A JAMA piece says "doctor + AI" is no longer the gold standard — and the AMA published a framework the same week saying the opposite
August 23, 2026 · 4 items
A JAMA piece says "doctor + AI" is no longer the gold standard — and the AMA published a framework the same week saying the opposite
Fierce Healthcare · Heather Landi · August 21, 2026Practice operations
The claim: a JAMA Perspective published August 17 argues that medicine is approaching the point at which AI alone will exceed both physicians and physician-AI hybrids at five fundamental cognitive medical tasks — gathering patient information, diagnosing, selecting tests, recommending treatment, and managing chronic disease. The authors reviewed all published articles on AI in medicine since January 1, 2024. Byline: Ezekiel J. Emanuel, MD, PhD (oncologist, bioethicist, former White House advisor), Neal Khosla (CEO, Curai Health), Vinod Khosla (Khosla Ventures), and researcher Abe Baker-Butler.
The mechanism they name, which is the uncomfortable part: it isn't just that AI gets good — it's that human oversight can introduce errors and degrade the performance of a superior system. Emanuel's own summary: "Physicians will not disappear. Surgery, deliveries, and other physical procedures still need human hands. But 'doctor plus AI' as the automatic gold standard of care no longer holds up against the evidence." Their timeline: deployable for real-world cognitive tasks in some, maybe many workflows by 2030.
The rebuttal, from the AMA's own CEO: John Whyte — "Zeke may be right about the tasks. But medicine is not a collection of tasks… It's why, even with autopilot, most of us would not board a plane without a pilot in the cockpit. Performing medical tasks well is not the same as practicing medicine well." He notes much of the cited evidence comes from simulated tasks, not real-world clinical encounters — a limit the JAMA authors themselves concede, along with liability, regulatory and reimbursement barriers that block AI-alone studies.
The same week, the counter-move: the AMA and the Digital Medicine Society (DiME) released a framework on the physician's role in the digital and AI era — "Individual tasks will evolve as technology advances, but the physician's enduring responsibilities will remain remarkably stable." The two documents agree on what must change (workflows, medical education, reimbursement, regulation, liability) and diverge sharply on who ends up in the chair.
Why it matters for an independent practice: This is the argument that will be used against your clients, by people with capital behind it — one of the four authors runs a fund and another runs an AI clinic company, which is worth saying out loud. But the strategic read for an independent regen or cash-pay practice is almost the reverse of the scare headline. Emanuel is careful to exempt procedures and physical work — which is exactly what an orthobiologics practice sells. If cognitive tasks commoditize toward AI, the defensible ground is the thing AI can't do: the injection, the exam, the judgment call made in the room, and the relationship that got the patient there. For MMR that reframes the whole positioning conversation — stop marketing a practice on information (which is being given away free by an answer engine) and market it on hands, access, and continuity. And Whyte's autopilot line is the single best piece of patient-facing language produced this month; it belongs in a client's FAQ. Fierce Healthcare — As AI advances, a debate grows over the future role of physicians · AMA + DiME — Defining the physician's role in the digital and AI era (Trusted)
"Was PHI used to train it?" is the wrong question — and asking it is both approving dangerous AI and blocking safe AI
Fierce Healthcare · Peter Grantcharov & David Knobel, Industry Voices · August 21, 2026Practice operations
The failure, stated plainly: hospital AI committees start and effectively end their review with one question — was protected health information used to train this model? The authors argue that question describes the development process, not the model's exposure risk, and a single checklist applied to every model fails in both directions at once.
Failure one — false confidence on generative models. A committee confirms a BAA is signed, data is encrypted and access is logged, then never examines the pathways that actually matter: whether users can enter arbitrary prompts, whether the model can be queried repeatedly and adaptively, whether it retrieves from live clinical records, whether prompts and outputs are retained, and whether it can reproduce information from its training or retrieval context.
Failure two — blocking the models that reduce exposure. A closed classifier with no prompt interface, no generative capability, no exposed weights and no output beyond a label or a redaction mask gets the same review as an open-ended LLM. Their example bites: a face-detection or out-of-body-detection model must train on identifiable frames to redact accurately. Block that, and you preserve the manual workflow in which more people view, handle and retain identifiable data for longer. As they put it, "Demanding proof of zero theoretical risk is not rigorous risk management. It is risk avoidance untethered from probability." They anchor it in HIPAA's own Expert Determination standard, which asks that re-identification risk be "very small" — not mathematically impossible.
The five questions they say every review should open with: (1) What can the model accept and produce? (2) Can a user probe it? (3) What is actually accessible — final outputs, or also confidence scores, embeddings, weights, unrestricted APIs? (4) What is the credible path to a patient? — "it may be theoretically possible" is not a complete analysis. (5) What risk is created by saying no? They flag the fifth as the one routinely skipped. Authors: Grantcharov is CTO of Aimbient (formerly Surgical Safety Technologies); Knobel is a partner at Hopp & Partners — both have a commercial interest in narrow clinical-video AI, so read pillar two with that in mind.
