drop the 2025 total below 2024 (Rock Health, 2025 Year-End Digital Health Funding Overview, Jan 2026).
• US digital health startups raised $7.4B across 244 deals in the first half of 2026, $1B more than in the first half of 2025 on about the same number of deals (Rock Health, H1 2026 Funding and Market Overview, Jul 2026).
• In the first half of 2026, 20 mega deals, just over 8% of all deals, absorbed 45% of digital health venture capital; Rock Health stopped labeling startups as AI-enabled from Q1 2026 because it says AI is now too common to count as a differentiator (Rock Health, H1 2026 Funding and Market Overview, Jul 2026).

Large rounds went to clinical AI and ambient documentation companies. Each figure below comes from the company's own announcement of its own round.

• Clinical AI company Aidoc raised a $150M Series E in April 2026, less than a year after its previous growth round, taking its total funding past $500M (Aidoc, Apr 2026).
• Ambient documentation company Abridge raised a $250M Series D in February 2025 (Abridge, Feb 2025).

How much are health systems spending on AI?

Spending is rising fast by the one investor survey that measures it, and health systems pay for most of it. Buying cycles are getting shorter for providers and longer for payers.

• US healthcare organizations spent an estimated $1.4B on generative AI in 2025, nearly triple their 2024 spend (Menlo Ventures, 2025 State of AI in Healthcare, Oct 2025).
• Health systems supplied about $1B of that spend (75%), outpatient providers $280M (20%) and payers just $50M (5%) (Menlo Ventures, 2025 State of AI in Healthcare, Oct 2025).
• 22% of US healthcare organizations have implemented domain-specific AI tools, 7 times the 2024 rate and 10 times the 2023 rate, led by health systems at 27%, outpatient providers at 18% and payers at 14% (Menlo Ventures, 2025 State of AI in Healthcare, Oct 2025).
• 85% of generative AI spending in healthcare goes to startups rather than incumbent vendors (Menlo Ventures, 2025 State of AI in Healthcare, Oct 2025).
• Health systems now take 6.6 months on average to buy AI, down from 8.0 months for traditional IT purchases, and outpatient providers take 4.7 months, down from 6.0 (Menlo Ventures, 2025 State of AI in Healthcare, Oct 2025).
• Payers move the other way: their average AI buying cycle has lengthened to 11.3 months from 9.4 months (Menlo Ventures, 2025 State of AI in Healthcare, Oct 2025).

Contract ceilings and vendor revenue give a second read on the money.

• The US Department of Veterans Affairs ambient AI enterprise contract, a multiple-award contract shared by all eligible vendors, carries a total ceiling of $775.72M over five years, according to awardee Abridge (Abridge, Sep 2026).
• Tempus AI, which sells AI-driven diagnostics and data, reported second-quarter 2026 revenue of $382.5M, up 22% year over year, with its Data and Applications business up 28% to $93.2M (Tempus AI, Q2 2026 results (SEC Form 8-K), Jul 2026).

How big will the healthcare AI market get?

It depends on whose model you read. Three research firms disagree by about 40% even on the 2026 figure and use different end years, so no single number is the market size.

• MarketsandMarkets values the global AI in healthcare market at $25.88B in 2025 and projects growth from $36.67B in 2026 to $194.79B by 2031, a 39.7% annual growth rate (MarketsandMarkets, Jun 2026).
• Grand View Research values the market at $36.7B in 2025 and projects $50.7B in 2026 rising to $505.6B by 2033, a 38.9% annual growth rate, with North America holding 54% of 2025 revenue (Grand View Research, Jun 2026).
• Precedence Research puts the market at $36.96B in 2025 and $51.20B in 2026, and projects $744.34B by 2035, a 35.02% annual growth rate (Precedence Research, Sep 2026).

All three are paid reports, and the figures above come from each firm's own public summary page. They define the market differently, so their end points cannot be compared as if they were the same year.

