AI in Healthcare: What It Does, Where the Evidence Is, and the Risks

Executive summary

AI in healthcare is the use of machine-learning systems — increasingly foundation and generative models — to support diagnosis, treatment, drug discovery, documentation and hospital operations. The evidence base is now substantial rather than speculative. The US FDA had authorized roughly 950 AI-enabled medical devices as of August 2024, up from six in 2015, and cleared a further 258 in 2025 (Stanford HAI AI Index 2025 and 2026). The market is heavily concentrated in imaging: across 1,016 authorizations, 76% were radiology devices and about 96% went through the 510(k) pathway, so most were cleared without new clinical trials (npj Digital Medicine, 2025). Where randomized evidence exists, results are strong — Sweden's MASAI trial of over 105,000 women found AI-supported mammography raised cancer detection by 29% while cutting radiologist workload by 44% (The Lancet / Lancet Oncology). In research, DeepMind's AlphaFold has predicted over 200 million protein structures and won the 2024 Nobel Prize in Chemistry, while the first fully AI-designed drug candidate has reached mid-stage trials (Nature Medicine, 2025). Adoption is climbing fast: 66% of US physicians reported using AI in 2024, up from 38% in 2023 (AMA, 2025), largely for ambient documentation that trials link to lower burnout (NEJM AI, 2025). McKinsey estimates generative AI could help unlock up to $1 trillion of improvement potential in US healthcare (McKinsey, 2023). But risks — bias, hallucination, thin trial evidence and regulation under the EU AI Act and FDA — remain central, and none of this constitutes medical advice.

Key findings

What is AI in healthcare, and why now?

AI in healthcare is the application of machine-learning systems — including deep-learning image classifiers and, increasingly, large multi-modal and generative models — to clinical and operational tasks such as diagnosis, treatment planning, drug discovery, clinical documentation and hospital administration. The World Health Organization defines the newest wave as large multi-modal models (LMMs) that can accept text, images, signals or other inputs and generate diverse outputs, and it issued dedicated guidance on their use in health in January 2024 (WHO, 2024).

Three forces explain why adoption accelerated after 2023. First, capability: foundation models reached clinically useful accuracy on language and imaging tasks, with the Stanford HAI AI Index 2025 noting a 2024 wave of medical foundation models including Med-Gemini and specialty models for echocardiography and radiology. Second, regulation caught up — the FDA had authorized about 950 AI-enabled devices by August 2024, giving hospitals cleared products to deploy (Stanford HAI AI Index 2025). Third, economics: McKinsey estimates generative AI could help unlock up to $1 trillion of improvement potential in US healthcare, much of it in the roughly 25% of spending that is administrative (McKinsey, 2023).

The result is measurable uptake. The American Medical Association found 66% of US physicians used AI in their practice in 2024, up from 38% in 2023 — a 78% one-year increase — with the average number of AI applications per physician rising from 1.1 to 2.3 (AMA, 2025). Healthcare AI has moved from pilot to point-of-care.

How is AI used in diagnosis and medical imaging?

Medical imaging is by far the largest clinical application of AI, accounting for 76% of all FDA-authorized AI/ML devices — 769 of 1,016 in a 2025 peer-reviewed taxonomy — because deep-learning models excel at pattern recognition in radiographs, CT, MRI and mammograms (npj Digital Medicine, 2025). About 84% of these devices use images as their core input. Typical uses include triaging suspected strokes and pulmonary emboli, flagging lung nodules, and prioritising worklists so urgent scans reach radiologists first.

The strongest evidence comes from breast screening. Sweden's MASAI randomized trial of over 105,000 women found AI-supported mammography increased cancer detection by 29% versus standard double reading, with sensitivity of 80.5% against 73.8% at matched specificity and no increase in false positives (The Lancet, 2025). Its 2023 interim analysis in Lancet Oncology reported a 44% reduction in radiologists' screen-reading workload — a rare randomized demonstration that AI can raise detection and cut labour simultaneously.

