What Medical AI Is Actually Being Used for Today

“Used” can mean several different things

Before describing adoption, it is necessary to define what counts as use.

Health questions submitted to a chatbot are directly observed consumer activity. Physician survey responses represent self-reported professional use, while integration into an electronic health record indicates organizational deployment. Regulatory authorization permits a medical device to be marketed for a defined purpose. Studies that measure effects on errors, workload, or outcomes provide evidence of evaluated use.

Those evidence states are related, but they are not interchangeable.

A product announcement does not prove adoption. Adoption does not prove clinical benefit. Regulatory authorization does not reveal how frequently a product is used. Strong performance on a benchmark does not establish that people will use the system effectively during a real health decision.

The clearest picture comes from keeping those distinctions visible.

Consumer-facing medical AI is already a health-information layer

One of the most visible forms of medical AI is the general-purpose assistant that people access through a phone or computer.

A Microsoft Research analysis examined more than 500,000 de-identified health-related Copilot conversations from January 2026. Health information and education accounted for about 40 percent of the sample. Other conversations involved symptoms, medical conditions, medications, fitness, emotional wellbeing, healthcare navigation, insurance, and medical paperwork.

Nearly one in five conversations involved personal symptom assessment or discussion of a condition. Some questions concerned another person, such as a child, parent, or partner. Mobile use leaned more toward personal health concerns, while desktop use included more research and paperwork.

This suggests that consumer-facing AI is already functioning as several things at once:

• a health-information interface;
• an early place to describe a symptom or concern;
• a caregiving aid;
• a way to understand medical terminology or test results;
• a navigation layer for providers, appointments, and insurance;
• and a workspace for health-related records and forms.

OpenAI reports an even larger platform-scale signal. Based on its de-identified analysis, more than 230 million people globally ask health and wellness questions on ChatGPT each week. That figure reflects health-related activity across ChatGPT generally. It is not a count of people using the dedicated ChatGPT Health product.

The ChatGPT Health announcement describes a separate health-focused space that can connect medical records and wellness applications so people can discuss test results, prepare for appointments, examine health patterns, or consider insurance questions. OpenAI states that it is intended to support medical care rather than replace it and is not intended for diagnosis or treatment.

Taken together, the Microsoft and OpenAI evidence shows that health use is no longer an edge case for conversational AI. People are already using these systems as an accessible interface to health information and healthcare complexity.

The available evidence does not show whether those conversations lead to better decisions, appropriate care, or improved health outcomes. It shows that the activity is occurring at substantial scale.

Information access is not the same as reliable diagnosis

Consumer health use can move quickly from general education into individualized decision-making.

Someone asking how a medication works may only be seeking background information. The same person might then enter symptoms, medications, and test results to decide what condition they have or whether urgent care is needed.

The visual boundary between those activities can be thin even when the safety requirements are very different.

A randomized study published in Nature Medicine found that the evaluated language models performed strongly when answering medical scenarios independently. However, people using those models did not outperform the control group when identifying the relevant condition or selecting the appropriate level of care.

The result illustrates an important distinction: a model can demonstrate substantial medical knowledge without reliably improving the decision made by the person interacting with it.

Consumer-facing medical AI is therefore most defensibly described today as an information, preparation, and navigation layer. It may help people organize questions, understand terminology, or prepare for a professional conversation. Widespread use should not be confused with evidence that autonomous diagnosis or triage is working safely at population scale.

Clinical medical AI in healthcare workflows

The professional side of medical AI looks different from consumer use.

In the American Medical Association’s 2026 physician survey , 72 percent of respondents reported incorporating at least one of 17 named AI use cases into their practice. Another 9 percent were uncertain which AI tools their practice offered, while 19 percent reported no use cases.

That distinction matters. The survey’s broader 81 percent figure combines physicians reporting at least one use case with those reporting awareness or uncertainty. It should not be presented as 81 percent verified deployment.

The most common reported use was summarizing medical research and standards of care, selected by nearly four in ten physicians. Other prominent uses included generating discharge instructions, care plans, progress notes, billing documentation, chart summaries, patient-portal responses, translations, and assistive diagnostic support.