Glossary

• AI: artificial intelligence, software that performs tasks such as reading images, predicting risk or writing text.
• LLM: large language model, the kind of AI behind chatbots such as ChatGPT.
• Generative AI: AI that produces new text, images or audio rather than only classifying data.
• Ambient AI scribe: a tool that listens to a clinical visit, with consent, and drafts the note for the clinician to review.
• EHR: electronic health record.
• Time-in-note: the time a clinician spends writing and editing notes in the EHR, as logged by the EHR.
• RVU: relative value unit, the measure Medicare uses to price physician work.
• Mini-Z: a short survey that scores physician well-being and burnout.
• 510(k): the FDA clearance pathway for devices shown to be substantially equivalent to a device already on the market.
• De Novo: the FDA pathway for new low-to-moderate-risk devices with no existing equivalent.
• PMA: premarket approval, the FDA pathway for high-risk devices, which requires clinical evidence of safety and effectiveness.
• ASTP/ONC: the Assistant Secretary for Technology Policy and Office of the National Coordinator for Health IT, part of HHS.
• HHS: the US Department of Health and Human Services.
• HTI-1: the 2024 ONC rule on health IT certification that set transparency requirements for decision support tools, including predictive AI.
• Sensitivity: the share of people with a disease that a test correctly flags.
• Specificity: the share of people without a disease that a test correctly clears.
• Interval cancer: a cancer diagnosed between scheduled screenings, after a screen that came back normal.
• AUC: area under the curve, a 0-to-1 score of how well a test separates positive from negative cases.

Methodology and Sources

Every statistic on this page was traced to the organization that produced the data: government agencies and regulators (FDA, HHS, ASTP/ONC, the Census Bureau, EUR-Lex), peer-reviewed studies in journals such as JAMA, the Lancet, NEJM AI and Nature Medicine, professional and safety bodies (AMA, ECRI, MGMA, WHO, NCSL), survey researchers (Pew Research Center, KFF), and companies reporting their own data, funding rounds or filings. Each link was opened on 2 October 2026 to confirm the page loads, the number appears on it and the date is right; counts from the FDA, HHS and NCSL files are our own tallies of the published data. Observational studies are reported as associations, trial results keep their comparators, market forecasts are attributed to the firm that made them, and non-US data is labeled. We do not cite statistics roundups. The page is refreshed monthly, starting with the FDA list, physician surveys, venture funding and the Census business survey.

