AI in healthcare covers algorithmic systems — rule-based, classical machine learning, and generative AI — used in clinical workflows, diagnostic decisions, patient interaction, and health-system operations. By 2026 the sector carries empirical evidence of both rapid real-world deployment and measurable unintended consequences, documented in the AMA physician sentiment survey (AMA Physician AI Sentiment Report (2026)) and the Lancet endoscopist-deskilling RCT (Lancet Endoscopist Deskilling Study (2025)), set against regulatory scaffolding spanning the FDA Software as a Medical Device framework, California AB 3030, and HIPAA. The sector serves as a sectoral test case for the AI Deskilling thesis, for augmentation-then-displacement dynamics, and for the deployment-versus-evaluation gap that separates task-level capability from health-system outcomes.
Categories
Major categories of healthcare AI:
- Ambient clinical documentation (scribes) — systems that listen to patient-physician encounters and generate draft clinical notes.
- Diagnostic and screening AI — radiology, pathology, ophthalmology, and dermatology image-interpretation systems.
- Computer-aided detection (CADe) / diagnosis (CADx) — real-time AI prompts in procedures (for example, colonoscopy polyp detection).
- Clinical decision support — AI-driven triage, risk scoring, and treatment-option recommendations.
- Patient-facing AI — chatbots, symptom checkers, and mental-health apps.
- Administrative AI — prior authorization, coding, denials management, and scheduling.
Cross-cutting dimensions
Deployment versus evaluation
A recurring finding across healthcare AI is that deployed systems are evaluated more rarely than enterprise software deployment norms would suggest. Retrospective evaluations frequently find performance degradation when systems move from development cohorts to real deployment populations, attributed to distribution shift, different scanners, and different patient demographics. The enterprise deployment gap (Enterprise AI Deployment Gap) applies here with added clinical-safety stakes.
Augmentation versus substitution
Most healthcare AI is structurally augmentative: the physician remains the decision-maker and the AI supplies input. As physicians integrate AI into workflows, the augmentation/substitution line moves — ambient scribes handle documentation with minimal physician edit, and diagnostic AI can influence human interpretation of the same image. The Lancet colonoscopy finding shows that even in augmentative deployment, unaided performance can erode.
Applications
Ambient clinical documentation (scribes)
Ambient scribes (Nuance DAX Copilot, Abridge, Suki, Augmedix, and others) record the physician-patient encounter and generate draft clinical notes. Uptake has been rapid: the AMA 2026 survey found chart documentation is now one of the top physician AI use cases. Evidence on quality is mixed — draft notes reduce documentation time but introduce hallucination risk that requires physician review. These systems generally operate below the FDA device threshold because they generate draft text rather than making a clinical recommendation. On June 11, 2026, Nvidia disclosed it was developing an AI healthcare model with Abridge trained for clinical conversations (Source: wsj.com).
Diagnostic AI — radiology, pathology, ophthalmology
Diagnostic AI for image interpretation is the most-studied category in healthcare AI:
- Radiology — chest X-ray, mammography, and CT interpretation systems have accumulated multiple FDA clearances. Deployment at scale is uneven, and several large-scale real-world studies have shown narrower gains than development-cohort studies predicted.
- Pathology — digital pathology with AI overlay is emerging; the FDA cleared the first AI-based prostate cancer detection tool (Paige Prostate) in 2021.
- Ophthalmology — IDx-DR (now LumineticsCore) was the first FDA-authorized autonomous AI diagnostic, cleared in 2018 for diabetic retinopathy screening without physician review.
CADe in colonoscopy
Computer-aided detection (CADe) systems in colonoscopy are real-time AI prompts that highlight potential polyps. The Lancet 2025 multi-site study found that endoscopists who regularly used CADe showed a decline in unaided adenoma detection rate, from 28.4% to 22.4%, compared to baseline. This is described as hard-endpoint evidence of AI deskilling in a professional domain: even augmentative AI deployment can erode the professional's unaided competency, a degradation invisible to productivity metrics because combined human-plus-AI performance remains normal.
Regulation
Healthcare AI is regulated through multiple channels: the FDA Software as a Medical Device (SaMD) framework, for systems meeting the statutory definition of a medical device; state laws such as California AB 3030; HIPAA, governing protected health information; CMS and payer coverage rules for AI-driven diagnostic and decision-support tools; and medical malpractice, under which physician liability is assessed through standard-of-care analysis when AI is used.