Beyond imaging, the Stanford HAI AI Index 2026 reports experimental multi-agent diagnostic systems reaching 85.5% accuracy on complex published case studies. Such figures come from curated benchmarks, not routine practice, and do not imply an AI should diagnose patients unsupervised. The consistent regulatory model is AI as a decision-support aid to a clinician, not a replacement — and none of this is medical advice.

How is AI changing drug discovery?

AI is compressing the earliest and most failure-prone stages of drug discovery — target identification, protein structure prediction and molecule design. The landmark tool is DeepMind's AlphaFold, which has predicted the 3D structures of over 200 million proteins, nearly all known to science, freely accessible to more than two million researchers via a database built with EMBL-EBI (DeepMind / EMBL-EBI). The achievement earned Demis Hassabis and John Jumper a share of the 2024 Nobel Prize in Chemistry.

AlphaFold 3, published in Nature in May 2024, extended prediction to how proteins interact with DNA, RNA, ions and small-molecule drugs — the interactions that matter for designing a therapeutic — reporting substantially higher accuracy for protein-ligand interactions than conventional docking tools (Nature, 2024). This structural foundation feeds generative chemistry platforms that propose novel drug candidates.

Clinical proof is now emerging. Insilico Medicine's rentosertib (ISM001-055), described as the first drug with both its biological target and its molecule discovered by generative AI, reported positive Phase IIa results in 71 patients with idiopathic pulmonary fibrosis in Nature Medicine in June 2025 and advanced toward Phase III in 2026. Industry reviews count over 170 AI-originated programs in clinical development, with reported Phase I success rates of 80-90% against a historical 40-65% — though the pipeline has produced no approved drug yet, so late-stage evidence remains the open question.

What is ambient AI clinical documentation?

Ambient AI clinical documentation — often called an 'AI scribe' — uses speech recognition and large language models to listen to a patient visit and draft the clinical note automatically, targeting the paperwork that drives physician burnout. Commercial systems include Microsoft DAX Copilot, Nabla and Abridge, and voice-based documentation had reached 29% of US physicians by 2025 (AMA, 2025).

Randomized evidence is now available. A three-arm pragmatic trial of 238 physicians across 14 specialties and about 72,000 encounters (NEJM AI, 2025) found physicians using Nabla cut documentation time by roughly 10% versus usual care, and both AI arms showed about a 7% improvement in burnout scores. A separate Abridge evaluation found clinicians spent 8.5% less total time in the electronic health record, with a reduction of about 30 minutes of documentation per provider per day.

The gains are real but bounded. The NEJM AI trial noted AI-generated notes 'occasionally' contained clinically significant inaccuracies — most often omissions or pronoun errors — which is why every vendor and regulator stresses that a clinician must review and sign each note. The Stanford HAI AI Index 2026 cites reports of up to 83% less time spent writing notes in some deployments, illustrating both the upside and the wide variance across settings.

How is AI used for administration and hospital operations?

The largest near-term financial opportunity for AI in healthcare is administrative, not clinical. McKinsey estimates administrative activities account for roughly 25% of total US healthcare spending, and that generative AI could help unlock up to $1 trillion of unrealized improvement potential across the industry (McKinsey, 2023). Applications include prior-authorization processing, medical coding and billing, claims management, patient scheduling, call-centre support and summarising unstructured records.

McKinsey's sizing puts potential annual savings at $60-120 billion for hospitals (a 4-11% cost reduction) and $20-60 billion for physician groups (3-8%), driven by clinical-operations automation, patient-flow optimisation and workflow tools (McKinsey, 2023). Operationally, hospitals use predictive models to forecast admissions, reduce no-shows, optimise operating-theatre scheduling and manage staffing.

Adoption of these back-office tools is spreading faster than high-risk clinical AI precisely because the safety bar is lower: summarising research and standards of care became the single most common physician AI use case at 39% in 2025, up from 10% in 2024 (Stanford HAI AI Index 2026). Administrative AI still touches sensitive data, so privacy and accuracy controls remain essential — 86% of physicians in the AMA survey cited data privacy as key to wider adoption.

