The claim making the rounds this year is tidy and wrong: that the Food and Drug Administration created a new regulatory class for autonomous diagnostic AI, a pathway for models that diagnose you outright while the specialist who used to sign off gets reassigned. There is no such class. There is no such pathway. What the FDA actually did on January 6, 2026 was quietly rewrite a guidance document about software that helps doctors think — and, in the same breath, reaffirm that software which looks at your medical images and tells you what is wrong with you remains a regulated medical device, full stop.
The document is the revised Clinical Decision Support Software Guidance, which supersedes the 2022 version. FDA Commissioner Marty Makary framed it as intended to, in his words, cut unnecessary regulation and promote innovation. The mechanics are narrower than the framing. Under Section 520(o)(1)(E) of the Federal Food, Drug, and Cosmetic Act — added by Section 3060 of the 21st Century Cures Act in 2016 — software escapes device regulation only if it clears four criteria, including that it does not analyze a medical image or signal, and that a clinician can independently review the basis for its recommendation rather than relying primarily on it.
The 2026 revision loosens exactly one meaningful bolt. FDA will now exercise enforcement discretion for tools that return a single recommendation where only one option is clinically appropriate, instead of demanding a menu of options for the clinician to weigh. The 2022 guidance had read the statute's plural word recommendations as a hard limit, which developers found unworkable in practice. FDA did not change its reading of the statute; it just agreed not to enforce it in that scenario. Analysts at Covington & Burling, who tracked the change closely, concluded it does not represent a fundamental deregulatory shift. Notably, the guidance is silent on AI itself, despite being announced as an AI action, and was issued as revised final guidance with no comment period.
The part that did not move
The American College of Radiology's read on the revision, published January 14, is the clarifying one: the 2026 update maintains that any function analyzing medical images to generate diagnostic recommendations is subject to FDA oversight, and it underscores that online AI services offering image interpretations directly to consumers generally fall within FDA jurisdiction. In other words, the exact capability at the center of the autonomous-diagnosis story is the capability FDA explicitly refused to deregulate.
Real autonomous diagnosis in the United States runs on a pathway that has existed since 2018, and it is small. IDx-DR — now marketed as LumineticsCore by Digital Diagnostics — received De Novo classification DEN180001, authorizing it to diagnose more-than-mild diabetic retinopathy from fundus images with no human reviewer. Its preregistered prospective trial enrolled 900 patients screened in primary care by operators with no prior imaging experience, and reported 87.2% sensitivity, 90.7% specificity, and 96.1% gradability against a level-1 reference standard that included OCT imaging. Eyenuk's EyeArt, cleared in 2020, reported 96% sensitivity and 88% specificity for more-than-mild retinopathy and 92% and 94% for vision-threatening disease. AEYE Health's AEYE-DS is the third. All three diagnose one condition, in one organ.
Melanoma is not on that list. DermaSensor received De Novo classification in 2024 as a prescription Class II device — and it is explicitly adjunctive, intended for physicians who are not dermatologists. Its pivotal DERM-ASSESS III data show 95.5% sensitivity across skin cancers with a 98.1% negative predictive value, but specificity between 20.7% and 32.5%, and 87.5% sensitivity for melanoma specifically. A device that flags roughly three in four benign lesions as suspicious is a triage instrument, not a replacement pathologist.
Reimbursement is the one place the autonomous premise holds up. CPT 92229, effective January 1, 2021, pays for point-of-care autonomous retinal imaging with automated analysis and report — explicitly without physician interpretation. CMS finalized a national rate around $50, though contractor pricing varies sharply: Cigna Government Services set $29.45 in Ohio and $28.42 in Kentucky, drawing a public objection from the American Academy of Ophthalmology.
Analysis: what nobody has answered
Autonomous clearance breaks the liability chain that malpractice law assumes. When LumineticsCore returns a negative and the patient goes blind, there is no clinician whose judgment can be evaluated against a standard of care, because the standard of care was the device. Digital Diagnostics has said it assumes liability for its own output — a corporate promise, not a legal regime. Product liability and medical malpractice are different doctrines with different discovery rules and different insurers, and no court has cleanly resolved which governs an unreviewed algorithmic diagnosis.
The equity case cuts both directions. CPT 92229 genuinely moved retinal screening into federally qualified health centers and pharmacies serving patients who would never reach an ophthalmologist. But retinal AI performance varies with fundus pigmentation, camera hardware, and pupil dilation, and gradability rates — that 96.1% figure — degrade fastest in exactly the low-resource settings the code was designed to reach. An ungradable image in a clinic with no referral pathway is a screening failure that never enters the sensitivity denominator.
The strongest clinician objection came from the AMA House of Delegates in June 2026, which adopted policy declaring AI an assistive tool and not an autonomous decision-maker. AMA CEO John Whyte, MD, MPH, put it directly: AI has enormous potential in healthcare, but it cannot replace physician judgment. On transparency he was blunter — AI should never function as an unaccountable black box. The AMA is now pushing for regular audits of AI clinical review tools, triggered by changes to models, training data, or guidelines.
What to watch: FDA plans to revise its Policy for Device Software Functions and Mobile Medical Applications in FY26, and that document — not the CDS guidance — is where the boundary between a symptom checker and a regulated diagnostic will get drawn. Watch also whether any autonomous device outside ophthalmology reaches De Novo. Eight years after DEN180001, none has.
“AI has enormous potential in healthcare, but it cannot replace physician judgment.”— John Whyte, MD, MPH, CEO, American Medical Association