What specific details about an AI system's training data and performance must health facilities include in their inventory under AB 2575?
Short answer
Under Assembly Bill 2575, health facilities in California must provide specific technical and clinical details about artificial intelligence tools upon request. The required inventory includes summaries of training data, demographic representativeness, known biases, validation processes, and qualitative measures of performance. These disclosures aim to ensure health care workers understand the limitations and reliability of the systems they use.
What the bill or law says
The bill amends the Health and Safety Code to require facilities, clinics, and physician offices to maintain an inventory of clinical decision support systems (CDSS). According to the proposed Section 1339.76, this inventory must be available to licensed health care professionals who use or view the system outputs. The text specifies that the information provided must include a summary of the training data or clinical research underlying the recommendations. This summary must address demographic representativeness and any known biases based on protected characteristics. Additionally, the inventory must contain a summary of the validation process and qualitative measures of performance. Other required details include the system’s developer, intended use, cautioned out-of-scope uses, and a link to the federal Certified Health IT Product List [https://legiscan.com/CA/text/AB2575/id/3449525].
How it works in practice
Health care facilities must update this inventory at least annually. By July 1, 2027, these entities must make the inventory available to staff upon request. Employers are also required to notify relevant staff members upon hire and annually of their right to request this information. This process ensures that doctors, nurses, and other clinical personnel can access critical data about the tools influencing their medical decisions. The requirement applies only to AI used for clinical decision-making, not for administrative tasks like documentation or automated messaging [https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260AB2575#94AMD].
What the source does not answer
The bill text does not specify the exact format or technical standard for presenting the training data summaries. It does not define what constitutes sufficient detail for "qualitative measures of performance," leaving some interpretation to the facilities. The source also does not outline the specific penalties for providing incomplete or inaccurate inventory information, focusing instead on the right to access the data and protection from retaliation for overriding AI recommendations. Furthermore, the text does not detail how small practices with limited resources will manage the administrative burden of maintaining these detailed records [https://leginfo.legislature.ca.gov/faces/billStatusClient.xhtml?bill_id=202520260AB2575].