The legal treatment of clinical AI has been theoretical for long enough that the theory got quite elaborate. Three malpractice suits filed this spring are now testing it against facts.

The cases point in different directions, which is what makes them collectively interesting. In two, the clinician followed a model recommendation that turned out to be wrong. In the third, the clinician overrode a correct model recommendation. The plaintiff theory in all three is essentially the same: the standard of care was not met.

The unresolved question underneath is whether a well-validated model's output becomes part of the standard of care, such that departing from it requires justification. Health system counsel have been arguing for years that it does not. Several plaintiff firms are now arguing that it does, at least where the model is in widespread use for the indication.

Nothing will be settled quickly. But the practical advice being circulated to clinicians is converging: document the reasoning for the decision, whether it agreed with the model or not, and do not document the model's output as though it were the reasoning.