How widely is AI adopted, and how many devices are FDA-authorized?

Adoption and regulatory authorization are both rising steeply, but from different baselines. The FDA's authorized-device count grew from six in 2015 to about 950 by August 2024, with 258 more in 2025, while physician use roughly doubled in two years (Stanford HAI AI Index 2025-2026; AMA, 2025). The table below summarises the key adoption and authorization statistics with their sources.

Two patterns stand out. First, authorizations are concentrated in radiology and cleared through the 510(k) pathway, which relies on 'substantial equivalence' to existing devices rather than new trials. Second, clinician adoption is broad but shallow — most physicians use one or two tools, dominated by documentation and research-summary assistants rather than autonomous diagnostic AI.

What are the main AI use cases in healthcare?

AI in healthcare spans clinical, research and operational domains, and each leading use case now has a real system and a piece of evidence behind it. The table below maps the primary use cases to what they do and a concrete, sourced example — the fastest way to see where the technology has moved beyond hype.

The pattern is that AI performs narrow, well-defined tasks within a clinician-supervised workflow. Imaging triage, structure prediction and note drafting are mature; autonomous diagnosis and generative treatment planning remain experimental and are not represented among routinely deployed, trial-backed tools.

What are the risks of AI in healthcare, and how is it regulated?

The central risks of AI in healthcare are bias, error (including generative hallucination), thin clinical evidence, privacy and accountability — and they are why the WHO, FDA and EU AI Act all treat health AI as high-stakes. The WHO's 2024 guidance on large multi-modal models issued more than 40 recommendations and warned, via Chief Scientist Dr Jeremy Farrar, that generative AI helps 'only if... risks are fully accounted for' (WHO, 2024).

Evidence quality is a live concern: across 1,016 FDA AI authorizations, only a small share were supported by randomized-trial data, and about 96% used the 510(k) pathway that does not require new trials (npj Digital Medicine, 2025). Bias is documented — models trained on non-representative data can underperform for under-represented groups — and the number of publications on ethics in medical AI quadrupled from 288 in 2020 to 1,031 in 2024 (Stanford HAI AI Index 2025), reflecting how seriously the field now takes it.

On regulation, two regimes matter most. In the EU, AI systems used as medical devices or in a medical-device safety component are classified as high-risk under the EU AI Act, layering AI-specific obligations (risk management, data governance, human oversight, transparency) on top of existing medical-device rules. In the US, the FDA regulates AI as Software as a Medical Device and has introduced Predetermined Change Control Plans so models can be updated safely after clearance. The table below maps the main risks to mitigations. None of this content is medical advice.

Where do Europe, the Nordics and Sweden stand?

Europe combines some of the strongest clinical AI evidence with the strictest regulation. The pivotal MASAI trial — the first large randomized controlled trial of screening AI — was run in Sweden across over 105,000 women, and its finding of 29% higher cancer detection with 44% less radiologist workload is now a global reference point for imaging AI (The Lancet / Lancet Oncology). Sweden's national screening infrastructure made such a trial feasible where fragmented systems could not.

Regulation is the defining European feature. The EU AI Act classifies AI used as, or within, a medical device as high-risk, meaning developers must meet requirements for risk management, high-quality data, logging, transparency and human oversight before deployment across the single market — obligations that stack on top of the Medical Device Regulation. This raises the compliance bar relative to the US 510(k) route but is designed to close the evidence and safety gaps that the FDA data expose.

The Nordic context is favourable for careful adoption: high-quality health registries, universal coverage and strong institutional trust make both validation and deployment easier than in more fragmented markets. The constraint is capacity — running trials and meeting EU AI Act obligations demands resources and expertise that health systems are still building, so the Nordic edge lies in trustworthy evidence rather than raw deployment speed.

What is the outlook for AI in healthcare?

