The signed law will require makers and health care sites using AI tools that guide care to make reasonable efforts to find bias risks and take steps to reduce them.
Makers and care sites must make reasonable efforts to spot care-guiding AI tools with known or reasonably expected risks of harm tied to traits protected by law. Makers must try to reduce those risks.
Health care sites must regularly check care-guiding AI tools with known or reasonably expected risks of harm tied to protected traits. They must take reasonable steps that fit the risk.
Makers must make uses, risks, and testing records available for covered AI tools on request or at first sale, whichever comes first, and after major updates when relevant. Care sites making and using tools only internally can skip detailed records if equivalent information is readily available.
Makers and care sites must make reasonable efforts to identify covered AI tools with known or reasonably expected bias risks. Makers must try to reduce those risks. This covers tools that help health facilities, clinics, and doctors' offices decide when to give care, diagnose, or treat. The harm must be based on a trait protected by state law. Worse access to care, quality of care, or health outcomes are examples. Routine scheduling, patient education, and payment tools are excluded when their independent work needs no professional license.
Care sites must regularly check covered AI tools and take reasonable steps that fit the risk to reduce biased harm. This applies to health facilities, clinics, and doctors' offices using care-guiding AI tools identified with known or reasonably expected bias risks. The risk must involve harm based on a trait protected by state law.
Makers must make information about covered AI tools' intended uses, risks, and testing available to care sites that use them. This applies to care-guiding AI tools identified with known or reasonably expected risks of harm tied to protected traits. The detailed records cover training data, testing, and ways to reduce bias risks. A care site that both makes and uses a tool only internally may skip the detailed records if equivalent information is readily available to it. Makers must make the information available on request or at first sale, whichever comes first, and, when relevant, when they release major updates.
No hearings or votes scheduled.