Forwarded this by a colleague? Subscribe free and get it every week.
THIS WEEK
Welcome to the first issue. It lands every Monday morning. Five minutes, five sections, no hype.
This week: the FDA asks how generative-AI devices should be evaluated, Washington state considers making physicians explicitly accountable for AI-assisted decisions, and state AI laws keep multiplying. That is the Spotlight, and it ends with one question worth asking every clinical-AI vendor.
In this issue:
ChatGPT can now read an Epic chart. Read-only, for now.
Ambient scribes are moving onto nursing units.
A rare FDA De Novo for an ECG model that flags the heart attacks standard criteria miss.
In the literature: the largest multisite AI-scribe study yet, the mammography trial with the strongest interval-cancer evidence so far, and a deterioration score wired straight to the rapid response team.
Spotlight: who is liable when the AI is wrong.
How-to: get a deeper answer out of OpenEvidence with one tap.
LATEST NEWS
ChatGPT can now read your Epic chart. Read-only, for now.
OpenAI announced on September 1 that approved ChatGPT for Healthcare workspaces can connect to Epic. A clinician can ask what changed since the last visit or which labs need review, and the model answers from the authorized record. It cannot write to the chart, place orders, or message patients. Access rides on your existing Epic permissions, and an administrator has to switch it on. A separate Healthcare Public Data plugin searches PubMed, ClinicalTrials.gov, DailyMed, RxNorm, and CMS coverage data.
Why it matters: Connectivity is not validation. Epic access shows the tool can retrieve authorized data; it says nothing about whether its summaries are complete or reliable. If your system enables it, ask whether access is audited, which parts of the record are exposed, whether prompts and responses are retained, and where to report a wrong chart summary. Organizational access does not mean approval for every clinical use.
Ivan Mehta, TechCrunch; OpenAI announcement
Ambient documentation is coming for the nursing flowsheet
Mercy, FMOL Health, Northeast Georgia, Mount Sinai, and Jefferson Health are piloting ambient AI in nursing workflows, and Northwell plans to add nurses by year end, Becker's reports. Mercy's three-unit pilot of Microsoft's Android tool cut flowsheet charting time 22% per shift, with system-wide rollout planned by mid-2027. How that time was measured, and whether nurses spent more time correcting generated entries, is not reported. These are health-system interviews, not a published study.
Why it matters: Nursing flowsheets are structured, repeated, device-fed entries, a different task from a physician's narrative note, so the physician-scribe results do not carry over automatically. If your unit pilots this, watch time spent correcting entries, not just entering them, and ask how omissions and wrong-field or wrong-patient errors are measured.
Naomi Diaz, Becker's Hospital Review
Powerful Medical's "Queen of Hearts" model, developed with Hennepin County emergency physician Stephen Smith, reads a 12-lead ECG for STEMI and STEMI-equivalent patterns, the occlusion MIs that fail classic criteria on a first tracing. De Novo is the pathway for a novel low- or moderate-risk device with no predicate on the market; fewer than ten AI devices a year go through it. The company claims about twice the sensitivity of standard reads with fewer false alarms. Those figures are the vendor's, and without the comparator, reference standard, setting, and confidence intervals they do not mean much yet.
Why it matters: An overread at first contact could catch the occlusion MIs a first ECG misses and shorten the trip to the cath lab. Before your ED adopts it, ask four things: was performance externally validated, in what prevalence and patient mix, what was the sensitivity for occlusion MI specifically, and how many additional activations or urgent cardiology reviews occurred per extra case caught.
Katie Palmer, STAT News
IN THE LITERATURE
JAMA — In a difference-in-differences study of 8,581 ambulatory clinicians at five academic centers, access to an AI scribe was associated with 16 fewer documentation minutes per eight scheduled patient-hours. After-hours EHR time did not change significantly, and gains were greatest in frequent users. Design: multisite longitudinal cohort. Bottom line: a modest workflow benefit, not evidence of better patient outcomes. Rotenstein LS et al.
The Lancet — In MASAI, 105,934 Swedish women randomized to AI-supported screening or standard double reading: 80.5% versus 73.8% sensitivity, a non-inferior interval-cancer rate, and 44% less reading workload. Design: randomized non-inferiority trial. Bottom line: strong evidence for AI-supported screening in this double-reading workflow; not evidence that autonomous reads are ready or that every program will see the workload gain. Gommers J et al.