These tasks are not equivalent. Drafting a note is different from recommending a diagnosis. Summarizing research is different from determining treatment. Translating patient instructions is different from estimating clinical risk.

But the pattern is consistent: AI is entering clinical work most visibly where healthcare retrieves, transforms, documents, and communicates large quantities of information.

OpenAI separately reports that millions of clinicians worldwide use ChatGPT each week for care consultation, writing and documentation, and medical research. The company’s clinician product announcement describes those as vendor-observed usage patterns rather than independently audited adoption statistics.

Physicians are also using specialized evidence tools

Some medical AI systems are designed specifically to answer clinical questions using cited medical literature.

OpenEvidence is one example. It provides verified healthcare professionals with literature-grounded responses to questions involving diagnosis, treatment choices, medication safety, guidelines, prognosis, and other issues that arise during care.

The company reports that its system has supported more than 200 million AI-powered clinical consultations from U.S. physicians and other frontline clinicians. It has also reported more than one million consultations per day from logged-in verified clinicians. Those figures are company-reported activity measures, not independently audited adoption statistics, and one consultation should not be treated as equivalent to one patient encounter.

A June 2026 preprint evaluating real point-of-care questions provides a more detailed view of this use. Researchers assembled 620 de-identified and rewritten questions submitted to OpenEvidence by physicians across 30 specialties. The questions included treatment and toxicity, clinical evaluation and management, diagnosis, and prognosis.

In blinded comparisons, 149 practicing physicians rated answers from OpenEvidence and three general-purpose models across accuracy, clinical utility, source quality, verifiability, and completeness. The specialized system received the highest ratings across those dimensions.

The result supports a broader point: medical AI does not need to become one universal system. A tool engineered around a defined clinical workflow, trusted evidence sources, and the needs of a specific professional group may perform differently from a general-purpose assistant.

The study still requires caution. The questions originated from the OpenEvidence platform, the company implemented parts of the data collection, and the work was released as a preprint. It evaluated answer quality, not whether using the system improved patient outcomes or clinical decisions.

OpenEvidence nevertheless represents a distinct and already active form of medical AI use: physicians consulting a specialized, literature-grounded system about questions that arise during practice.

Ambient documentation is a representative clinical use

Ambient clinical documentation is one of the clearest examples of generative AI entering routine healthcare work.

An ambient scribe captures a patient-clinician conversation and generates a draft visit note. The clinician remains responsible for reviewing, correcting, and approving the documentation.

In a quality-improvement study involving 46 clinicians across 17 specialties , ambient-scribe use was associated with about 20 percent less time spent on notes per appointment and 30 percent less after-hours work time per workday. Clinicians also provided a mixture of positive, negative, and qualified feedback about the generated notes.

The significance of ambient documentation is not that the system has become the clinician. It is that AI is converting an existing clinical interaction into a draft record.

The same general pattern applies to chart summaries, discharge instructions, referral letters, coding support, and portal messages:

The system produces a draft, summary, classification, or recommendation. A healthcare professional or organization remains responsible for how that output enters care.

Hospitals were already using predictive AI before generative AI arrived

Not all current medical AI is conversational.

Predictive models have been integrated into hospital systems for years. They can classify patients or produce risk scores involving readmission, inpatient deterioration, outpatient follow-up, appointment attendance, billing, scheduling, and other clinical or operational functions.

An analysis from the federal health IT office found that 71 percent of responding non-federal acute-care hospitals reported using predictive AI integrated with their electronic health record in 2024, up from 66 percent in 2023.

The most common applications involved predicting inpatient health trajectories or risks. Some of the fastest-growing uses involved billing automation and scheduling. Adoption was substantially lower among small, rural, independent, government-owned, and critical-access hospitals than among larger or system-affiliated institutions.

This corrects a common public image of medical AI.

Some of the most widely distributed systems are not assistants that a patient or clinician deliberately opens. They operate within an electronic health record or administrative workflow and generate a score, flag, recommendation, or priority.

