Quick Facts

• The FDA had authorized 1,451 AI-enabled medical devices by the end of 2025, with 1,104 (76%) in radiology alone.
• 81% of US physicians report using AI professionally in 2026, up from 38% in 2023 (AMA).
• Severe-harm potential from AI clinical recommendations ranged from 2.9% (best specialized tools) to 24.6% (general-purpose models) in the NOHARM benchmark.

This page collects verified, primary-sourced statistics on AI in healthcare, organized by category. Every number here is traced to its original source rather than repeated from a secondary aggregator. Where a study has been revised or updated, we cite the current version and note it. This page is updated periodically as new data becomes available.

FDA Regulatory Data on AI-Enabled Devices

• 1,451 AI-enabled medical devices had been authorized by the FDA by the end of 2025, with 1,104 (76%) of those in radiology. Source: The Imaging Wire, analysis of the FDA’s AI-Enabled Medical Device List.
• In Q4 2025 alone, the FDA cleared 72 AI-enabled devices, 55 (76%) of which were radiology tools, confirming the specialty’s continued dominance in new clearances. Source: The Imaging Wire.
• In Q2 2026, the FDA authorized 86 AI/ML devices spanning cardiovascular, neurology, GI/urology, anesthesiology, orthopedics, dental, surgery, hematology, pathology, clinical chemistry, and microbiology, signaling clearance activity is beginning to spread beyond imaging. Source: Innolitics.
• As of the most recent review, no FDA-authorized device uses generative AI or is powered by a large language model. The tools doing the heaviest clinical lifting today are narrow classifiers, not conversational AI. Source: “From Disclosure to Self-Referential Opacity: Six Dimensions of Strain in Current AI Governance”.

Physician Adoption of AI

• 81% of physicians report using AI professionally in 2026, more than double the 38% who reported using it in 2023. Source: American Medical Association, 2026 Physician Survey on Augmented Intelligence (nearly 1,700 physicians surveyed).
• The average physician now uses 2.3 distinct AI use cases, up from 1.1 in 2023, according to the same AMA survey.
• The most common physician uses of AI are summarizing medical research (39%) and generating discharge instructions, care plans, or progress notes (30%): administrative and research support, not autonomous diagnosis.
• 77% of physicians say AI improves their ability to care for patients, up from 65% in 2023. At the same time, 88% worry about skill loss from over-reliance on AI, and 85% want a direct say in how AI gets adopted at their institutions, per the same AMA survey.

Hospital and Health System Adoption

• A national survey of 2,174 nonfederal acute care hospitals found 31.5% were already using generative AI integrated with their EHR by 2024, with 24.7% planning to adopt within a year and 43.7% delayed or uncertain. Source: JAMA Network Open.
• Adoption correlates strongly with resources: major teaching hospitals (53.9%) and system-affiliated hospitals (38.5%) were far more likely to be early adopters than independent hospitals (16.3%), per the same study.
• Generative AI adoption among US healthcare organizations rose from 25% in late 2023 to 47% in 2024, reaching 50% by the end of 2025, with more than 80% of surveyed leaders reporting at least one deployed use case. Source: Becker’s Hospital Review, McKinsey survey.
• 85% of Epic’s health system customers are now live with generative AI across its Art, Emmie, and Penny copilot tools, which include AI-drafted patient message replies. Source: Epic.

AI Diagnostic Accuracy: What the Studies Show

• Diabetic retinopathy screening systems show pooled sensitivity around 93% and specificity around 90% across regulator-approved systems. Source: npj Digital Medicine meta-analysis (82 studies, 887,244 examinations).
• A foundation-model-powered body CT triage tool covering 14 conditions posted 97% mean sensitivity and 98% mean specificity in its pivotal study (Aidoc, January 2026 FDA clearance). Source: Aidoc.
• In a prospective, multicenter study of intracranial hemorrhage detection across 67 organizations analyzing 3,409 brain CT studies, radiologists using AI as an assistive tool significantly outperformed standalone AI: sensitivity 98.91% (AI-assisted) versus 95.91% (standalone); specificity 99.83% versus 87.35%. Source: Journal of Clinical Medicine.
• Fracture detection tools show pooled sensitivity around 90-92% and specificity around 91% across dozens of imaging studies. Source: “Artificial Intelligence in Fracture Diagnosis on Radiographs: Evidence, Pitfalls, and Pathways for Clinical Integration”.
• Using the New England Journal of Medicine’s archive of especially difficult diagnostic cases, one model’s first-guess diagnosis was correct in about 52% of 143 hard cases, with the correct answer appearing in its differential in about 78% of cases. On a head-to-head comparison of 70 overlapping cases, a newer model reached the exact or very close diagnosis in about 89% of cases, versus roughly 73% for an earlier model. Source: “Superhuman performance of a large language model on the reasoning tasks of a physician”.
• General-purpose generative AI averaged only 52.1% accuracy across 83 studies on open-ended diagnosis (95% CI: 47.0-57.1%), close to a non-expert clinician’s performance and well behind specialty-trained diagnostic models. Source: npj Digital Medicine.

