Most clinical AI systems are trained on pre-pandemic assumptions about human biology.
Post-COVID physiology has changed – and the models have not.
Health AI systems are being deployed at scale under an assumption that the underlying biology of patient populations is stable. That assumption is no longer safe.
Since 2020, widespread viral exposure, immune dysregulation, vascular injury, and autonomic disturbance have altered baseline physiology across large segments of the population. These changes are subtle, heterogeneous, and poorly captured by existing diagnostic labels.
As a result, many AI systems now operate in an environment where:
- training data reflects a different biological regime
- labels are noisier and less specific
- validation metrics overestimate real-world reliability
This creates a silent failure mode: models appear robust, yet degrade precisely in complex, high-risk patients. The issue is not algorithmic sophistication, but biological mismatch.
My perspective is informed by clinical experience before, during, and after the pandemic, as well as by analysing trends across health systems rather than focusing on individual cases. My emphasis is on examining mechanisms and trajectories—investigating how physiology evolves over time and under stress—instead of relying on snapshot diagnoses or static classifications. Rather than being concerned with whether an algorithm functions in theory, my primary interest lies in assessing its potential for undetected failure at scale.
Dr Robert Pearl is one of the most experienced system-level voices in modern healthcare:
- Former CEO of Kaiser Permanente, responsible for care delivery to over 12 million people
- Professor at Stanford University School of Medicine and Business
- Author of ChatGPT, MD and Mistreated
First Published Paper on COVID-19 Autoimmunity
Involving Plasma/Serum ACE-2