Sources

• Abridge: Series D funding announcement
• Abridge: VA ambient AI enterprise contract
• Afshar et al., NEJM AI: Pragmatic randomized trial of ambient AI documentation and practitioner well-being
• Aidoc: Aidoc raises $150 million Series E
• AMA: 2026 Physician Survey on Augmented Intelligence
• Asgari et al., npj Digital Medicine: Framework for assessing clinical safety and hallucination rates of LLMs for medical text summarisation
• ASTP/ONC Data Brief 80: Hospital trends in the use, evaluation and governance of predictive AI, 2023 to 2024
• Bagla et al., JMIR: Patterns of AI use by hospitalists
• Bean et al., Nature Medicine: LLMs and members of the public in medical scenarios
• Blease et al., BMJ Health & Care Informatics: UK GPs and ambient AI scribes
• Blease et al., Digital Health: UK GPs and generative AI, 2025 survey
• Bracic et al., JAMA Network Open: Patient preferences for AI-assisted diagnosis
• Busch et al., JAMA Network Open: Multinational attitudes of hospital patients toward AI
• Cavalier et al., JAMA Network Open: Patient perceptions of AI-drafted portal message replies
• Chen et al., Lancet Digital Health: Effect of using an LLM to respond to patient messages
• Ciurescu et al., Diagnostics: Meta-analysis of AI in mammography screening
• Doximity: 2026 State of AI in Medicine
• ECRI: Misuse of AI chatbots tops annual list of health technology hazards
• ECRI: ECRI expands problem reporting network and calls for more data on AI errors in patient care
• Elsevier: Clinician of the Future 2026, full report
• EUR-Lex: AI Act, Regulation (EU) 2024/1689, consolidated text of 27 July 2026
• FDA: Artificial Intelligence-Enabled Medical Device List
• Goh et al., JAMA Network Open: LLM influence on diagnostic reasoning, randomized trial
• Goh et al., Nature Medicine: GPT-4 assistance for management reasoning, randomized trial
• Gommers et al., Lancet: MASAI trial, interval cancer and screening accuracy
• Gong et al., JMIR: Meta-analysis of AI and cognitive workload and burnout in health care
• Grand View Research: AI in Healthcare Market report
• Grolleau et al., JAMA Network Open: AI-drafted hospital course summaries
• Hernström et al., Lancet Digital Health: Screening performance and characteristics of breast cancer detected in the MASAI trial
• HHS: AI Use Case Inventory
• Holmgren et al., JAMA Network Open: AI scribes and physician financial productivity at UCSF
• Kashiouris et al., Applied Clinical Informatics: Ambient AI rollout across 14 emergency departments
• KFF: Tracking Poll on Health Information and Trust, use of AI for health information and advice (Mar 2026)
• KFF: Tracking Poll on Health Information and Trust, use of social media and AI for health information (Jun 2026)
• Lång et al., Lancet Oncology: MASAI trial, first safety analysis
• Lawrence et al., JAMA Network Open: Patient consent for ambient AI documentation
• Lee et al., JAMA Health Forum: Recalls of FDA-cleared AI medical devices
• Lin et al., JAMA Health Forum: Benefit-risk reporting for FDA-cleared AI devices
• Lukac et al., NEJM AI: Randomized trial of two ambient AI scribes at UCLA
• MarketsandMarkets: Artificial Intelligence in Healthcare Market report
• Mehta et al., npj Digital Medicine: Transparency of FDA summaries for AI-enabled devices
• Menlo Ventures: 2025 State of AI in Healthcare
• MGMA Stat: Are AI tools making clinicians more productive?
• Microsoft: FY26 Q4 earnings call
• Muralidharan et al., npj Digital Medicine: Scoping review of FDA-approved AI devices and demographic reporting
• Naved et al., BMC Oral Health: FDA-cleared AI dental devices
• NCSL: Artificial Intelligence 2025 Legislation
• Nong and Platt, JAMA Network Open: Patients' trust in health systems to use AI
• Obermeyer et al., Science: Dissecting racial bias in an algorithm used to manage the health of populations
• Olson et al., JAMA Network Open: Ambient AI scribes and clinician burnout across 6 health systems
• Omar et al., Nature Medicine: Sociodemographic biases in medical decision making by LLMs
• ONC: HTI-1 Final Rule, Federal Register
• OpenAI: Launching Health in ChatGPT
• Pew Research Center: 60% of Americans would be uncomfortable with provider relying on AI in their own health care
• Pew Research Center: Americans want transparency when AI is used in their healthcare
• Pew Research Center: From diagnoses to treatments, why Americans use AI chatbots for health
• Poon et al., JAMIA: Adoption of AI in large US health systems
• Potnis et al., JAMA Internal Medicine: FDA-cleared AI products for breast cancer screening
• Precedence Research: Artificial Intelligence in Healthcare Market report
• Qian et al., BJU International: FDA-cleared AI devices in urology
• Ren et al., JAMA Network Open: Clinical evidence and recalls of FDA-authorized AI devices
• Rock Health: 2025 year-end digital health funding overview, a tale of two markets
• Rock Health: H1 2026 funding and market overview
• Rock Health: Health AI insights from the 2025 Consumer Adoption Survey
• Rotenstein et al., JAMA: Changes in clinician time expenditure and visit quantity with adoption of AI-powered scribes
• Shah et al., JAMIA Open: Patient experience with ambient AI scribes at Stanford Health Care
• Stanford HAI: 2026 AI Index Report, Medicine chapter (FDA data)
• Tempus AI: Q2 2026 results, SEC Form 8-K Exhibit 99.1
• U.S. Census Bureau: Business Trends and Outlook Survey, national data
• U.S. Census Bureau: Business Trends and Outlook Survey, sector data
• WHO: AI ethics and governance guidance for large multi-modal models
• Wright et al., JAMIA: Enterprise-wide ambient scribing rollout at Vanderbilt