The State of AI in Healthcare: A Comprehensive Review
A comprehensive review of recent developments reveals a significant expansion of artificial intelligence (AI) tools across multiple domains of healthcare. These developments range from clinical diagnostic studies and new telehealth programs to increased public use of AI for health information and related policy proposals. The integration of AI presents a complex landscape of demonstrated capabilities, ongoing limitations, and varying stakeholder perspectives.
AI in Clinical Diagnosis and Medical Research
Diagnostic Accuracy Studies
A study published in the journal Science examined an AI reasoning model developed by OpenAI. Researchers from Harvard Medical School and Beth Israel Deaconess Medical Center found that the AI model matched or outperformed physicians in diagnostic accuracy at three stages: triage, emergency department care, and hospital admission.
"The model outperformed a large physician baseline." — Raj Manrai, study author
The study tested the AI on real-world emergency department cases, case reports from the New England Journal of Medicine, and clinical vignettes. The AI used only electronic health records and information available to physicians at the time of diagnosis. In one case, the AI correctly attributed a patient's worsening symptoms to lupus-related heart inflammation.
Dr. David Reich of Mount Sinai Health System commented that while the AI appeared accurate and potentially ready for clinical use, questions remain about integration into workflows.
The study's authors stated that the findings do not support replacing physicians with AI. Caveats include that the study relied solely on text data, and real-world clinical practice involves images, sounds, and nonverbal cues. The authors noted that further rigorous, forward-looking trials are needed to assess clinical impact.
Rare Disease Diagnosis
A study published in NEJM AI reported that OpenAI's o3 Deep Research model assisted researchers at Boston Children's Hospital in identifying diagnoses for 18 children with previously undiagnosed rare diseases. The research, a collaboration between the hospital's Manton Center for Orphan Disease Research and OpenAI, analyzed the genomes of 376 patients.
The model provided a diagnosis at a rate of approximately 4.8%. Diagnoses included:
- 10 cases of rare neurodevelopmental diseases
- 4 neuromuscular disorders
- 2 sudden deaths in children
- 2 early childhood psychosis illnesses
Seven diagnoses were "rediscoveries" — diagnoses previously identified by one treatment team but not shared globally.
"Almost 5% new diagnoses is a significant number." — Catherine Brownstein, scientific director of the Manton Center
Researchers provided the model with clinicians' notes, symptom descriptions, and filtered gene lists. A human team reviewed all outputs to confirm diagnoses. Adam Rodman, not involved in the study, said a diagnostic yield of 5% is meaningful and could serve as a screening tool. Chunhua Weng of Columbia University called the study a contribution but cautioned that results require rigorous human review. The study was financially supported by OpenAI.
AI in Primary Care Access
AI-Supported Telehealth Program
Mass General Brigham (MGB) launched Care Connect, an AI-supported telehealth program, in September. The program aims to address a primary care provider shortage. Patients interact with an AI agent via an app to describe symptoms. The AI compiles a summary for a primary care doctor, who provides care through video appointments, often within one to two days.
Care Connect operates 24/7 with 12 remotely located physicians. It is designed for common urgent care needs, colds, rashes, sprains, mild to moderate mental health concerns, and issues related to chronic diseases. The AI tool assists by suggesting diagnoses and treatment plans.
MGB committed $400 million over five years to primary care services, including the Care Connect initiative. By February, MGB plans to extend services to all insured residents of Massachusetts and New Hampshire. The program is not for emergencies or physical exams.
Some primary care doctors within MGB expressed concerns that the investment in AI should instead be directed towards increasing salaries and retaining staff. They stated that Care Connect could diminish access to in-person primary care over time.
Dr. Steven Lin of Stanford University School of Medicine suggested AI tools are safest for urgent care issues but acknowledged AI-generated summaries can enhance physician efficiency.
National Primary Care Shortage Context
A national shortage of primary care providers affects approximately 17% of adults in America, according to reports. This issue is particularly pronounced in Massachusetts, where the primary care workforce is declining faster than in most other states, according to a January 2025 report.
AI in Rural Healthcare Proposal
Dr. Mehmet Oz, head of the Centers for Medicare and Medicaid Services (CMS), proposed using AI to address the rural healthcare crisis. The proposal is part of the Trump administration's $50 billion plan for rural healthcare modernization. Oz suggested deploying digital avatars for basic medical interviews, robotic systems for remote diagnostics, and drones for medication delivery. He specifically suggested replacing some in-person obstetric care with AI-guided devices, such as robots conducting ultrasounds.
CMS issued a statement indicating Oz's comments emphasized the need to "responsibly explore tools" to extend the capabilities of licensed clinicians, not to replace them. CMS supports evidence-based, patient-centered AI tools under clinical oversight.
The proposal follows federal Medicaid spending reductions under the One Big Beautiful Bill Act. Over 190 rural hospitals closed between 2005 and early 2024. A 2024 CDC report found rural residents are more likely to die early from five leading causes compared to urban populations, attributed to limited access to providers, longer travel times, and higher poverty rates.
Carrie Henning-Smith of the University of Minnesota criticized the proposal, arguing AI avatars would remove human connection from healthcare and expressed concerns about implementing unproven technology in underserved populations.
She noted logistical challenges such as unreliable broadband. Matt Faustman of Honey Health argued AI tools could assist rural communities by automating administrative burdens. Oz has not presented a comprehensive implementation plan.
