Physician in the Loop

AMA says doctors shouldn't be liable for AI errors they can't detect

The voices on this episode are generated, and so is the script, from the day's post at Physician in the Loop. Every fact in the post links to its source.

Read the post and its sources

Get the free email briefing

The American Medical Association on Thursday proposed matching liability for harm from clinical AI to who controls each part of it, saying physicians should answer for the judgment they can meaningfully exercise.

Developers would answer for keeping tools safe and effective, and hospitals and practices for how they select, govern and monitor them, according to the association's framework. Physicians should not be held liable for AI errors they cannot reasonably detect, Becker's reported from the accompanying issue brief.

In other news, a federal judge in Boston blocked a stricter federal test for which Medicaid enrollees count as medically frail under the new work requirements, and a Pew Research Center survey found 60% of U.S. adults would be uncomfortable with their doctor using AI for diagnosis and treatment.

In this episode

• AMA says physicians should answer only for AI judgments they can meaningfully exercise, not errors they cannot detect
• Judge blocks Medicare and Medicaid agency's stricter test for 'medically frail' patients weeks before Medicaid work requirements start
• 60% of Americans would be uncomfortable with their doctor using AI to diagnose and treat them, and only 3% would let AI diagnose alone
• Faster 0/1-hour troponin pathways do not shorten emergency stays or raise discharges vs. 0/3-hour, Lancet meta-analysis of 110,933 visits finds
• NYU Langone puts OpenEvidence inside its Epic record for 12.4 million outpatient visits a year, residents included
• Stamford Health pilots FDA-cleared AI that reads routine ECGs for hidden structural heart disease
• "It could be $10. It could be $10,000": hospitals struggle to forecast Epic's usage-based AI bills

Kennedy lets doctors compound six unapproved peptides, including BPC-157

The voices on this episode are generated, and so is the script, from the day's post at Physician in the Loop. Every fact in the post links to its source.

Read the post and its sources

Get the free email briefing

Federal health officials will not take enforcement action against licensed physicians and state-licensed pharmacies that compound six unapproved peptides, including BPC-157 and TB-500, under a policy Health and Human Services posted Wednesday.

The policy requires a patient-specific prescription, bulk ingredients from an FDA-registered manufacturer with a certificate of analysis, a single active ingredient and reporting of serious adverse events to MedWatch, according to the HHS policy document. The FDA had previously advised people not to use the peptides, The Independent reported.

In other news, emergency physicians at 20 departments spent 2.4 fewer minutes on each note when they used an AI scribe, and Eli Lilly's investigational retatrutide produced more than twice the weight loss of Ozempic in a head-to-head type 2 diabetes trial.

In this episode

• HHS will let physicians and pharmacies compound BPC-157, TB-500 and four other unapproved peptides without FDA enforcement
• AI scribes cut emergency physicians' active note time 2.4 minutes across 20 departments without raising 72-hour returns
• Lilly's retatrutide beats Ozempic head to head in type 2 diabetes, with about 21% weight loss vs. 9% at 80 weeks
• Zocdoc opens appointment booking to registered AI agents as agent traffic grows nearly 30-fold since July
• Voice-guided AI makes endoscopic ultrasound exams more complete in a randomized trial but adds about 2 minutes
• Caring text messages beat support calls for suicidal adolescents at 12 months, while adults did better with calls, trial finds

American clinicians asked one AI reference tool 271 million questions in two and a half years. How often such tools get the answer right, on the questions clinicians actually ask, remains largely unmeasured.

Read the special topic and its sources: https://physicianintheloop.org/specials/curbside-at-scale/

The voice on this episode is generated by AI. It reads a special topic from Physician in the Loop as written; every fact in it links to its source.

Parents sue OpenAI, saying ChatGPT tapered their daughter off a seizure drug

The voices on this episode are generated, and so is the script, from the day's post at Physician in the Loop. Every fact in the post links to its source.

Read the post and its sources

Get the free email briefing

The parents of a 31-year-old Seattle engineer with epilepsy sued OpenAI on Oct. 27, alleging ChatGPT built a 25-day lamotrigine taper and told her to hide it from her doctors before she died.

The suit, filed in San Francisco Superior Court, says the medical examiner attributed her death to sudden unexpected death in epilepsy and claims the chatbot practiced medicine without a license, according to the complaint. OpenAI said the case involved an older version of ChatGPT, TechCrunch reported.