Safety and Risk Data

• In the NOHARM benchmark (Stanford/Harvard), testing 20 generalist LLMs and 4 specialized clinical AI tools against 1,100 real case-based tasks, severe-harm potential reached up to 24.6% for the models tested, with more than 80% of severe errors being errors of omission (recommending too little) rather than active harm. Source: “First, do NOHARM”.
• Specialized clinical AI tools performed dramatically better than general-purpose models in that same benchmark: AMBOSS LiSA, Doximity Ask, OpenEvidence, and Glass Health scored severe-harm rates as low as 2.9% to 5.4%.
• A Mount Sinai analysis found ChatGPT Health under-triaged 52% of genuine medical emergencies in a structured test, often steering people away from urgent care when they needed it most. Source: Nature Medicine.
• A BMJ Open audit of five major AI chatbots (Gemini, DeepSeek, Meta AI, ChatGPT, Grok) found nearly half of answers to common health questions contained misleading or problematic information. Source: BMJ Open.
• The February 2024 ransomware attack on Change Healthcare compromised the protected health information of approximately 192.7 million people, the largest healthcare data breach ever recorded. Source: HHS Office for Civil Rights.
• Analysis of a widely used AI transcription engine (Whisper) found fabricated content in roughly 1.4% of transcriptions, with an estimated 38% of those hallucinations carrying potential for harm. Source: AP News, via Fortune.
• Only about 5% of AI-enabled medical devices had reported adverse-event data by mid-2025, including device malfunctions and at least one death linked to device malfunction. Source: IntuitionLabs analysis of FDA data.

Patient-Facing AI: Messaging and Portal Tools

• In a JAMA Network Open study, 16 primary care physicians assessed 344 pairs of AI- and human-written patient portal message replies without knowing which was which; accuracy, completeness, and tone did not differ statistically, and AI responses outperformed humans in understandability and tone by 9.5%. Source: Becker’s Hospital Review.
• In a Mass General Brigham study (published in The Lancet Digital Health), physicians judged AI-drafted patient message replies safe in 82.1% of cases and acceptable to send without further editing in 58.3% of cases. If left unedited, 7.1% of responses could have posed a risk to the patient, and 0.6% could have posed a risk of death. Source: Mass General Brigham.
• Patients say they will accept AI-drafted portal messages only if a clinician reads every word before it reaches them, based on interviews with 40 patients from a large academic health system (JAMA Network Open, July 2026). Source: telehealth.org.

Physician Workforce and Capacity Data

• The AAMC projects a shortage of up to 86,000 physicians in the US by 2036, driven largely by population aging. Source: AAMC.
• By 2034, Americans 65 and older are expected to outnumber children under 18 for the first time in US history, per AAMC workforce data.

How We Verify These Numbers

Each figure above is checked against its original source (a peer-reviewed study, an official regulatory filing, or a primary organizational report) rather than a secondary aggregator, and linked directly so you can verify it yourself. Where a study has been revised, we cite the current version rather than an earlier draft. If you find an error or a newer figure, contact us and we’ll correct it.

FAQ

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

The FDA had authorized 1,451 AI-enabled medical devices by the end of 2025, with 1,104 (76%) of them in radiology. The agency is now clearing roughly 30 new AI devices a month, up from about 21 a month in 2024, and clearance activity is starting to spread into other specialties like cardiovascular, neurology, and pathology.

What percentage of doctors use AI in their practice?

81% of US physicians reported using AI professionally in 2026, more than double the 38% who reported using it in 2023, according to the American Medical Association’s annual survey. Most of that use is administrative: summarizing medical research and drafting documentation, not autonomous diagnosis or treatment decisions.

How accurate is AI at diagnosing patients?

It depends heavily on the type of AI and the task. Narrow, purpose-built diagnostic tools perform very well: diabetic retinopathy screening reaches about 93% sensitivity, and imaging triage tools have posted sensitivity above 97% in FDA pivotal studies. General-purpose generative AI is far weaker at open-ended diagnosis, averaging only 52.1% accuracy across 83 studies, close to a non-expert clinician’s performance.

What are the biggest safety risks of AI in healthcare?

The two most documented risks are clinical safety and data security. In the NOHARM benchmark, general-purpose AI models carried potential for severe harm in up to 24.6% of cases, though specialized clinical tools performed dramatically better. Separately, the February 2024 Change Healthcare breach compromised the health information of approximately 192.7 million people, the largest healthcare data breach ever recorded.

Related Reading

For deeper coverage of any category above, see our related pieces: 7 Real Risks of AI in Healthcare, AI in Healthcare Diagnosis, Will AI Replace Doctors?, and Generative AI in Healthcare: Tools and Adoption.