The near-term outlook is continued fast growth concentrated in imaging, documentation and administration, with slower, evidence-gated expansion into autonomous clinical decision-making. FDA authorizations are running at hundreds per year (258 in 2025) and physician adoption is doubling every roughly two years, so the direction of travel is not in doubt (Stanford HAI AI Index 2026; AMA, 2025).

Three variables will shape how far it goes. Evidence: the field needs more randomized trials like MASAI to justify high-risk clinical use, since most current devices are cleared without them (npj Digital Medicine, 2025). Regulation: the EU AI Act's high-risk regime and the FDA's evolving Software-as-a-Medical-Device framework will determine what reaches patients and how updates are governed. Economics: whether McKinsey's up-to-$1-trillion opportunity materialises depends on integration into workflows and billing, not just model quality (McKinsey, 2023).

The honest summary is that AI in healthcare has crossed from promise to partial proof. Imaging triage, protein prediction and ambient documentation are backed by real data; autonomous diagnosis and AI-designed drugs are advancing but unproven at scale, with no AI-discovered drug yet approved. As the WHO stresses, the benefits are conditional on managing bias, error and accountability — and nothing here is a substitute for professional medical advice.

Scoreboard (machine-readable data)

Every headline indicator with its value, period, source and confidence. Free to reuse under CC BY 4.0.

↓ CSV · ↓ JSON

Methodology & verification

This report synthesises primary and peer-reviewed sources on AI in healthcare and attributes every statistic to a named source with its year and scope. Regulatory device counts come from the US FDA and the Stanford HAI AI Index (2025 and 2026); the specialty breakdown and pathway data come from a peer-reviewed taxonomy of 1,016 authorizations in npj Digital Medicine (2025). Clinical-outcome figures are drawn from randomized controlled trials where available — MASAI (The Lancet / Lancet Oncology) for mammography and a three-arm trial (NEJM AI, 2025) for ambient documentation — and benchmark or preclinical figures are labelled as such. Drug-discovery claims cite DeepMind/Nature (AlphaFold) and Nature Medicine (rentosertib). Economic estimates are McKinsey's. We deliberately distinguish trial-backed results from benchmark accuracy and from market projections, and we do not present any figure as clinical guidance. Where curl verification returned bot-block responses (The Lancet, McKinsey), URLs were confirmed as live via independent search. Figures were checked against source publications as of 30 July 2026. This report is not medical advice.

Data dictionary

Frequently asked questions

Glossary

Software as a Medical Device (SaMD)

Software intended for a medical purpose that performs that purpose without being part of a hardware device; the FDA category under which most healthcare AI is regulated. About 96% of FDA AI clearances used the 510(k) pathway. ↗

Large multi-modal model (LMM)

An AI model that can accept multiple data types (text, images, signals) and generate diverse outputs; the WHO's 2024 health-AI guidance focuses on LMMs and their benefits and risks in medicine. ↗

AlphaFold

DeepMind's AI system for predicting protein 3D structure, which has mapped over 200 million proteins and won the 2024 Nobel Prize in Chemistry; AlphaFold 3 extends prediction to molecular interactions used in drug design. ↗

Ambient clinical documentation

AI that listens to a clinical encounter and drafts the medical note automatically ('AI scribe'); randomized evidence links it to lower documentation time and burnout, with mandatory clinician review of outputs. ↗

Sources

• Stanford HAI — 2026 AI Index Report, Medicine chapter (and 2025 Science & Medicine)
• npj Digital Medicine — How AI is used in FDA-authorized medical devices: a taxonomy across 1,016 authorizations (2025)
• The Lancet / Lancet Oncology — MASAI randomized trial of AI-supported mammography screening
• Google DeepMind / Nature — AlphaFold and AlphaFold 3 (Nature 630:493-500, 2024)
• WHO — Ethics and governance of AI for health: guidance on large multi-modal models (2024)
• AMA — 2 in 3 physicians are using health AI, up 78% from 2023 (2025)
• McKinsey — Tackling healthcare's biggest burdens with generative AI (2023)

Related reports & data

• AI and Jobs
• The EU AI Act, explained
• National AI Power Index