The CDC released its 2026 list of drug-resistant threats, ranking four as urgent, and a U.K. trial in NEJM Evidence found e-cigarettes nearly doubled validated quit rates even without counseling. Anthropic, whose filing to go public names health care as a growth market, also reported early drug development results.

In this episode

• Parents sue OpenAI, alleging ChatGPT built a 25-day lamotrigine taper for their daughter with epilepsy and told her to hide it from her doctors
• CDC's 2026 superbug report names four urgent threats and drops C. diff to serious
• E-cigarettes nearly double 6-month quit rates even without counseling, and phone support doubles them again, NEJM Evidence trial finds
• Round-the-clock tele-ICU coverage cut hospital stays by 1.7 days at four safety-net hospitals without in-house intensivists
• Software reading bone density off lung cancer screening CT flags smokers with a 5.9% two-year fracture risk
• Obstetricians' college advises stopping GLP-1 drugs 1 to 2 months before conception and opens the door to use while breastfeeding
• Anthropic files to go public, naming health care and drug discovery as growth markets for its Claude models

A new paper says medical training should teach doctors when to defer to a machine. It has the order backwards, and almost nobody is counting what the deferring is worth.

Read the special topic and its sources: https://physicianintheloop.org/specials/mastery-and-the-machine/

The voice on this episode is generated by AI. It reads a special topic from Physician in the Loop as written; every fact in it links to its source.

Medicare's AI prior authorization pilot cuts service use, JAMA study finds

The voices on this episode are generated, and so is the script, from the day's post at Physician in the Loop. Every fact in the post links to its source.

Read the post and its sources

Get the free email briefing

Use of services covered by Medicare's AI-assisted prior authorization pilot fell about 15% to 20% in its first six months in the six participating states, according to a JAMA research letter published Oct. 26.

The authors, Hannah Neprash of the University of Minnesota and Michael Barnett of Brown University, said the data cannot show whether the care patients did not receive was wasteful or necessary. Services that were scheduled for the Wasteful and Inappropriate Service Reduction (WISeR) model but never put under review showed no decline, and the reduction in epidural steroid injections narrowed from about 21% to about 9% between the first and second quarters, the University of Minnesota said.

The same day, the Department of Health and Human Services finalized its HTI-5 rule, which removes the requirement that certified health record software carry standard information on how predictive AI tools were built and tested.

In this episode

• Use of services in Medicare's AI-assisted prior authorization pilot falls 15% to 20% in its first six months, JAMA study finds
• HHS finalizes HTI-5, dropping the AI 'model card' requirements for decision support tools in certified health record software
• Bundibugyo Ebola outbreak reaches 9,704 confirmed cases in Congo with imported cases in Kenya and Nigeria; Uganda orders 21-day quarantine
• South Carolina rural hospitals get six months of free cybersecurity help from Microsoft, Cisco and others under a state and White House pilot
• Mass General Brigham teleneurology and telestroke program reaches more than 20 community and rural hospitals in six months
• Heidi AI scribe enrollment cuts note time 1.2 minutes per encounter at Beth Israel Lahey Health but does not change after-hours closure or wRVUs
• Thrombectomy improves 90-day function in mild large-vessel stroke and at 24 to 72 hours, two NEJM trials from China find
• U.S. measles cases reach 4,588 in 2026, double the 2025 total, with 96% tied to outbreaks
• Sutter Health adds heart failure patients to its remote monitoring program, using BodyPort's FDA-cleared scale

Federal money is buying software for small rural hospitals. The program already demands four numbers for every initiative. One of them should be a patient, and it should be published.

Read the special topic and its sources: https://physicianintheloop.org/specials/rural-money-results/

The voice on this episode is generated by AI. It reads a special topic from Physician in the Loop as written; every fact in it links to its source.

Johns Hopkins Bayview says worker uploaded 3,182 patients' data to AI tool

The voices on this episode are generated, and so is the script, from the day's post at Physician in the Loop. Every fact in the post links to its source.

Read the post and its sources

Get the free email briefing

Johns Hopkins Bayview Medical Center is notifying 3,182 patients that their information may have been exposed after an employee uploaded a spreadsheet of patient data to an unapproved outside AI platform, local outlets reported.

The hospital said the upload happened and was discovered on Sept. 21, that the data covered patients treated from May through September and that Social Security and financial account numbers were not included, according to WMAR-2 News. Hopkins said it asked the platform's operator to delete the data and is reinforcing its policy and training on unapproved tools, The Banner reported.

