AI in Healthcare: What's Real in 2026

Where AI is actually used in medicine today

The clearest picture comes from what regulators have authorized. Through the end of June 2026, the US FDA had authorized 1,614 AI-enabled medical devices — and 1,230 of them (76%) are radiology devices, mostly tools that flag suspicious findings in scans^[1]. Cardiovascular devices are the next largest group^[2]. Beyond imaging, the fastest-growing uses are administrative: AI documentation assistants that turn clinician-patient conversations into notes, and workflow tools. The pattern across all of it: AI handles repetitive, well-bounded tasks — screening, measurement, drafting — while humans keep the decisions.

What "FDA authorized" does — and doesn't — mean

Authorization is not the same as proof of patient benefit. Of the 1,247 AI devices on the FDA list as of July 2025, 96% were cleared via the 510(k) pathway — a showing of "substantial equivalence" to an existing device, not a new clinical trial^[2]. That matters, because cleared technology can still perform poorly in the real world (see our radiology guide for the sepsis-model cautionary tale).

Regulation is also still being built. The FDA finalized its Predetermined Change Control Plans (PCCP) guidance in December 2024, letting manufacturers pre-specify planned AI model updates, and issued draft lifecycle-management guidance in January 2025^[3]. Notably, as of the July 2025 list update, no authorized device used large language model (LLM) technology — the chatbot-class AI behind ChatGPT has not crossed the medical-device line^[4]. In the UK, the MHRA applies the same logic: whether software is a medical device depends on the manufacturer's declared intended purpose. General chatbots are not medical devices; software claiming to diagnose or treat is regulated^[5].

Drug discovery: real progress, long timelines

The 2024 Nobel Prize in Chemistry went, in part, to Demis Hassabis and John Jumper for AlphaFold2, which predicts protein structures — structures for "virtually all the 200 million proteins" researchers have identified, used by more than two million people across 190 countries^[6]. It is a landmark research tool, not a therapy: predicting a protein's shape is one early step in a drug pipeline that still takes years.

The most advanced clinical result so far: in June 2025, Nature Medicine published a phase IIa trial of rentosertib (ISM001-055), a lung-fibrosis drug whose target and molecule were designed with generative AI. In 71 patients over 12 weeks, the 60 mg daily group showed a mean FVC change of +98.4 mL versus −20.3 mL on placebo^[7]. That is an encouraging early signal — and the investigators themselves stress that larger, longer trials are needed. Phase IIa is not approval, and most candidates still fail.

Will AI replace doctors?

No credible evidence points that way. Even in the most favorable screening trials, at least one human clinician remains in the loop, and the authors of the largest randomized mammography trial wrote explicitly that their results "do not support replacing healthcare professionals with AI"^[8]. The realistic direction is augmentation: AI triages routine work and flags findings, while clinicians keep judgment, communication and responsibility. For the trial evidence behind that conclusion, see our AI radiology guide.

Frequently asked questions