The Question

AI diagnostic interface displaying a medical scan with highlighted regions of concern alongside clinical data

Consider a patient with diabetes in a place with no eye specialist. Diabetic retinopathy is a leading cause of blindness, and preventable if caught early. A camera in a primary care clinic and an AI system called IDx-DR (now LumineticsCore) can screen for it without a specialist reading the image. In its pivotal trial of 900 patients, it detected more-than-mild diabetic retinopathy with 87.2% sensitivity and 90.7% specificity.

IDx-DR became the first autonomous AI diagnostic system authorised by the FDA — meaning no clinician needs to interpret the result — in April 2018. Since then, the FDA's list of AI-enabled medical devices has grown to more than 1,400 by the end of 2025, about three-quarters of them in radiology. The question is no longer whether AI can perform well on narrow diagnostic tasks — peer-reviewed studies show it can. It is whether it will become the main diagnostician, and what that means for doctors and patients.

What the Evidence Shows

The published performance data is worth stating precisely. In September 2021, the FDA authorised Paige Prostate, the first AI software to flag areas of a prostate biopsy image most likely to contain cancer. In the study behind the authorisation, 16 pathologists read 527 slide images with and without the tool; with it, they detected cancer on individual slides 7.3% more often, with no effect on readings of benign slides. The tool was authorised as an aid to pathologists, not a replacement.

The cardiology findings are equally striking. Apple Watch's ECG feature received FDA clearance in 2018 to help detect atrial fibrillation. In a 2019 Mayo Clinic study in Nature Medicine, AI analysis of standard 12-lead ECGs from 52,870 patients identified weak heart pumping (left ventricular dysfunction) with an area under the curve of 0.93 — a condition that often goes undetected until symptoms appear. Among patients without the condition, those flagged by the AI had four times the risk of developing it later.

"In this trial, the availability of an LLM to physicians as a diagnostic aid did not significantly improve clinical reasoning compared with conventional resources."

— Goh E et al., JAMA Network Open, 2024

Google DeepMind's AlphaFold represents a different category of achievement: fundamental biology. In 2020, AlphaFold 2 largely solved the problem of predicting protein structures, and DeepMind's database now offers free access to more than 200 million predicted structures. Its creators, Demis Hassabis and John Jumper, shared the 2024 Nobel Prize in Chemistry. It speeds up research, but it does not diagnose patients.

AI is already better than doctors at some narrow tasks. Being better at medicine is a much bigger claim.

Why This Is Happening

The global health worker shortage is severe. The World Health Organization projects a shortfall of about 10 million health workers by 2030, concentrated in low- and lower-middle-income countries. Training more doctors takes a decade, and wealthy countries keep attracting doctors from poorer ones. AI symptom checkers such as Ada Health are used by millions of people. But the sector has had setbacks: Babylon Health, once valued at $4.2 billion for its AI triage, collapsed into bankruptcy in 2023 amid doubts about its technology.

AI can learn from more images than any human sees in a career. In a 2018 Stanford study in PLOS Medicine, an algorithm called CheXNeXt, trained on more than 100,000 chest X-rays, performed as well as radiologists on 10 of 14 conditions, better on one, and worse on three. It does not tire or get distracted. But real-world performance often falls short of study results, and a 2024 randomised trial in JAMA Network Open found that giving doctors access to a large language model did not significantly improve their diagnostic reasoning — even though the model alone scored higher than the doctors.

Regulatory frameworks are adapting. The FDA has authorised more than 1,400 AI-enabled devices, mostly through its 510(k) and De Novo pathways. In December 2024 it finalised guidance on "predetermined change control plans", which let manufacturers update AI systems in pre-agreed ways without a full new submission. But almost all authorised tools are designed to assist clinicians, not to replace them.


What Could Happen

AI as primary diagnostician in most specialties by 2035 Possible · 35%

AI becomes the standard first-contact diagnostician in image-based fields such as radiology, pathology, dermatology, and ophthalmology. Human specialists shift toward oversight, complex cases, and the relational side of medicine. In low-income countries, AI screening helps close the specialist gap for common conditions such as diabetic eye disease and tuberculosis.

Liability frameworks block deployment in wealthy countries while low-income countries adopt freely Most likely · 45%

US and EU liability law poses an unresolved question: when an AI diagnostic error harms a patient, who is liable — the physician, the hospital, or the company that built the system? Health systems in wealthy countries adopt a cautious "AI-assist only" posture requiring human confirmation, partially negating the AI's advantage. Meanwhile Rwanda, India, and Southeast Asia — with no equivalent liability culture — deploy autonomous AI diagnostics at scale, a paradox in which the world's poorest patients receive the most advanced diagnostic care.

AI over-diagnosis creates a new medical crisis Less likely · 20%

AI systems optimised for sensitivity over specificity generate high false-positive rates, producing a wave of unnecessary biopsies, surgeries, and anxiety that overwhelms the systems AI was supposed to relieve. Regulators raise the specificity bar, slowing deployment. Medicine settles on a cautious model — AI handles screening and triage, humans confirm anything triggering invasive follow-up — which proves sustainable but falls short of the transformative access gains predicted.

Our Assessment
We assign 35% probability — possible that AI becomes the primary diagnostic tool in at least three major medical specialties within 15 years. We have lowered this forecast from 79%. Our earlier version cited study results, figures, and a New England Journal of Medicine quote we could not verify, and it overstated how well AI performs. The real evidence is strong for narrow tasks, but authorised tools are almost all assistants, randomised trials show mixed results in practice, and liability, payment, and trust remain unresolved. The likeliest outcome is AI as a powerful assistant, with humans still making most diagnoses.

What Can We Do

Rural health worker using a smartphone diagnostic tool to screen a patient in a low-resource setting

The AI medicine transition will happen with or without public engagement — but its shape, fairness, and safety are being decided now in regulatory agencies and legislatures.

Demand that AI diagnostic tools address health equity, not just affluent markets. Commercial incentives push toward wealthy markets with high reimbursement, while the greatest need is in low-income countries. Supporting public and philanthropic efforts to deploy and test AI screening where specialists are scarce is where individual advocacy has real leverage.

Push for mandatory bias audits on AI medical systems before deployment. Systems trained mostly on data from some groups can perform worse for others. A 2019 study in Science found that a widely used algorithm for allocating extra care underestimated the needs of Black patients because it used health spending as a stand-in for illness; fixing the bias would have raised the share of Black patients flagged for extra help from 17.7% to 46.5%. Performance broken down by demographic group should be a condition of approval.

Engage with the liability debate before courts set the precedent. If legislatures do not act, the rules on who is responsible when AI gets a diagnosis wrong will be set case by case in the courts. Patients, not just physicians and technology companies, deserve a voice in that debate.

Use what is already available — carefully. AI symptom checkers and chatbots can help you prepare questions and understand terms before an appointment. They are not replacements for clinical care, and they can be confidently wrong, so check anything important with a clinician.