Virtual Seminar
Friday September 11, 12pm-1pm Register Here
Health systems are deploying AI faster than anyone can determine whether it works, for whom, and at what cost. The standard response has been to evaluate models more rigorously before release, yet a year of conversations across the MIT Critical Data community, spanning clinicians, engineers, patients, and communities in dozens of countries, points to a harder truth. The data that trains these systems carries the imprint of the care environment that produced it, errors resist detection because the causal analysis required to find them exceeds what most institutions can perform, and risk compounds unpredictably once multiple AI components interact with each other and with the humans around them. Evaluation must therefore move from the model to the system into which it deploys, and no single organization can deliver that assessment on its own. Implementation science for clinical AI has to reckon with questions the field has largely avoided: whose care patterns the machinery encodes, why patient-centeredness requires an actual redistribution of power, and why capacity building belongs in barbershops and churches as much as in academic medical centers. Underneath it all sit incentive structures that reward speed over understanding, in medicine as much as in the society it serves.
About Leo Anthony Celi, MD, MS, MPH
Leo Anthony Celi is part of the team behind the Medical Information Mart for Intensive Care or the MIMIC database, now stewarded by over 100,000 credentialed students and researchers across six continents. The group has championed open science not as a methodological preference but as a moral imperative: knowledge generated from patients must be returned to communities, not sequestered behind institutional gatekeeping. The work is anchored on the conviction that science must earn, and continually re-earn, its place in the public trust, and insists that rigorous inquiry must be humble enough to learn from lived experience and ways of knowing that resist quantification.
Through workshops and hands-on data analysis across the globe, his team surfaces that modeling data without understanding how it came to be - what political, economic, and cultural forces shaped its collection, its silences, and its distortions - is epistemic violence. Such is like cooking without knowing the ingredients: whether there are nuts, shellfish, meat, or worse, cyanide. The dish may taste exquisite to those who sit comfortably at the table, but it is poison to those whose conditions were undertreated, whose pain was disbelieved, whose bodies were rendered invisible by the very systems the data was drawn from. A model trained on such data does not merely inherit these harms; it industrializes them, encoding historical injustice into algorithmic certainty and deploying it at scale.
Add to Calendar2026-09-11 19:00:002026-09-11 20:00:00AI Seminar Series - Leo Anthony Celi, MD, MS, MPH
Virtual Seminar
Friday September 11, 12pm-1pm
Register Here
Health systems are deploying AI faster than anyone can determine whether it works, for whom, and at what cost. The standard response has been to evaluate models more rigorously before release, yet a year of conversations across the MIT Critical Data community, spanning clinicians, engineers, patients, and communities in dozens of countries, points to a harder truth. The data that trains these systems carries the imprint of the care environment that produced it, errors resist detection because the causal analysis required to find them exceeds what most institutions can perform, and risk compounds unpredictably once multiple AI components interact with each other and with the humans around them. Evaluation must therefore move from the model to the system into which it deploys, and no single organization can deliver that assessment on its own. Implementation science for clinical AI has to reckon with questions the field has largely avoided: whose care patterns the machinery encodes, why patient-centeredness requires an actual redistribution of power, and why capacity building belongs in barbershops and churches as much as in academic medical centers. Underneath it all sit incentive structures that reward speed over understanding, in medicine as much as in the society it serves.
About Leo Anthony Celi, MD, MS, MPH
Leo Anthony Celi is part of the team behind the Medical Information Mart for Intensive Care or the MIMIC database, now stewarded by over 100,000 credentialed students and researchers across six continents. The group has championed open science not as a methodological preference but as a moral imperative: knowledge generated from patients must be returned to communities, not sequestered behind institutional gatekeeping. The work is anchored on the conviction that science must earn, and continually re-earn, its place in the public trust, and insists that rigorous inquiry must be humble enough to learn from lived experience and ways of knowing that resist quantification.
Through workshops and hands-on data analysis across the globe, his team surfaces that modeling data without understanding how it came to be - what political, economic, and cultural forces shaped its collection, its silences, and its distortions - is epistemic violence. Such is like cooking without knowing the ingredients: whether there are nuts, shellfish, meat, or worse, cyanide. The dish may taste exquisite to those who sit comfortably at the table, but it is poison to those whose conditions were undertreated, whose pain was disbelieved, whose bodies were rendered invisible by the very systems the data was drawn from. A model trained on such data does not merely inherit these harms; it industrializes them, encoding historical injustice into algorithmic certainty and deploying it at scale.
GastroenterologyAmerica/Los_Angelespublic