Why it matters for an independent practice: A solo or small practice doesn't have an AI committee — the doctor is the committee, usually deciding in a vendor demo. That makes these five questions more useful here than in a health system, because they're short enough to actually use. Two direct applications. First, they upgrade the vendor screen we've been building all month: the FDA's autonomy × consequence grid (Aug 22) tells you how much the tool decides; these five tell you what could leak — together that's a one-page diligence sheet MMR could hand a client this week. Second, question five is the one that protects the practice from itself: refusing an AI intake or redaction tool is not the safe default if the alternative is a front-desk workflow where five people handle the same chart by hand. And note the asymmetry that should make everyone cautious — a promptable model wired into your records is the high-risk case, and it is also the one being sold hardest. Fierce Healthcare — Hospital AI committees are asking the wrong privacy question
Google will now let you put a button on your own site that makes you a reader's Preferred Source — and Preferred Sources get a badge inside AI Overviews
Search Engine Journal · Matt G. Southern · August 20, 2026Patient acquisition
What shipped: websites can now embed an interactive Preferred Sources button on their own pages. Clicking it adds the site as a Preferred Source and returns the reader to where they left off on your page. The old badge did the opposite — it pushed readers into Google's source-preferences tool and left them there. Button code is in Google's Search Central documentation.
Why the badge matters beyond Top Stories: per Google's documentation, sites selected as Preferred Sources are more likely to appear in Top Stories — and in AI Overviews and AI Mode, selected sites can be highlighted with a "preferred" badge. That is a rare case of a publisher-controlled signal reaching inside an AI answer.
The adoption number: Google says more than 600,000 unique sources have been selected through Preferred Sources so far — up from more than 345,000 in May. Roughly a 74% increase in about three months.
Two more changes in the same announcement: Discover will let users tell Google in their own words what they want more or less of via the three-dot menu on any card (rolling out "in the coming days," no confirmed date), and the Google News Android app now lets users customize daily audio briefings by topic, with each story attributed and linked back to the full article.
Why it matters for an independent practice: Almost everything we've tracked this month has been about visibility being decided for the practice — Grounding Drift, generative UI eating tool pages, citation shares collapsing overnight. This is the one lever that runs the other way: a signal the patient sets and the practice can ask for directly. Concrete moves this week. First, if a client site already carries the old Preferred Sources badge, check whether it needs a fresh embed — the new button is a different thing and the old one leaves readers stranded in Google's settings. Second, the honest bar: this only pays off where a practice is genuinely publishing — a physician-authored blog, a monthly patient-education post, a newsletter. For a five-page brochure site there's nothing for a reader to prefer. That makes this a decent forcing function in a client conversation: the practices already doing the content work now have a compounding mechanism, and the ones that aren't just got a concrete reason to start. Third, the place to put the button is at the end of the article a patient just found useful, not in the footer. Search Engine Journal — Google Expands Personalization Across Search, Discover & News · Google Search Central — Preferred Sources button code
ChatGPT ads switch on in 31 more countries today — and OpenAI quietly stopped blocking ads near medical advice back in April
Search Engine Land · Anu Adegbola · August 19, 2026Patient acquisition
What goes live today: ChatGPT Ads expand to 31 European countries — Germany, France, Spain, Italy, Sweden, Norway, Denmark, the Netherlands, Austria and more. OpenAI began testing ads in the US in February and has since added eight markets; this is its largest geographic expansion so far. Buying at launch runs through OpenAI's Ads Solutions team, agency partners and technology partners, with self-service via Ads Manager expected later.
The guardrails OpenAI states: ads appear only for Free and Go users — Plus, Pro and Enterprise stay ad-free. Ads are labeled and kept separate from ChatGPT's answers; advertisers do not receive users' conversations; and OpenAI says advertising does not influence the answers ChatGPT generates.
This is no longer an experiment — it's an ad platform. In the same update OpenAI says it has moved beyond CPM and CPC bidding to conversion optimization, added geo-targeting and custom audiences, shipped the OpenAI Pixel and a Conversions API, and added third-party measurement integrations. It says tens of thousands of marketers have now advertised on ChatGPT. That is the full performance-marketing stack, built in roughly six months.
The part with your clients' names on it, and it's older but it's the reason this matters: back in April, per OpenAI's own ad-policy changelog, medical, legal and financial advice contexts stopped being blocked by default — replaced by a "more precise approach." Health advertisers are now approved one at a time through manual review, with named eligible categories including medical testing, hospitals and urgent care, dental, vision and supplements. Legal services run the opposite way and are still barred. An SE Ranking study of 50,006 US commercial prompts collected July 23 found ads on 28.69% of prompts in its healthcare category — though note that measures commercial healthcare queries, not the personal-health conversations OpenAI's placement rules protect, and SE Ranking sells ChatGPT ad tracking. (SEJ, Matt G. Southern, Aug 12 — flagged as outside the freshness window and carried only as context.)
Why it matters for an independent practice: Zero medicine in today's news, so extrapolate. A paid channel is now forming inside the tool your patients use to decide where to go — and unlike Google, this one sits in the middle of a conversation where the patient has already explained their symptoms, their budget and their hesitation. Three things follow. One: health is conditionally open, so "can we even advertise there?" is now a real question with a real answer, and it's answered on OpenAI's ad-policies page — not the February launch post, which still says ads won't run near regulated topics and has not been corrected. Two: the rules have changed four times in five months, so any media plan written today can be stale before it runs; treat eligibility as a thing you re-check, not a thing you learn once. Three, and this is the trust question: the moment a practice can pay to appear next to an AI's health answer, "ChatGPT recommended us" stops being an authority signal and starts being an ambiguous one — for us, for competitors, and for the patient trying to tell the difference. Get ahead of that with clients now, because the first person to buy the placement will absolutely put it in a testimonial. Search Engine Land — ChatGPT Ads are expanding to 31 European countries · OpenAI — Advertising policies (the page to plan against)
