The AI Doctor Debate Is Really a Capacity Story

The question is narrower than “AI doctors”

Nature’s June 3 piece on whether “AI doctors” are becoming good enough frames a real shift, but the headline version is too broad. The evidence does not show general-purpose autonomous doctors ready to replace clinicians. It shows AI moving from benchmark demos into supervised workflow tests: taking histories, drafting differentials, checking treatment plans, reviewing charts, and operating in sandboxed electronic health-record environments.

That distinction matters. Medicine is not just naming a diagnosis from text. It includes physical examination, procedures, follow-up, cost-sensitive management, liability, patient trust, and the ability to know when the available data are bad. Current AI systems are being tested around pieces of that work, not the whole job.

The access case is real

The reason this debate matters is not technological novelty. It is capacity. The Association of American Medical Colleges projects a U.S. physician shortage of up to 86,000 by 2036, alongside sharp growth in older populations that use more care. The bottleneck is also clinician attention: the AMA reported that physicians averaged a 57.8-hour workweek in 2024, including substantial indirect care and administrative time, with 43.2% reporting at least one burnout symptom.

If AI can safely absorb even part of history-taking, documentation, chart review, triage, inbox routing, or error checking, it could expand useful clinical capacity without waiting a decade to train more physicians. That is a more practical claim than the robot-doctor story. It also sets a lower but still meaningful bar: the tool does not have to be autonomous to matter. It has to reduce avoidable work, reduce errors, or help scarce clinicians spend time on the cases that need them most.

The evidence points to supervised scaling

The strongest recent studies are not proof of safe autonomous deployment. They are evidence that some clinical tasks are becoming more machine-assistable.

A June 17 Nature paper on MIRA tested an autonomous medical AI agent in a sandboxed EHR using more than 500 MIMIC-IV emergency-department cases, 11 tools, and more than 85,000 ordering and intervention options. The authors reported physician-level or better performance on diagnosis and treatment quality, but the caveats are central: sandbox conditions, possible benchmark overlap, and a simulated workflow are not the same as live clinical accountability.

Google Research’s AMIE feasibility study, posted to arXiv, tested a conversational diagnostic AI with 100 adult urgent-care patients before appointments. No real-time human safety supervisor had to stop a consultation. AMIE’s differential included the final diagnosis in 90% of cases, with 75% top-three accuracy. But primary-care physicians did better on the practicality and cost-effectiveness of management plans. That is exactly the boundary: useful pre-visit work, not full clinical replacement.

The OpenAI/Penda Health primary-care study in Nairobi, also on arXiv, is closer to the near-term model. Across 39,849 visits at 15 clinics, clinicians using AI Consult had fewer diagnostic and treatment errors under independent physician review. The system was designed as a safety net that preserved clinician autonomy. That is less dramatic than an AI doctor. It is also more plausible.

Regulation shows where adoption really is

The FDA’s public list of AI-enabled medical devices shows medical AI is already being institutionalized, but mostly in device-like workflows. A July 2025 FDA Law Blog analysis counted 1,247 listed AI-enabled devices, with 77% authorized through the radiology panel and 96% through 510(k) clearance. That does not describe broad autonomous medicine. It describes concentrated deployment where data are digital, outcomes are easier to measure, and regulation fits existing device pathways.

The unresolved questions are practical, not philosophical: whether results generalize across hospitals, languages, EHR systems, and patient populations; whether tools improve hard outcomes rather than intermediate metrics; who is liable when AI advice is wrong; and whether systems reduce workload or create more review burden. The useful reading is skeptical but not dismissive. AI is not becoming a doctor. Some parts of doctoring are becoming scalable software workflows.