Their effects can still be consequential. A model may influence which patient is reviewed first, who is contacted for follow-up, which appointment receives attention, or where limited resources are directed.

Generative AI is moving into the electronic health record

Generative systems are now entering the same institutional infrastructure.

A national survey study of 2,174 non-federal U.S. hospitals found that 31.5 percent reported using generative AI integrated with their electronic health record in 2024. Another 24.7 percent planned to implement it within one year.

That study measured organizational integration. It did not establish that every hospital used the same application, used it at the same intensity, or obtained the same result.

A hospital reporting generative-AI integration could be supporting documentation, summarization, patient-message drafting, coding, research, administrative work, or another function. The result should not be converted into a claim that nearly one-third of hospitals were using ambient scribes or any other single product category.

The more defensible conclusion is broader: generative AI is becoming part of hospital information infrastructure, but its functions, maturity, and intensity vary substantially by organization.

Clinical copilots are beginning to operate as supervised safety nets

Some implementations move beyond information retrieval and documentation into supervised clinical decision support.

Penda Health integrated an OpenAI-powered clinical copilot into primary-care workflows in Kenya. The system was designed to provide recommendations at defined points during a visit and to alert clinicians when it identified a possible omission or error, while leaving the clinician in control.

In a study covering 39,849 patient visits across 15 clinics , clinicians with access to the system had a 16 percent relative reduction in diagnostic errors and a 13 percent relative reduction in treatment errors compared with clinicians without it.

OpenAI partnered with Penda on the work and participated in the analysis, so that relationship should remain visible when interpreting the findings. The study also did not find a statistically significant difference in short-term patient-reported recovery or additional-care rates.

This is more consequential than drafting a note, but the operating model remains supportive. The copilot did not independently control the encounter. It surfaced recommendations inside an existing clinician-led workflow.

Conversational clinical AI remains supervised and experimental

Google’s AMIE research shows another possible direction: conversational AI used before a clinical visit.

In a prospective single-centre feasibility study , AMIE conducted pre-visit clinical history-taking with 100 adult patients before ambulatory primary-care appointments. A physician monitored the interactions and was prepared to intervene under predefined safety criteria.

The system produced a transcript and summary for the treating clinician. Google reported that no safety stops were required and that clinical reviewers rated several aspects of the system’s differential diagnoses and management plans as broadly comparable with those produced by primary-care clinicians.

The study was designed to evaluate feasibility, safety, and acceptance. It involved one centre, live physician oversight, and no controlled comparison with the normal workflow. It does not establish that the system improves care or is ready for unsupervised deployment.

That distinction is important. Medical AI can be genuinely used in a clinical study without yet being routine clinical infrastructure.

Administrative healthcare work is part of medical AI too

Medical AI is not limited to conversations between patients and clinicians.

Healthcare organizations are also applying AI to coverage policies, prior authorization, claims appeals, medical coding, provider verification, care coordination, clinical-trial operations, biomedical literature, and regulatory work.

Anthropic’s healthcare and life-sciences offering , for example, includes connectors for Medicare coverage policies, ICD-10 codes, provider registries, and PubMed. It describes workflows involving prior authorization, appeals, coding, ambient documentation, chart review, and care coordination.

These materials demonstrate product availability and intended workflow categories. They do not independently establish adoption levels or clinical benefit.

The relevance is not that every major AI company now has a healthcare product. It is that the same functional divisions are appearing repeatedly across the market:

• patient information and navigation;
• clinical evidence retrieval;
• documentation and communication;
• administrative and reimbursement work;
• hospital prediction and prioritization;
• supervised clinical decision support;
• and scientific or regulatory research.

That repeated structure is more informative than a catalogue of vendors.

AI-enabled medical devices form another distinct layer

Medical AI also exists inside regulated devices and device software.

The U.S. Food and Drug Administration’s AI-enabled medical-device list identifies products that have met applicable premarket requirements for authorization in the United States. These systems can analyze medical images, physiological signals, measurements, and other defined clinical inputs.

