Good morning, {{first_name|Doctor}}. Last issue: 1,357 cleared devices, three with outcome data. This week, the other side of that ledger. Not what the tool adds, but what happens to the clinician after it arrives, measured in the one specialty that counts detection for a living.
In today's Med-AI Brief:
🩺 Research: detection fell six points with the AI switched off
⚖️ Regulation: the evidence tool that left the UK
🌍 Use cases: the Gulf builds, and the duty sits with the employer
🛠️ Practice: the two-minute habit that protects your own read
RESEARCH
🩺 The endoscopists got worse
The brief: Four Polish endoscopy centres introduced AI polyp detection at the end of 2021. When the same endoscopists went back to scoping without it, adenoma detection fell from 28.4% to 22.4%. Published in The Lancet Gastroenterology & Hepatology.
The details:
1,443 standard non-AI colonoscopies, 795 before AI arrived in the unit and 648 after. Absolute difference 6.0 percentage points, p=0.0089.
Observational and retrospective, nested inside the ACCEPT trial. The authors put it forward as a hypothesis, not a verdict, and this population is not ACCEPT's screening cohort. Case mix, staffing and secular trend cannot be ruled out.
Not an isolated signal. In a JAMIA study, 40 clinicians read 200 knee MRIs for ACL rupture. AI lifted accuracy from 87.2% to 96.4%. Of the errors that survived, 45.5% were traced to automation bias: the clinician deferring to a wrong AI call.
A 2026 scoping review in ESMO Real World Data and Digital Oncology pulls the same pattern together across specialties.
Clinicians already suspect it. Wolters Kluwer's 2026 survey found 74% naming loss of critical thinking or decision-making skill as one of the greatest risks. Only 27% could describe their organisation's AI governance policy.
Why it matters: Read the direction of travel, not the effect size. This is one observational study and it may not replicate. But notice what it threatens. Every current UK permission for these tools rests on the clinician as the safety control: the MHRA's position on ambient voice sits on you verifying the output, and the ACL data shows AI raising accuracy overall while still routing nearly half the remaining errors through misplaced trust. A control that quietly degrades with use is not a control, it is an assumption. Nobody costs six percentage points of detection into a business case, because nobody measures it after go-live.
REGULATION
⚖️ The tool that left rather than comply
The brief: OpenEvidence, the clinical evidence engine used daily by more than 40% of US physicians, withdrew from the UK and the EU on 27 April, citing regulatory uncertainty including the EU AI Act.
The details:
The company held multi-year content deals with NEJM Group and the JAMA Network, and raised $250m at a $12bn valuation in January.
Under the AI Act it would sit in the high-risk tier, with audit, transparency and bias-evaluation duties attached. The UK has guidance and principles, no statutory equivalent, and a National Commission still to report.
A commentary in Intelligence-Based Medicine works through the regulatory logic. This was a commercial withdrawal, not enforcement.
Why it matters: Regulatory uncertainty is not a neutral holding position. It has a price, and this is what the price looks like: the cited, journal-licensed tool leaves, and the gap fills with a general chatbot that cites nothing. That is the trade your registrar is making at 2am whether or not anyone sanctioned it. When the Commission's recommendations land, this is the case to hold them against.
USE CASES
🌍 The Gulf builds, and puts the duty on the employer
The brief: UAE hospitals are expanding AI across mammography, stroke, sepsis detection, neonatal care and triage, reported on 26 August. Cleveland Clinic Abu Dhabi runs Transpara on digital mammograms ahead of the reporting radiologist.
The details:
Deployments span Abu Dhabi, Dubai and Sharjah, from Sheikh Shakhbout Medical City to Al Jalila Children's. The announcements list capability. None of them list outcomes.
Abu Dhabi's Department of Health has had an AI policy since 2018 and added a Responsible AI Standard in 2025. It puts duties on the provider organisation: governance structures, regular audits of AI functionality with reporting obligations, explainability, and graceful degradation so a failing system alerts and stands down rather than drifting.
Why it matters: Two models sit side by side. Abu Dhabi makes the employer prove the system stays safe. The UK, for now, makes the clinician catch it. Given the first story, the employer-side model is the one that would notice deskilling, because audit is a standing duty rather than a one-off assurance at procurement. If you work across both systems, that difference is worth knowing before you sign anything.
PRACTICE
🛠️ Two minutes: read first, then look
Take your own view before the overlay loads. On the next AI-assisted read, film, slide or draft note, commit to your answer, even silently, before you see what the model says. Then compare.
Keep a tally of disagreements for a month: how often you were right, how often it was, how often you changed your mind and why. It costs seconds per case, and it is the only local evidence you will have that your own performance is holding. It also makes you considerably better company in the governance meeting.
QUICK HITS
📰 Everything else
Oncology: The Moderna and Merck personalised cancer vaccine would be the first approved medicine whose mRNA instructions are chosen by an algorithm.
Records: Stanford and Penn are deploying LLM chatbots that query and summarise the patient record. Striking single cases so far, no published error rates.
EU: The AI Act's Article 50 transparency duties have applied since 2 August. If a patient is interacting with an AI system, it has to be disclosed. That deadline was not among the ones deferred.
Evidence: Trials of clinical AI still cluster in well-resourced centres and exclude the patients most likely to be misread. Worth asking any vendor which groups their validation set left out.
How was today's brief?
✉️ One favour: hit reply and tell me whether anyone has measured your unassisted performance since AI arrived in your workflow.
Until next week,
Saeed
Know a colleague who'd use this? Send them your referral link.
Nothing here is clinical or legal advice. Check any tool against your own information-governance rules before it touches patient data.