THE MED-AI BRIEF
October 13, 2026 · 5-MINUTE READ · Read online
Good morning, {{first_name|Doctor}}. On Wednesday The Lancet published the first peer-reviewed study of a diagnostic chatbot taking histories from real primary care patients before they saw their doctor. Ninety-eight conversations, no safety stops. The number worth repeating at handover is a different one: 57%.
In today's Med-AI Brief:
🩺 Research: the AI took the history. Did it steer the doctor?
🌍 Use cases: UpToDate can now read the chart. Can it read it right?
⚖️ Regulation: same name, new algorithm, no clearance
🛠️ Practice: read the AI summary second, not first
RESEARCH
🩺 The AI took the history. Did it steer the doctor?
The brief: On 8 October The Lancet published a prospective feasibility study from Beth Israel Deaconess Medical Center in Boston. Google's AMIE (a large language model built for diagnostic conversation) took a history by text chat from adult patients up to five days before their primary care appointment, then offered possible diagnoses for them to discuss with their clinician.
The details:
114 patients enrolled; 98 completed both the chat and the appointment. Single centre, run April to November 2025.
A physician watched every conversation live and could stop it. None needed stopping. Supervisors logged one hallucination and stepped in to add clinical clarification five times.
Clinicians read the AI transcript or summary before the visit in 44 cases. 75% said it helped them prepare; 57% said it may have influenced their clinical approach. One clinician rated an interaction as somewhat harmful.
Health outcomes were not measured.
Why it matters: This is a safety and feasibility study, under live supervision, not evidence that the tool improves care. Its own authors frame it that way: "AI tools must first be shown to operate safely and acceptably in real care workflows," says Adam Rodman. The 57% is the finding to sit with. In more than half of the reviewed cases, a summary written by software shaped the consultation before the clinician met the patient. That is anchoring (fixing on the first idea offered), and it is the risk that will matter as patient-facing triage spreads, including the AI triage tool NHS England is adding to the NHS App. Patients' concerns about confidentiality and the chatbot's honesty also persisted after using it.
USE CASES
🌍 UpToDate can now read the chart. Can it read it right?
The brief: On 8 October Wolters Kluwer and Suki launched Ask, the first product built on the new UpToDate Expert AI API. A clinician asks a question by voice or text inside the record, and the answer combines UpToDate content with the patient's own chart.
The details:
Ask pulls medications, allergies, vitals, lab results and chronic conditions from the record without a manual search, and answers with linked sources and dosing considerations.
It runs inside Epic, Oracle Health, athenahealth and MEDITECH.
The companies say clinicians preferred answers that used patient context 83% of the time in early evaluations. No sample size or method has been published.
Why it matters: Clinical reference has always lived in a separate tab. Once it reads the chart, the question changes from "is the source good?" to "did it read the right part of the record?" A missed allergy or a stale eGFR now changes the answer, not just the context. The launch is in US health systems; no UK or Gulf availability has been announced. If it reaches your trust or hospital, the evaluation to ask for is how often it pulls the wrong chart data, not how often clinicians like the answer.
REGULATION
⚖️ Same name, new algorithm, no clearance
The brief: On 6 October the FDA posted a warning letter to O.N. Diagnostics over version 3.0.0 of VirtuOst VFA, its CT-based bone analysis software. The verdict: a new algorithm, and a move to the web, each needed a fresh clearance.
The details:
Version 3.0.0 added a new machine learning model for vertebral landmarking, moved the software from a desktop application to a web-based platform, and migrated its technology stack.
The FDA said the new algorithm "fundamentally changes how the device performs, fails, and affects patients" and asked the company to stop distributing that version until the issues are resolved.
The original clearance (K171435) covered the earlier software, not the update.
Why it matters: It is a US letter about one product, but the principle travels. An AI tool is not a fixed object; the version your department evaluated may not be the one running today. The UK is writing its own answer: the government's response to the National Commission commits the MHRA to draft guidance by December 2026 on predetermined change control plans (agreed rules for how a model may be updated after approval). For Gulf hospitals buying tools on the strength of a US clearance, the practical question for the vendor is simple: which version was cleared, and is it the one we are running?
PRACTICE
🛠️ Two minutes: read the AI summary second, not first
AI-written histories, pre-visit summaries and scribe notes are arriving in records before you meet the patient. The Lancet study suggests they shape the consultation. Keep the shaping deliberate:
Form your own first impression from the patient's opening account before you read the AI's differential.
Check the one item that would change management (a red flag, an allergy, a current medication) against the patient or the primary record.
If you relied on AI-generated history, say so in your note, as you would for a history taken by a colleague.
Thirty seconds each. The aim is not to ignore the summary. It is to make sure you are agreeing with it, not inheriting it.
QUICK HITS
📰 Everything else
AI Airlock, phase 3: The MHRA's regulatory sandbox opened applications on 6 October. Apply before 31 October for the current wave. The priority this round is post-market surveillance, and there is no fee. Worth forwarding to whoever leads AI projects in your trust.
Gulf diary: No new Gulf regulatory announcements this week. Two big dates are close: the Department of Health's Future Health Summit at ADNEC, Abu Dhabi, 20 to 22 October, and the Global Health Exhibition in Riyadh, 26 to 29 October. We will report what regulators say at both.
This Thursday: The Health Foundation's AI in the NHS 2026 runs online on 15 October.
Watching: NHS England and NHS Resolution indemnity guidance for AI (promised, no date); MHRA change-control guidance (December); HSSIB's ambient voice investigation.
How was today's brief?
✉️ One favour: hit reply and tell me whether an AI-written summary has reached your patients' records yet, and whether you read it before or after you saw them.
Until next week,
Saeed