This category should be kept separate from consumer chatbots, enterprise assistants, and hospital administrative models.

An authorized device operates under a defined intended use and a medical-device regulatory framework. A general-purpose chatbot answering a health question does not automatically become a medical device because the conversation concerns medicine.

Authorization also does not prove widespread use. The FDA list identifies products authorized for marketing. It does not show how frequently each product is purchased, activated, used in routine care, or monitored after deployment. The FDA also notes that the list is not comprehensive.

What medical AI is actually being used for

The strongest available evidence supports several current-use categories.

Consumer information and navigation

People use conversational systems to ask about symptoms, conditions, medications, test results, fitness, emotional wellbeing, caregiving, insurance, providers, appointments, and medical paperwork.

Clinical evidence retrieval and synthesis

Physicians use specialized systems to search medical literature, compare evidence, investigate treatment and safety questions, and obtain cited responses to point-of-care clinical questions.

General clinical knowledge work

Clinicians use broader assistants to summarize research, review standards of care, organize information, and prepare material for professional review.

Documentation and communication

AI generates draft notes, chart summaries, discharge instructions, referral letters, billing documentation, translations, and patient-portal responses.

Prediction and prioritization

Hospitals use predictive models to classify risk, identify patients for follow-up, support scheduling, and prioritize clinical or operational attention.

Administrative healthcare work

AI is being integrated around coverage review, prior authorization, claims appeals, coding, revenue-cycle work, and care coordination.

Supervised clinical decision support

Clinical copilots and research systems are being evaluated as tools that identify possible errors, collect histories, generate differential diagnoses, or prepare recommendations while clinicians remain responsible for decisions.

Bounded medical-device functions

Authorized AI-enabled devices perform defined tasks involving images, signals, measurements, classification, detection, or other specified medical functions.

What is not established

The evidence does not support describing medical AI as one system that independently manages care from first symptom to final treatment.

It does not establish that:

• consumer chatbots reliably diagnose or triage people;
• widespread use produces better health outcomes;
• physician self-report confirms organizational deployment;
• hospital integration means every clinician regularly uses the system;
• regulatory authorization proves routine adoption;
• benchmark performance predicts effective human use;
• or one general-purpose model is the most accurate medical AI across every task.

The more consequential the task, the more important workflow design, validation, monitoring, escalation, and human accountability become.

The AMA survey reflects that concern. Physicians identified validated safety and efficacy, privacy assurances, clearer liability frameworks, post-market monitoring, education, and practical implementation guidance as important conditions for broader adoption.

The clearest answer today

Medical AI is already being used.

At the consumer level, people use assistants to explain medications, understand results, prepare for appointments, and navigate insurance.

In clinical practice, physicians search medical literature, investigate point-of-care questions, draft notes, summarize charts, prepare discharge instructions, and respond to patient messages.

Hospitals use AI to predict inpatient risk, identify patients for follow-up, support scheduling, and integrate generative tools into electronic health records.

Higher-consequence uses are also emerging. Supervised copilots can flag possible clinical omissions, research systems can collect pre-visit histories, and authorized software can analyze medical images or physiological signals.

What is less established is the idea of a single artificial system replacing the judgment, responsibility, context, and relationships that hold those activities together.

For now, medical AI is becoming a distributed support layer across healthcare.

Its most common roles are not necessarily the most cinematic ones. They are information retrieval, evidence synthesis, documentation, navigation, prediction, communication, administration, and bounded technical assistance.

That may be the more important development to watch.

A revisitable assessment

This assessment describes documented use available in July 2026. The functional categories may remain relatively stable while adoption, evidence, regulation, and the balance between consumer and clinical use continue to change.

New AI Health will revisit the question as stronger real-world adoption data, comparative evaluations, safety findings, and measured outcomes become available.

Editorial boundary

This article distinguishes consumer activity, physician self-report, organizational deployment, regulatory authorization, and evaluated clinical use. None should be treated as proof that a system improves patient outcomes unless an appropriate study measured that result.

This publication is informational. It is not medical advice, a clinical recommendation, procurement guidance, or a regulatory determination.