FDA Software as a Medical Device (SaMD) framework
Key elements of the FDA's SaMD framework:
- Pre-market review — via 510(k), De Novo, or PMA depending on risk classification.
- Predetermined Change Control Plans (PCCP) — FDA guidance finalized in 2024 that permits manufacturers to specify in advance the kinds of algorithm updates allowed without a new submission, the agency's formal response to the locked-versus-continuously-learning algorithm problem.
- AI/ML-Based SaMD Action Plan (2021) — a framework document outlining the agency's regulatory approach.
- Good Machine Learning Practice (GMLP) — a 2021 joint FDA-Health Canada-MHRA document with 10 guiding principles.
As of 2026, the FDA has cleared or authorized several hundred AI-based medical devices, concentrated in radiology and cardiology. Generative AI applications based on large language models have been treated more cautiously; most ambient-scribe tools are explicitly positioned below the device threshold to avoid regulation.
California AB 3030 (healthcare generative AI disclosure)
California AB 3030 (Calderon, signed 2024) requires health facilities using generative AI to communicate clinical information with patients to disclose that the communication was AI-generated and to provide contact information for a human provider. It is the first US state-level AI-specific regulation of generative-AI patient communication. It applies to licensed clinics, health facilities, and physician offices, and excludes communications read and personally approved by a physician. The interaction with FDA's device framework contributes to the fragmentation discussed in AI Compliance Industry / Regulatory Fragmentation.
HIPAA implications
HIPAA's Privacy Rule and Security Rule govern protected health information (PHI) flowing through AI systems:
- Business Associate Agreements — required between covered entities and AI vendors that access PHI.
- De-identification — PHI used for training or evaluation must be de-identified per Safe Harbor (18 identifiers) or Expert Determination.
- Breach notification — AI-system breaches exposing PHI trigger notification requirements.
- Training-data ambiguity — whether model training on PHI constitutes a "use" or "disclosure" is not fully resolved under current OCR guidance.
Policy instruments
| Instrument | Scope | Status |
|---|---|---|
| FDA SaMD framework | Clinical decision devices | Active, ~hundreds cleared |
| FDA PCCP guidance | Algorithm update rules | Final 2024 |
| California AB 3030 (Healthcare AI Disclosure) | GenAI clinical communications | CA only |
| HIPAA Privacy/Security Rules | PHI handling | Pre-AI, retrofitted |
| CMS coverage decisions | Payment for AI tools | Emerging |
| Medical board guidance | Standard of care | Uneven |
| ONC HTI-1 (Decision Support) | EHR-embedded AI transparency | Active |
Professional-society and agency activity (2026)
The American Medical Association's House of Delegates adopted policies, made public June 11, 2026, calling for health AI tools to support rather than replace physicians (Source: insideaipolicy.com). The Department of Health and Human Services published a request for information on June 8, 2026 seeking input on using AI to evaluate substance-use and mental-health programs in real time (Source: insideaipolicy.com).
Deskilling and augmentation debate
The AMA survey reports physicians as subjectively concerned about skill erosion — 88% at least mildly concerned, and 70% concerned for trainees — while simultaneously reporting productivity and burnout-relief benefits. The Lancet study indicates the subjective concern is empirically validated in at least one procedural domain. Read together, the two pieces of evidence suggest healthcare is exhibiting an augmentation-first, deskilling-after pattern in real time, a pattern predicted but unconfirmed in software and customer-support contexts and directly observable in clinical practice.
Relationships
- depends-on: AMA Physician AI Sentiment Report (2026) — the sentiment-survey evidence
- depends-on: Lancet Endoscopist Deskilling Study (2025) — the hard-endpoint evidence
- depends-on: AI Deskilling — the mechanism made visible in the Lancet finding
- related: AI and Productivity — task-level productivity evidence with healthcare specifics
- related: California AB 3030 (Healthcare AI Disclosure) — the state-level disclosure regime
- related: Enterprise AI Deployment Gap — deployment friction applies with safety stakes
- related: Algorithmic Accountability and Bias Audits — FDA SaMD as a sectoral accountability regime
- related: AI Liability — physician and developer liability under malpractice and product regimes
- instance-of: sectoral AI governance under partial federal and state regulation
- related: AI for Science — AI in drug discovery and genomics is the high-stakes science-adjacent frontier of AI in healthcare