Forwarded this by a colleague? Subscribe free and get it every week.
THIS WEEK
This week's Spotlight is a Stanford–Harvard review of where clinical AI stands in 2026. On controlled tasks, models outperformed physicians, and clinician–AI teams still didn't beat the model alone. That leaves the bedside question open, and it raises a sharper one for anyone who teaches.
In this issue:
Epic pauses most development after an AI found MyChart flaws
Six physician societies answer the HHS Secretary on AI
Medicare asks whether AI could run parts of the wellness visit
Spotlight: the state of clinical AI, and what it means for teaching
How-to: set up ChatGPT for Clinicians
LATEST NEWS
Epic pauses most development after an AI found MyChart flaws
Epic has paused most product development to fix security flaws, Zack Whittaker reports in TechCrunch. CEO Judy Faulkner told Modern Healthcare the pause would likely last about six weeks. The flaws surfaced when Epic ran Anthropic's Mythos model against its software. Epic's security chief, Stirling Martin, told The New York Times that some MyChart configurations could let outsiders reach patient records without a log entry. Epic has not described the bugs, and neither report says whether anyone exploited them.
Why it matters: If your organization runs MyChart, ask your IT security team whether your configuration is affected and when fixes land, so you have an answer when patients ask.
Disclosure: Anthropic also makes Claude, one of the AI tools that help draft this newsletter. I pay for it like any subscriber and have no other relationship with the company.
Six physician societies answer the HHS Secretary on AI
The AMA, AAFP, AAP, ACOG, ACP, and ACS issued a joint statement after HHS Secretary Robert F. Kennedy Jr. said AI can give patients a second opinion "much better informed than any doctor in the country," Susan Morse reports in Healthcare Finance News, citing The New York Times. The societies said such claims undercut physician expertise and patient trust, and that safety and clinical judgment should guide how these tools are built. It is a position statement, not new data.
Why it matters: Expect patients to arrive with an AI second opinion. Ask what they typed in and what came back, then walk through it with them.
Medicare asks whether AI could run parts of the wellness visit
In its proposed 2027 physician fee schedule, CMS asks whether AI tools could gather pre-visit data, flag beneficiaries for more assessment, and draft follow-up steps for the Medicare annual wellness visit, Joyce Frieden reports in MedPage Today. It also asks what blocks AI companies from running those tools in partnership with enrolled providers (proposed rule docket). The AAFP objects to preventive care delivered outside the primary care relationship. These are questions, not policy; the final rule is expected around Nov. 1.
Why it matters: If you do wellness visits, the question for any vendor is who owns the follow-up when the software flags a patient, and whose name is on the order.
IN THE LITERATURE
Anesthesiology — Wu et al.
Design: Open-label RCT at two centers in Taiwan; 100 adults having major noncardiac surgery with arterial lines.
Finding: Treating when the Hypotension Prediction Index reached 85 was not superior to treating when MAP fell to 73 mmHg under the same protocol. Median time-weighted hypotension was 0.07 versus 0.16 mmHg (P = 0.12). Vasopressor dose, length of stay, and 30-day mortality did not differ.
Limitation: Small, unblinded, and not designed to show equivalence.
Vendor: HPI is Edwards Lifesciences' FDA-authorized algorithm. Trial funding is not stated in the abstract.
Bottom line: The algorithm has not shown it beats a higher pressure trigger. If your department pays for HPI, ask what it adds over a MAP alarm at 73.
NEJM AI — Nahass, Hanna, et al.
Design: Quasi-experimental, staggered before-and-after cohort; 23,132 adult admissions with an Epic Deterioration Index of 60 or higher at 11 RWJBarnabas hospitals in New Jersey, 2022–2024.
Finding: After the score was paired with chart banners, staff education, and automatic pages to the rapid response team, inpatient mortality fell from 23.1% to 18.6% (absolute difference −4.5 points, 95% CI −5.6 to −3.5; adjusted OR 0.82, 0.74–0.91). Rapid response activations rose from 25.3% to 37.5%. Lower mortality was concentrated at scores of 60 to 79 (adjusted ORs 0.75 and 0.78); above 80, the confidence intervals crossed 1.
Limitation: No randomization. The post period held far more academic-center patients (60% vs. 18%) and far fewer winter admissions (18% vs. 36%). Mortality also fell in the middle-risk tier, which got chart banners but no page, and near the threshold the page's own effect was not significant (OR 0.85, 95% CI 0.6–1.2).
Vendor: The Deterioration Index is Epic's proprietary model. Author disclosure forms are posted with the article.
Bottom line: A multipart rollout around Epic's score was associated with fewer deaths, but the study can't say which part did the work. If your hospital turns on the score, ask what response is wired to it, not just where the threshold sits.
npj Digital Medicine — Farrag et al.
Design: Five AI evidence-search tools (Consensus, Ai2 Paper Finder, ChatGPT, Gemini, Claude) tested with 15 query wordings against a non-public reference set the authors built in advance.
