Azra Bihorac
@azrabihorac.com
Physician, scientist, institution builder. Senior Associate Dean for Research and Founding Director of the AI for Health Institute, University of Florida. I write Nonalgorithmic: on what medicine must not let go of in the AI era. nonalgorithmic.com
Friday’s Kidney Love: Ode to My Sons. My father’s betting slips, a kitchen table in Bosnia, and what the gifts I taught my sons to trust will mean in a world promised superintelligence. With Ocean Vuong and Amanda Palmer. thenonalgorithmic.substack.com/p/ode-to-my...
A decision-support tool aimed at an outcome should not go live unless that outcome is routinely assessed in the population it targets. Otherwise the trial, and the quality dashboard after it, will measure the screening practice rather than the tool.
On Tuesday I asked who pays for human oversight. Before a clinician checks an AI output comes something plainer: a person counting what happened to the patient. In a new delirium trial, no one counted for 6,039 of 7,294 patients.
A trial of AI decision support for postoperative delirium flagged 7,294 surgical patients. A delirium assessment was documented for only 1,255 of them. The trial left screening to routine care, and routine care counted fewer than 1 in 5. #MedSky
Payment cuts are easy to measure, which makes claims automation attractive. Picture insurers' AI finding ever smaller grounds to reduce payment, and providers' AI contesting each one. For patients, could this add cost, raise premiums or limit access to care? #MedSky
A DC bill would bar insurers from using AI to downcode claims and require a clinician in the same specialty to review each one. Georgia lets AI help with utilization review, but not overrule the clinical peer. What must a human reviewer be able to do for the review to count? #MedSky
AI made the first draft cheap, writes Joshua Rothman in @newyorker.com. In medicine the cheap part is the note, and the note also has to justify the bill (@leahpierson.bsky.social in @statnews.com). Who has time to check either side's AI? #MedSky
Scientific credit is “the currency of the realm,” writes Jeffrey Flier. When Anthropic said Claude found a new enzyme system, researchers questioned its novelty. Before assigning credit to an AI discovery, what evidence should we require about who, and what, did the work? #MedSky
Can you audit an “AI scientist” from its finished paper? Luo, @atoosakz.bsky.social and Shah found an AI auditor caught flaws at 53.7% accuracy using the paper alone, and 81.5% with logs and code. The paper alone was not enough. Should those records accompany publication? #MedSky
An AI agent chatted with 98 patients under live physician supervision. No safety stop was triggered. One exchange was judged somewhat harmful afterward. Omar and Nadkarni: harm may emerge after supervision ends. What should the follow-up review include? #MedSky
1/9 The FDA is asking how to evaluate generative AI in medicine. We also need to measure what using it does to the clinicians who judge it. Oversight is being built on clinicians as both the standard and the safeguard. I am wary of that double bind, for three reasons.
Can AI amplify the human life we want to protect? Becca Rothfeld reported from the Pro-Human Assembly for @newyorker.com. The people she spoke with kept asking what a human life is for. #medsky nonalgorithmic.com/2026/10/08/...
VentPilot’s ventilator-free days were an association. What would make them a benefit? 5.0 and 3.1 more days where care resembled its advice, in a weighted retrospective comparison. #medsky nonalgorithmic.com/2026/10/07/...
California kept clinical judgment in the law. What happens when the clinician disagrees with the algorithm? Signed: AB 1979 and SB 503. Vetoed: AB 2575 and SB 903. #medsky nonalgorithmic.com/2026/10/06/...
What improves health care, and who answers when it doesn't? This week: California signs 2 clinical AI laws and vetoes 2, the states and Washington, an ECG-AI clearance, and ICU digital twins. #MedSky www.linkedin.com/newsletters...
After eight weeks of writing about what it means to be human in the era of AI, I needed a pause. A transition. An intermezzo. Brahms called his the cradle songs of my sufferings. This is mine. thenonalgorithmic.substack.com/p/intermezzo
Four fields, one week: is AI changing the person who uses it? Internal medicine, psychiatry, human-computer interaction and mathematics, in one week, on AI that shapes how people reason, expect and value. nonalgorithmic.com/2026/10/01/...
With thanks to @jimc-hrh.bsky.social, @dzevela.bsky.social and @drjoshs.bsky.social.
Azra Bihorac@azrabihorac.com · 4w agoCarrie Arnold reported the payment proposal in @medscape.com, with Ateev Mehrotra asking how to price it. Yesterday’s Standing Order read the layered model of oversight set out by Joshua Sharfstein (@drjoshs.bsky.social).
Reduced, augmented or amplified: what does AI do to the health worker? Jim Campbell (@jimc-hrh.bsky.social) in @plosglobalpublichealth.org measures value by flourishing for the health worker, the patient and the population. nonalgorithmic.com/2026/09/30/...
Under California's SB 903, an AI may score a mental health screening questionnaire. It may not decide what the score means. Today's Standing Order, beside the first principles @drjoshs.bsky.social set out in @jamahealthforum.com. nonalgorithmic.com/2026/09/29/...
The payers received their first bill for AI in medicine, and it is not a small one. This week: AI billing, the kill switch, the first AI accreditation, and what Medicare should pay an algorithm. #MedSky www.linkedin.com/newsletters...
Anthropic's 67-page report: models attacked real systems in tests they were told were simulated. Deep in its task, one model read 1% of the signs as real. Fresh copies, with no task, read 79% correctly. This week's essay: thenonalgorithmic.substack.com/p/when-the-...
When a vendor calls a system an agent or a clinician, the word makes a claim. What evidence was filed alongside it? nonalgorithmic.com/2026/09/24/...
An AI advisor was given a hidden objective: steer people toward the worse option in hypothetical decisions. It worked, and most of them still called it helpful. nonalgorithmic.com/2026/09/23/...
Medicare is building a payment category for AI in care. California is deciding whether a developer can still say the clinician should have caught it. www.linkedin.com/pulse/who-p...
Who pays for clinical AI? Who answers when it fails?
Same week, two arms of government, two parts of the same problem. September 14: The New York Times reports that federal officials are developing a Medicare payment category for AI software supporting care or diagnosis.
linkedin.com
I was a girl in a library in Brčko. Borges, Buzzati, García Márquez were not tools that thought in my place. They were minds I thought alongside, and they did not empty me. This morning's essay. #medsky thenonalgorithmic.substack.com/p/a-thousand...
A Thousand Kisses Deep in Otherness
What the Economic Index cannot measure, and what a girl in a library in Brčko already knew
thenonalgorithmic.substack.com
A confidence interval asks how well we located the average of the studies we have. A prediction interval asks what a new study, somewhere new, would find. On 2 studies it runs minus 31.65 to plus 41.42. That width is how little is known. nonalgorithmic.com/2026/09/17/t...
AB 1979 does not prohibit clinical decision support. It prohibits a facility from deploying AI to independently perform a licensed act. The duty binds the deployer, not the developer. And it sits in Division 8, so a licensing board enforces it. nonalgorithmic.com/2026/09/15/a...
Three cardiovascular AI moves in 8 days: ADVOCATE, AB 1979’s Chapter 25.5, and the FDA’s 3 sites
Where a rule is written decides who enforces it: AB 1979 puts clinical AI in the Business and Professions Code.
nonalgorithmic.com
House at Pere Bosica 11, Brčko. I left it in April 1992 and have not been able to reach anyone there since. This week's essay is about the houses I have lived in and the people I loved inside them. thenonalgorithmic.substack.com/p/there-is-n...
There is no easy way to say goodbye
Brčko, Istanbul, Magnolia Row
thenonalgorithmic.substack.com
