At Fordham for Ignatian AI, a conference built around discernment: not what AI can do, but whether, when, and how we should use it. We need more meetings like this. #AIEthics #ResponsibleAI

From Copilots to Clinical Intelligence at DHNY Summit today, with Alexi Nazem, Rich Ziade, and Pete Clardy. Whatever else we cover, I'll circle back to the patient-clinician relationship. #DHNY26

MAUDE barely sees AI device failure: 98% of reports come from under five devices. Miscalibration is not an adverse event. #MedicalAI #FDA

CDRH suggests a prospective study may not be needed for every GenAI device. But only an RCT gives you a counterfactual, and you cannot synthesize a workflow. #MedicalAI #FDA

We don't ask screw makers to sit for boards. We test the constructs to failure. FDA already runs that regime - so why reach for licensing for GenAI?

FDA's GenAI device risk framework has two axes: activity and consequences. Neither is probability. The error rate enters later. #MedicalAI #FDA #ClinicalAI

FDA wants to credential GenAI devices like clinicians. But a device that passes its boards has no license to lose. Comments due Oct 19. #MedicalAI #FDA

Hospital LLM logs: most of the work is admin, not medicine. Time to optimize for HealthAdminBench over MedHELM? #MedicalAI #HealthIT Paper:

ICLR, AAAI, and now NeurIPS 2026 are running randomized experiments on their own peer review. Medical journals are not. #PeerReview #AI

Mortality 23.1% to 18.6%, 11 hospitals. The model: a logistic regression already in 187 Epic instances. The workflow did the work. #MedAI #ClinicalAI

Favorite recent medical AI paper: a negative RCT. 9,000+ patients in real clinics, patient-centered endpoint. Design matters. #MedAI #ClinicalAI

Benchmarks exist because most fields lack real endpoints. Medicine has them. So what is "medical superintelligence"? On ARISE MAST: #MedicalAI

An AI agent can be endlessly responsive and still not be responsible. Medicine's oldest infrastructure isn't a drug or a device - it is names. Whose name is on the chart? #MedicalAI #AIinHealthcare

Only 34% of claims survived independent reanalysis of the same data. In ML, releasing code didn't predict reproduction. Evaluating medical AI sits between benchmarks and trials. #ClinicalAI

61% word accuracy decoding text from MEG, no surgery. But nine healthy volunteers who can already type is not a locked-in patient. #MedicalAI #BCI

Who you ask changes the answer. 400+ physicians rated the same medical LLM outputs. Ratings shifted by seniority and practice setting - enough to flip model rankings. #MedAI #ClinicalAI #LLM

We studied the temporal lobe and reinvented the Recurrent Memory Transformer - ML reached it via NLP. What else is the brain hiding? #Neuroscience

Congrats to Prof. Jinseok Lee of Kyung Hee on a Presidential Commendation for ICT - incl. intl collaboration with NYU. Great people + shared rigour + common mission = global science. #ClinicalAI

OpenAI's unit distance proof ran on shared math. So does medical AI. What if the answer to value capture is to stop sharing?

AI disproved a 1946 Erdős conjecture in one shot. 9 mathematicians still had to digest it. The last mile, not the model. #MedAI