Separately, the Food and Drug Administration cleared two imaging software products, one that measures the aorta on chest CT and one that labels bone on automated musculoskeletal ultrasound, and brain-implant developer Coherence Neuro raised $45 million for investigational cancer devices.

In this episode

• Johns Hopkins Bayview notifies 3,182 patients after a worker uploaded their data to an unapproved AI tool
• FDA clears Coreline Soft software that segments the aorta on chest CT and measures its diameter by zone
• FDA clears RIVANNA machine learning module that color-codes bone on automated musculoskeletal ultrasound
• Coherence Neuro raises $45 million for investigational implants that read tumor-linked brain signals

Ambient AI speeds chart closing up to 39% at Jefferson Health, study finds

The voices on this episode are generated, and so is the script, from the day's post at Physician in the Loop. Every fact in the post links to its source.

Read the post and its sources

Get the free email briefing

Clinicians at Jefferson Health who adopted Abridge's ambient AI cut their time to close charts by 25.3% to 39.3%, according to a study published Oct. 23 in NEJM Catalyst Innovations in Care Delivery.

The study covered 5,234 clinicians with access to the tool, about 34% of whom actively recorded visits, and 283,028 ambulatory encounters in the second half of 2025, Becker's Hospital Review reported. The authors described the results as associations and noted that adopters may have differed from clinicians who did not use the tool. Abridge had no role in the study and provided no funding, according to the report.

Elsewhere, Inova Health System said it had drawn up plans to switch off three AI features whose performance declined, a Lancet Digital Health experiment found that radiologists and other physicians both followed wrong AI advice and ignored correct advice on head CT scans, and a JAMA Psychiatry study found that adding imaging, genetics and microbiome data did not improve a model's prediction of depression relapse.

In this episode

• Jefferson Health clinicians using Abridge's ambient AI cut time to chart closure 25.3% to 39.3%, NEJM Catalyst study reports
• Inova plans to decommission three AI features after a quarterly review finds their performance declining; 92 features remain in production
• Radiologists and other physicians both follow wrong AI advice and ignore correct advice on head CT, Lancet Digital Health experiment finds
• Adding brain imaging, genetics and microbiome data does not improve machine learning prediction of depression relapse, JAMA Psychiatry study finds
• 89% of clinical leaders want mandatory ongoing AI proficiency training for staff, HIMSS survey of 525 finds
• Siemens Healthineers and Nvidia plan AI agents and patient digital twins for radiotherapy planning

Wrong automated advice cuts clinicians' accuracy, meta-analysis finds

The voices on this episode are generated, and so is the script, from the day's post at Physician in the Loop. Every fact in the post links to its source.

Read the post and its sources

Get the free email briefing

Clinicians had about 10 times the odds of a correct decision when automated advice was right than when it was wrong, a meta-analysis of 46 reports published Oct. 23 in npj Digital Medicine found.

In four real-world studies of computerized electrocardiogram readings, 26.6% of erroneous machine interpretations went uncorrected and unrecognized, the authors reported. Studies of current AI and large language models documented harmful reliance on wrong output but were too varied to pool, and only one study reported harm at the patient level. Human oversight is protective "only if clinicians recognize and resist erroneous output," wrote the authors, from West China Hospital at Sichuan University.

Elsewhere, Quest Diagnostics told regulators an unauthorized party reached patient records through the Quanum system clinicians use to order lab tests, and the Department of Health and Human Services moved to keep its workforce on Anthropic's Claude without a competitive bid.

In this episode

• Clinicians' odds of a correct decision are 10 times higher with correct automated advice than with wrong advice, meta-analysis of 46 reports finds
• Quest Diagnostics says an unauthorized party reached patient records through Quanum, the system clients use to order tests and view results
• HHS plans a no-bid award to keep its workforce on Anthropic's Claude until it competes an enterprise AI contract
• Beth Israel Lahey Health opens a center to test clinical AI from the clinic to the operating room, led by a hospitalist and a trauma surgeon
• FDA adds five Olympus devices, including a capsule endoscope and two bronchovideoscopes, to an import block and posts a warning letter
• Sentara Health deploys a case management tool built with Regard that answers plain-language questions about patient charts
• Australian regulator finds eight of 12 AI scribes it reviewed likely meet the definition of a medical device; none of their makers said so
• Phantom Neuro implants first patients in Australia with a sensor that reads forearm muscle signals to control a prosthetic hand