Public Use of AI for Health Information
Prevalence of Use
A Gallup poll published in late 2025 found that approximately one-quarter of U.S. adults reported using an artificial intelligence tool for health information or advice within the 30 days prior to the survey. This finding is supported by at least three other recent surveys.
- A KFF poll from late February found about 3 in 10 U.S. adults had sought health information from AI tools in the past year
- About 8 in 10 had sought information from a doctor or healthcare professional
- A Pew Research Center survey from October found about 2 in 10 U.S. adults get health information at least sometimes from AI chatbots
- About 85% get it from healthcare providers
User Motivations and Demographics
According to the Gallup survey, about 7 in 10 recent AI health users cited wanting quick answers, additional information, or curiosity as their reason. Majorities used AI for research before or after seeing a doctor. A smaller share cited barriers to accessing professional care, including cost, inconvenience, or lack of time.
Specific reasons included:
- Wanting help outside normal business hours (about 4 in 10)
- Not wanting to pay for a doctor's visit (about 3 in 10)
- Not having time for an appointment, past negative experiences, or embarrassment (about 2 in 10 for each)
The KFF survey indicated younger adults and people with lower incomes were more likely to use AI for health information due to cost or access difficulties.
Trust and Privacy Concerns
The Gallup poll reported that about one-third of recent AI health users said they "strongly" or "somewhat" trust the accuracy of AI-generated health information, about one-third distrusted it, and another third were neutral.
A KFF poll found about three-quarters of U.S. adults expressed concern about the privacy of personal health information provided to AI tools.
Dr. Bobby Mukkamala, president of the American Medical Association, stated AI should be considered a tool and not a substitute for medical care. Dr. Karandeep Singh of UC San Diego Health described AI tools as an upgraded version of web searches for health information, providing an executive summary. Tech experts have noted AI chatbots can sometimes provide false information.
AI Chatbot Performance Studies
A study published in Nature Medicine simulated user interactions with AI chatbots for medical scenarios. Participants correctly identified hypothetical conditions approximately one-third of the time after consulting AI tools. Only 43% made correct decisions regarding next steps.
Researchers from Oxford University noted that users often lack knowledge on how to effectively prompt AI models. An example cited involved different user descriptions for the same severe headache leading to vastly different AI advice.
Another study found that AI bots "under-triaged" 52% of emergency medical cases, meaning they downplayed the seriousness of the ailment. Researchers noted AI struggled with time-sensitive scenarios. An OpenAI spokesperson stated one of the studies utilized an older version of ChatGPT and did not reflect current usage.
KFF Poll on Vaccine Misinformation
A KFF poll found that US adults who frequently seek health advice from AI chatbots are more likely to believe vaccine misinformation. Among frequent AI health users, 35% believe it is 'definitely or probably true' that MMR vaccines cause autism, compared to 20% of non-AI users.
The correlation persisted after controlling for age, race, education, and political partisanship. Adults who use social media for health information at least weekly are more than twice as likely to believe the MMR-autism myth.
AI in Mental Healthcare
Workforce Concerns and Labor Actions
Psychologist Vaile Wright of the American Psychological Association noted that fear and anxiety about AI in the mental healthcare field include concerns about job replacement.
In March, 2,400 mental health care providers for Kaiser Permanente in Northern California participated in a 24-hour strike. The strike addressed concerns about changes to the triage system. Ilana Marcucci-Morris, a licensed clinical social worker at KP, stated that a 10-15 minute screening by a licensed clinician is now conducted by unlicensed lay operators following a script or via E-visits.
Kaiser Permanente stated its use of AI does not replace clinical expertise and confirmed it is evaluating AI tools from Limbic to assist members in accessing care.
Current AI Applications
Wright reported that AI adoption in mental health care has focused on administrative tasks such as documentation, billing, and updating electronic health records. Companies like Blueprint offer AI assistants for session summaries and patient progress tracking. Limbic develops AI assistants for health systems, with its chatbot Limbic Care trained in cognitive behavioral therapy skills.
Dr. John Torous of Beth Israel Deaconess Medical Center indicated that widespread clinical use of AI is not yet common due to a lack of testing and high implementation costs.
Statements from Health Tech Companies
OpenAI launched ChatGPT Health in January, described as a platform with enhanced security for medical data sharing. OpenAI states that hundreds of millions of individuals consult ChatGPT weekly for wellness advice and that health is one of the most common uses of ChatGPT. OpenAI stated its technology should not be used for self-diagnosis.
Anthropic offers similar features through its Claude chatbot. Both companies state their large language models are not substitutes for professional medical care.
Allon Bloch, CEO of K Health, stated that technology and AI are essential to addressing healthcare challenges related to cost, quality, and access.
Reported Limitations and Risks
Data Privacy
Sharing personal medical information with AI companies through chatbots is not protected by the federal HIPAA law. OpenAI and Anthropic state that user health information is kept separate, subject to additional privacy protections, and not used for model training.
Dr. Singh noted most AI tools have settings users can adjust to prevent data use for training models, but this requires user vigilance. An incident was noted where private ChatGPT conversations were discovered indexed on a public website without users' knowledge.
Clinical Limitations
The diagnostic study relying solely on text data did not account for images, sounds, or nonverbal cues.
The emergency department represents only a small portion of patient care. LLM outputs require verification by human experts. Experts emphasized that individuals experiencing urgent symptoms should seek immediate medical attention rather than consulting a chatbot.