Finding: A single query found a median of 7% to 42% of the relevant evidence, depending on the tool. Pooling all 15 wordings raised that to 46% to 72%. Twelve percent was never found by any tool, and conference abstracts were missed far more often than journal articles.
Limitation: One reference set; results may differ by topic.
Vendor: The University of Florida authors declare no competing interests.
Bottom line: One AI query is not a literature search. Rephrase and rerun before you trust "no evidence found."
SPOTLIGHT
Clinical AI in 2026, and what it means for the people we teach
The Stanford–Harvard ARISE network has published the peer-reviewed version of its State of Clinical AI Report 2026 (Worth, Goh, Chen, and colleagues, BMJ Digital Health & AI). It is a narrative review of 82 studies, and the authors nominated 38% of them. Read it as a map, not a meta-analysis.
On curated cases, the capability is real. In one study of hard NEJM case conferences, OpenAI's o3 named the right diagnosis first 60% of the time; physicians with unrestricted search managed 24% on a set of 302 such cases. But accuracy is brittle. When cases arrived as dialogue instead of tidy summaries, GPT-4 fell from 82% to 63%.
The workflow results matter more. In a randomized trial of 70 physicians, AI help raised diagnostic accuracy from 75% to as high as 85%, but the AI alone scored 87%. In another trial, physicians shown deliberately wrong AI output scored 14 points lower, even after a 20-hour AI literacy course. And in an observational study, endoscopists' adenoma detection fell from 28.4% to 22.4% after a period of AI use.
A companion paper in the same journal is for those of us who teach. Patel, Hosamani, and colleagues, educators at one U.S. academic center, describe three risks: losing skills, never building them, and learning the tool's mistakes. A learner can check retrieved data against its source. Checking an AI-generated plan takes the skill to build one. Their stage-by-stage framework is untested.
The takeaway: Trust AI task by task, not on a general impression. Form your own read before you open its answer. If you supervise trainees, have them commit to a differential and plan before the tool weighs in, then talk through the gap. It is the old teaching question, "What do you think is going on?", asked before the screen answers it.
HOW-TO
Before you try this: check your organization's approved-tool list, and never put patient information into a tool it hasn't approved.
Set up ChatGPT for Clinicians, and know where it stops
OpenAI now offers a free ChatGPT workspace for verified U.S. clinicians, separate from any personal ChatGPT account. Details below are as of October 2026, from its product page and help article.
What you get. Clinical search that cites journal, authors, and date. Deep research reports across the medical literature. Starter prompts and reusable skills for referral letters, prior authorization letters, and patient instructions. CME credit on eligible clinical questions. Read-only connectors to PubMed, ClinicalTrials.gov, DailyMed, openFDA, and other public sources. OpenAI says your content is not used to train its models. Image generation and the Epic plugin are not included.
Who qualifies. The help article lists MD/DO, NP, PA, and pharmacists. The product page also names psychologists and other licensed clinicians. The verifier makes the call.
Start at the sign-up page. Sign in with an existing ChatGPT account or create one.
Verify with your NPI. A third-party provider checks it, and your license has to be verifiable.
Confirm and accept. Attest that you're a licensed clinician and accept the services agreement.
Select Get started with ChatGPT. You land in the new workspace. Switch back to your personal one from the account menu.
Before you type a patient detail. Don't enter PHI unless a BAA is in place. OpenAI lets an individual sign one under Settings > Agreements, but only if you are authorized to. If you are employed, that authority usually sits with your organization, and a BAA you sign yourself does not put the tool on its approved list. Check that list first.
A prompt to copy. No patient details needed.
You're assisting a licensed clinician. Summarize current guideline recommendations for managing GLP-1 receptor agonists before elective surgery.
Cite each source with the organization or journal, authors, and year.
Separate guideline recommendations from expert opinion.
List what remains uncertain or contested.
Don't assume any patient details I haven't given.What a usable answer looks like. Each recommendation names its source and year, says whether it is a guideline or expert opinion, and flags what is still contested. If an answer skips any of those, ask again.
How to check it. Open at least two of the citations and confirm they say what the answer claims. Then ask in different words. This week's npj study found that single queries to AI search tools, ChatGPT among them, missed most of the relevant evidence.
When to stop. Never act on AI-generated patient-specific dosing or contraindications without checking a primary reference. The search quality and the CME are OpenAI's claims; this issue cites no independent evaluation of either. For the full checklist, our free guide Safe Prompting for Clinicians still applies.
I have no financial relationship with OpenAI.
THAT'S IT FOR THIS WEEK
How was this issue?
Send us a tool or a how-to. Using something in clinic that other clinicians should know about? Submit it here and it may show up in a future issue. We ask before we publish anything and credit you by name only if you want us to.
Know one colleague who’d use this? Forward this email, or send them the link below.
Loading Dose AI is for educational purposes only and is not medical advice. Nothing here should guide the care of a specific patient, and reading it does not create a physician-patient relationship. AI tools help gather and draft each issue; I read every source and edit every word. Opinions are mine and not my employer's. Full disclaimer