Coedee Health & Policy
The Line That Just Moved
For most of AI’s history in medicine, a human always had the final say. That line is now being redrawn, clearance by clearance, in ways most patients haven’t noticed yet.
There’s a quiet distinction in healthcare AI that most headlines skip past: the difference between “assistive” and “autonomous.” An assistive AI system flags a suspicious spot on a scan and hands the decision to a doctor. An autonomous system makes the call itself. For years, regulators kept nearly every AI tool firmly on the assistive side of that line. That’s changing, gradually and then suddenly, and this year is when the shift became impossible to ignore.
A Regulatory Milestone Nobody’s Grandmother Has Heard Of
The U.S. Food and Drug Administration now has more than 1,500 AI-enabled medical devices on its authorized list — a number that has grown explosively from fewer than 30 total AI-enabled clearances through 2015. The vast majority still move through the FDA’s 510(k) pathway, the standard route for devices that resemble something already on the market, rather than the stricter De Novo or full Premarket Approval tracks reserved for genuinely novel technology.
But a small, growing subset of those clearances belong to a different category entirely: autonomous diagnostic systems that don’t require a clinician to review the AI’s output before it reaches a patient. The landmark case is a retinal-screening tool, originally cleared under the name IDx-DR and now marketed as LumineticsCore, which became the first FDA-authorized fully autonomous AI diagnostic system back in 2018. It analyzes retinal photographs and identifies diabetic retinopathy — a leading cause of preventable blindness — without a physician reviewing the image first. In its premarket clinical study of 900 patients, it achieved 87.4% sensitivity and 89.5% specificity for detecting disease severe enough to require referral, a performance bar high enough that primary care offices and pharmacies can now run the screening directly, without first sending a patient to a specialist.
The Next Frontier: AI That Talks Back
The more recent milestone is arguably a bigger deal, and it happened quietly. In late June, clinical AI company UpDoc announced it had received FDA clearance for what’s being described as the first authorized Software as a Medical Device incorporating a patient-facing large language model — meaning the AI doesn’t just analyze an image behind the scenes, it directly converses with patients as part of a cleared medical function. Legal analysts tracking the clearance describe it as establishing a genuinely new regulatory pathway, one that future clinical AI companies incorporating conversational LLMs into patient-facing products will likely study closely as a template.
Even with this clearance, the operative expectation remains “human-in-the-loop.” Clinical AI, for now, is expected to function as a support tool for clinicians rather than an independent decision-maker — a limit rooted as much in state-level practice-of-medicine law as in FDA policy itself.
Why Regulators Are Moving Carefully — and Why That’s the Right Call
The caution isn’t bureaucratic foot-dragging; it reflects genuine unresolved questions that the industry hasn’t answered yet. A recent survey of physicians found that more than 80% now use AI professionally in some capacity, more than double the adoption rate from just a few years earlier — but the same survey found that roughly 86 to 88% of those physicians still flag safety validation and data privacy as active concerns, and ranked clear liability frameworks as a top regulatory priority. That tension — rapid, enthusiastic adoption paired with persistent structural worry — is precisely the environment regulators are trying to write rules for in real time, rather than after the fact.
Liability is the sharpest edge of that debate. If an autonomous diagnostic tool misses a cancer that a human radiologist would have caught, current frameworks still mostly assume there’s a licensed clinician somewhere in the chain who bears responsibility. Fully autonomous systems complicate that assumption in ways the legal system hasn’t fully worked through — a gap researchers studying the issue have flagged as one of the central unresolved questions facing the entire category, alongside unresolved questions of accountability when a state-level pilot program, such as one recently announced testing an AI system for medication prescribing, doesn’t have a clear predicate to fall back on.
