Key takeaways
- AI in healthcare spans four broad areas: medical imaging, drug discovery, clinical decision support, and operational efficiency.
- The FDA has authorized over 1,000 AI/ML-enabled medical devices as of 2025, up from just a handful a decade ago.
- DeepMind’s AlphaFold solved a 50-year-old problem in protein structure prediction and catalyzed a wave of AI drug discovery.
- Clinical AI faces heavy regulation, high evidence bars for deployment, and serious workflow-integration challenges that slow adoption.
- Large language models are being piloted for clinical documentation and summarization — the single biggest current deployment area by revenue.
Why healthcare is different
AI in healthcare has to clear bars that consumer AI does not. Mistakes can kill. Training data carries demographic biases that lead to real inequity. Regulators — FDA in the US, EMA in Europe, PMDA in Japan — require evidence before deployment. And clinical workflows are rigid, legacy-laden, and reimbursement-sensitive. As a result, AI progress in healthcare has lagged the consumer-tech hype but is now producing real, deployed systems.
The Stanford AI Index 2025 documents the acceleration: the FDA authorized its first AI-enabled medical device in 1995, approved six by 2015, and passed 223 by the end of 2023. The curve steepened sharply from 2020 onward.
Medical imaging
The single most mature application of AI in healthcare is imaging. Radiology, pathology, ophthalmology, dermatology — anywhere a trained clinician interprets images — is a natural fit for convolutional networks and vision transformers.
Systems like IDx-DR (approved in 2018) autonomously screen for diabetic retinopathy from fundus photos. Google Health’s mammography model demonstrated radiologist-level breast-cancer detection. Aidoc, Viz.ai, RapidAI, and similar companies deploy triage tools that flag strokes and pulmonary embolisms in CT scans, moving patients to care faster. For a deeper look at one modality, see our ai radiology coverage.
The deployment pattern is usually AI as a second reader or triage aid, not replacement. Regulatory, liability, and clinical workflow reasons all push toward AI as assistant rather than autonomous agent.
Drug discovery
Drug discovery is increasingly AI-first. The transformative moment was AlphaFold 2, DeepMind’s 2020-2021 system that predicted protein structures at near-experimental accuracy. AlphaFold solved a fundamental biology problem that had resisted 50 years of work. The AlphaFold 3 release in 2024 extended this to protein-ligand and protein-protein complexes, directly useful for drug design.
Beyond structure prediction, AI is applied at multiple drug-discovery stages: target identification (which protein to drug), hit discovery (finding candidate small molecules), lead optimization (improving candidate properties), and clinical-trial design. Companies like Insilico Medicine, Recursion, Isomorphic Labs, and Exscientia have brought AI-designed drug candidates into human trials. The first AI-designed drugs are now in Phase 2 and Phase 3, though no AI-discovered drug has yet been approved by FDA as of 2026.
The underlying techniques draw from our deep learning toolkit — graph neural networks for molecular structures, transformers for protein sequences, diffusion models for de-novo molecule generation.
Clinical decision support
Hospitals increasingly use AI to support clinicians with flagging, triage, and risk scoring. Sepsis prediction models, deterioration-risk scores, readmission predictors, and triage models run in background processes across major health systems. Epic’s embedded AI, and similar features in Cerner (now Oracle Health), push predictions into clinician workflows via the electronic health record.
Evidence of real-world benefit has been mixed. The original Epic sepsis model was criticized in a well-known 2021 JAMA paper for poor performance in external validation, a cautionary tale about deploying models without rigorous independent evaluation. More recent cohorts have improved, but the lesson — validate in your population, not just the training one — stuck.
Clinical documentation and summarization
The biggest emerging use case by revenue is AI scribing — using LLMs and speech recognition to draft clinical notes from patient-doctor conversations. Products from Abridge, Nuance DAX, Microsoft (which acquired Nuance), Ambience, and Suki run in thousands of clinics. They reduce documentation burden by 30-60 minutes per day per clinician, a widely-reported result that has driven fast adoption even where other AI tools have stalled.
LLMs are also being piloted for patient messaging, prior authorization drafting, and medical summary generation. These areas do not require FDA clearance (they are not clinical decision support) but do require HIPAA-compliant deployment and careful risk management.
Genomics and personalized medicine
AI helps interpret the massive data from modern sequencing. Variant classification (is this genetic change benign or disease-causing?), polygenic risk scores (how much does this person’s genome elevate their disease risk?), and cancer-treatment selection (which drug matches this tumour’s profile?) all use machine-learning models. The technology is moving into routine oncology and rare-disease workflows.
Challenges that have not gone away
Bias and fairness
Training data reflects existing healthcare inequities. Models trained primarily on one demographic underperform on others. A widely cited 2019 Science paper found a widely-used risk algorithm systematically underestimated illness severity in Black patients. Mitigations — diverse training data, bias audits, population-specific validation — are now expected, though not universally practiced.
Distribution shift
A model trained at one hospital may fail at another because imaging equipment, patient demographics, or coding practices differ. Robust deployment requires ongoing monitoring and often site-specific fine-tuning.
Regulatory and liability uncertainty
Who is liable when an AI-assisted diagnosis is wrong — the clinician, the hospital, the software vendor? Legal frameworks are still emerging. The EU AI Act classifies medical AI as high-risk, imposing additional conformance requirements. FDA continues to refine its Good Machine Learning Practice guidance.
Workflow integration
A brilliant model that adds three clicks to a clinician’s workflow will not get used. Decades of health-IT pain — EHR usability, alert fatigue, interoperability — shape how AI products must be built to stick.
What to expect in the next few years
Multimodal models that ingest imaging, text, and structured data together will become standard. Foundation models specialized for medicine (Med-PaLM, MedGemini, biomedical Llama variants) will move from benchmark demos to deployed assistants. Clinical-grade agents that can pull information from charts, draft orders, and flag abnormalities will start being piloted. None of this happens on the timescale of consumer AI; healthcare’s adoption curve is measured in multi-year deployments. For broader industry context, see our ai industry coverage.
Frequently asked questions
Can AI replace doctors?
Not on any realistic timeline, and that is not the goal of most clinical AI work. AI is most successful when it augments clinicians — reducing documentation burden, catching things a human might miss, prioritizing scarce attention. Full autonomy exists only in narrow screening applications (retinal disease, some dermatology) where the decision is binary and refer-out-to-specialist is the fallback.
Is my medical data used to train AI?
It depends on your provider, jurisdiction, and consent. Major AI-scribe deployments are typically structured so that audio is processed and discarded without being used for training. HIPAA in the US and GDPR in Europe constrain what can be done with identifiable patient data. Large-scale model training typically uses de-identified public datasets plus licensed data from partner institutions, not arbitrary patient records. Always check a provider’s privacy notice if you are concerned.
Has an AI-designed drug been approved?
As of early 2026, no AI-designed small-molecule drug has received full regulatory approval. Several are in advanced clinical trials — notably from Insilico, Exscientia, and Recursion — with readouts expected over the next one to three years. The industry consensus is that the first AI-designed drugs will clear approval within the current decade, though defining what counts as “AI-designed” (versus AI-assisted) remains fuzzy.
