Assemblymember Liz Ortega's AB 2575 did not have an ordinary passage. On 28 August the California Senate read it a third time and refused passage, 18 ayes to 10 noes — short of the 21 votes needed in a 40-seat chamber. Reconsideration was granted immediately and unanimously, 40-0. On 31 August it passed 21-10, the Assembly concurred the same day, and the bill went to the governor.
What the bill requires
Health facilities would have to disclose information about clinical decision-support systems on a health worker's request: training data, known biases, validation and performance metrics. Workers would be protected from retaliation for declining to follow an AI recommendation. The main duties — system inventory and detailed disclosure — start 1 July 2027. Physician violations go to the Medical Board of California and the Osteopathic Medical Board; retaliation goes to the Labor Commissioner.
The clause that reallocates liability
Buried in the text is the provision with the largest commercial consequence: in a civil action, a defendant cannot rely on a health worker's failure to override an AI output as a basis for breaking causation. Today, "the clinician approved it" is the standard defence available to a hospital or a vendor. Removing it moves exposure back toward the institution that deployed the system and the company that sold it.
Two framings to avoid
First, the 28 August vote is being described as the Senate rejecting the bill. Procedurally it was "refused passage" with reconsideration granted the same day — the accurate description is failed and revived, not defeated. Second, 21-10 is the bare minimum, not a comfortable margin; ten senators did not vote at all. A bill that squeaks through on its second attempt is a weaker signal to a governor than the raw "passed" suggests.
Status
Not law. It is one of roughly two dozen AI bills sitting with Governor Newsom after the legislature adjourned, and he has until 30 September to act.