The Global Deadline Nobody’s Talking About
While the U.S. debate plays out clearance by clearance, Europe is moving on a fixed calendar. Most of the high-risk AI obligations under the EU AI Act take effect this month — August 2026 — with full compliance for medical device AI required by August 2027. The bloc’s Medical Device Coordination Group has already published guidance clarifying how the AI Act interacts with existing medical device regulation, and the FDA’s own draft guidance, issued last year, pushed manufacturers toward new transparency requirements: a clear disclosure that a device uses AI, details on what data goes in and comes out, documented performance measures, and disclosure of known sources of bias. The agency has also encouraged the use of standardized “model cards” — essentially structured spec sheets for how a given AI model was built and tested — as these tools proliferate.
What This Actually Changes for Patients
The practical upshot, right now, is narrower than the headlines about “AI doctors” might suggest — but it’s real and it’s already reaching underserved communities. Walking into a rural pharmacy or a primary care office and getting an on-the-spot AI-driven retinal screen or skin assessment, with a result in minutes rather than a weeks-long wait for a specialist referral, is no longer a hypothetical. It’s deployed, cleared technology operating today in exactly the kind of access-limited settings where the wait for a human specialist has historically been the barrier to catching disease early.
What hasn’t arrived — and, based on both the regulatory posture and the clinician sentiment data, isn’t arriving imminently — is a world where an AI system independently decides your treatment plan without a licensed professional in the loop. The FDA’s own guidance is explicit that AI-enabled devices exist on “a continuum of decision-making roles,” and the agency has so far stopped short of authorizing systems that don’t ultimately rely on a human to interpret outputs and make the final clinical call, autonomous diagnostics like LumineticsCore being a narrow, carefully validated exception rather than the new norm.
The Bigger Picture
What’s happening in medical AI regulation right now is a genuinely useful preview of how autonomy questions will get resolved across every high-stakes industry AI touches, from finance to transportation to legal services. The pattern is consistent: narrow, well-validated autonomous capability gets approved first, in the specific contexts where the evidence is overwhelming and the failure mode is well understood, while broader autonomous decision-making stays gated behind human oversight until liability, transparency, and bias questions get real answers rather than good intentions. Healthcare, with its unusually rigorous evidence bar and its unusually high stakes, is simply the industry working through that process in public first — and everyone else building AI into consequential decisions should be watching closely.
The Prescribing Question Regulators Haven’t Answered Yet
One area still sitting well outside the current comfort zone is autonomous prescribing — AI systems that would not just flag a diagnosis but independently decide on and issue medication. As of the most recent tracking, no autonomous AI prescription service has received FDA clearance. Legislative interest is building regardless: a bill introduced in Congress would, for the first time, explicitly allow an AI system to qualify as a “practitioner” eligible to prescribe drugs, provided it’s authorized by a state or by the FDA. Utah has separately announced a pilot program working with a clinical AI company to explore algorithmic prescribing recommendations, a move that has already raised pointed questions from clinicians about liability, oversight, and what happens when an algorithm’s recommendation and a doctor’s judgment diverge.
Researchers studying the ethics of autonomous prescribing generally don’t argue the concept is inherently reckless — most agree it could work, provided systems are demonstrably effective, safe, and genuinely serve patient interests. The sticking point is that accountability and liability frameworks for that scenario simply don’t exist yet in any settled form, which is exactly why no regulator has been willing to be first to clear a fully autonomous prescribing product, even as adjacent categories like autonomous diagnostics continue to expand.
What to Watch Over the Next Year
Three developments are worth tracking closely for anyone following this space. First, whether the FDA finalizes its total-product-lifecycle guidance for AI devices, which would set ongoing monitoring requirements for models that continue learning and changing after their initial clearance — a meaningfully different challenge than approving a static piece of software once. Second, how the EU’s August 2026 high-risk obligations play out in practice for device makers who sell into both U.S. and European markets simultaneously, since compliance burdens that diverge across the two systems will shape where companies choose to seek clearance first. Third, whether any state moves forward on legislation like the federal prescribing bill before the FDA itself has settled the underlying safety questions — a sequencing mismatch that could create real regulatory confusion if it happens.