Little Hoover Commission hears state officials, industry and labor on responsible use of generative AI

California state officials, private vendors, academics and labor representatives outlined the benefits and risks of adopting generative artificial intelligence in state government at a Little Hoover Commission hearing on May 23 in Burlingame.

"We want to have a responsible adoption of generative AI in the state of California," Secretary of Government Operations Amy Tong told commissioners, describing the governor's executive order and a sequence of deliverables: a benefits-and-risks report, procurement guidance and a sandboxed environment for proofs of concept. Tong said the state has issued procurement guidelines and a six-month request-for-innovative-idea vehicle that allows vendors to demonstrate capabilities using state data for a nominal fee.

Tong said the state's approach centers on three priorities: protecting service delivery and workforce impacts through training and upskilling; testing limited use cases in an isolated, monitored sandbox; and coordinating procurement and technical standards with federal partners such as GSA and NIST. "We are taking this very seriously," Tong said, urging testing in narrow, controlled pilots before production deployments.

Industry witnesses described use cases and technical safeguards. Ed Abbo of C3.ai outlined back-office uses that can speed benefit adjudication and inspections, and urged retrieval-augmented generation designs to prevent hallucinations and make outputs traceable to source documents. As Abbo put it, "you can't run an organization on hallucination." He cited deployments in Riverside County and San Mateo County that sped up property assessments and investigations.

Meta's representative outlined the company's open-source model strategy and related safety tooling, saying openness can accelerate research and security testing while acknowledging tradeoffs. "Open source is a risk-based decision," Meta said in testimony, describing license restrictions and safety tools they provide to developers.

From the platform side, Google's Paige Bailey emphasized application-layer design: steerability, interpretability and safety; user interfaces that surface confidence and source citations; and education tools that treat models as tutors rather than answer engines. "We need to make the decision-making processes of AI understandable," she said.

Russell Wald of Stanford's Institute for Human-Centered AI urged public investment in academic compute and dataset access, citing the National AI Research Resource (NARE) pilot at the federal level and state bills that aim to create compute and data hubs. Wald warned that academia is losing talent and capacity to industry because frontier research requires large compute resources and rich datasets.

Labor testified that workers must be "in the loop" during design and procurement. Irene Green of SEIU Local 1000 said rank-and-file input, training, collective bargaining and transparency are essential to avoid deskilling, inappropriate surveillance or biased outcomes. Public commenters, including nurses and disability advocates, urged pre-deployment safety testing for clinical use, accessibility and stronger privacy protections.

The hearing highlighted tensions policymakers must weigh: procurement influence and market effects, the need for government data and compute for public-interest research, the potential efficiency gains for customer service and inspections, and the risks of bias, hallucination and workforce disruption. Witnesses repeatedly recommended controlled pilots, human oversight, traceability of AI outputs and stronger academic-government partnerships.

The commission said it will hold a second hearing in Sacramento focused on AI and disadvantaged communities. Commissioners also discussed follow-up work, including workforce reports and scheduled July deliverables.

The hearing record and staff background memo will be posted to the commission's website and YouTube channel; commissioners asked staff to make the background memo publicly available as a resource for future hearings and public engagement.