NEJM AI — Routing Epic's Deterioration Index straight to rapid-response notification across 11 hospitals was associated with in-hospital mortality among high-risk admissions falling from 23.1% to 18.6% (23,132 admissions). Design: quasi-experimental staggered pre-post; association, not proof. Bottom line: the intervention was the model plus a response pathway, not the score. Before copying it, ask who receives the alert, the expected response time, alert volume, and escalation rules. Nahass TA et al.
SPOTLIGHT
Who is liable when the AI is wrong? Increasingly, you.
Medscape's Steph Weber pulled the regulatory picture together this month, and the through-line is uncomfortable. The rules for medical AI are being written now, and the liability is landing on whoever signs the note.
Start with the FDA. On August 18 its device center published a discussion paper on how generative-AI devices should be evaluated, with public comment open until October 19 under docket FDA-2026-N-7874. The proposed approach borrows from medical training: benchmark the model's clinical knowledge and safety behavior before market, confirm it under clinical conditions, then re-test after updates, the way we sit boards and recertify. The paper says plainly that it is not policy. It is 26 questions and an invitation. Clinicians are on the list of people the FDA wants to hear from. Very few of us will comment. The vendors will.
Two caveats. First, the FDA regulates devices, not generative AI as such. Whether a tool counts as a device turns on its intended use and on the clinical-decision-support exclusion, and many decision-support tools and every general-purpose chatbot currently fall outside it. Second, the states are moving faster. More than a dozen passed healthcare-AI laws this year, including California's limit on chatbots using clinical titles and, from 2027, rules on AI-assisted utilization review.
The part that should change your behavior is a draft policy from the Washington Medical Commission. It would hold physicians fully responsible for AI-assisted decisions, require meaningful review of AI output, and back a clinician's right to decline a tool when the vendor cannot supply reliability data. It is one state's draft, not a national rule, and liability will still turn on jurisdiction, employment, contracts, and what the vendor claimed. But it may become a template for other boards. The Federation of State Medical Boards plans draft AI guidance for public comment in early 2027. Health-law scholar Sara Gerke, quoted by Weber, points out that vendors have every reason to oversell and may never be asked to back the claims up.
The takeaway: we already work this way with drugs and devices. Nobody prescribes off a press release. An AI tool that touches a clinical decision deserves the same standard. Ask the vendor for the intended-use statement, the validation population, absolute error rates and subgroup performance, known failure modes, the monitoring plan, and a change log after model updates. Then ask who receives the safety reports and whether your institution can switch off a new version. That list fits on an index card, and it is worth bringing to the next committee meeting. A vendor who cannot answer has told you something.
Steph Weber, Medscape Medical News; Steve Alder, HIPAA Journal; FDA discussion paper; Washington Medical Commission draft (June 30, 2026 packet)
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.
Match the OpenEvidence model to the question
OpenEvidence, which by its own count has 1.12 million license-verified U.S. clinicians on the platform, split into three models on September 8, all free to verified clinicians on the web and in the iOS and Android apps. (Model names and availability are as of September 2026; they will change.) The difference between the three is time and search depth, not a different accuracy claim. Most users will never touch the selector. Try this next week:
Find the selector. Open OpenEvidence and look for the model selector by the question box. Osler is the new default: about five seconds, built for the question in the hallway.
Switch to Sackett when the evidence is the question. "Is there RCT evidence for X in this population?" It takes about 30 seconds and will often ask you a clarifying question first. Answer it. That is the point.
Run Snow for the hard case. Competing comorbidities, a differential you cannot close, a drug question with thin evidence. It follows several lines of inquiry through the literature and returns a report in about five minutes. Start it, see your next patient, read it after.
Open a source before you act. Whatever the model, click through to at least one cited paper. For anything touching dosing, contraindications, pregnancy, renal impairment, or a rare diagnosis, one source is a floor, not a ceiling.
A prompt worth trying in Sackett: "In an adult with community-acquired pneumonia and a severe beta-lactam allergy, what randomized or guideline evidence supports the available empiric regimens? Separate guideline recommendations from trial evidence, name the populations studied, and list the evidence gaps."
The stop rule: do not use the answer alone for patient-specific dosing or contraindications. Verify the primary source, the current guideline, and your institution's formulary or protocol.
Skip Darwin. It is a research preview, by application only.
One reminder: at most institutions none of these tools is approved for patient information. Ask the general question, not the identified one.
Nathan Eddy, MobiHealthNews; OpenEvidence announcement. Loading Dose AI has no financial relationship with OpenEvidence.
THAT'S IT FOR THIS WEEK
One click on how this issue landed helps us make the next one better.
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