Officials outline AI training, sandboxing and vendor rules as legislators weigh access

The state's chief AI officer and the chief data officer briefed the Data and Impact Subcommittee on current AI work, training plans and governance safeguards, underscoring that tools should be used with traceability and data protections in place.

Chris Sarbocha, introduced as the chief AI officer, described efforts to build role-based "skills" for models, run coffee talks and office hours, and create a community of practice for agency staff. He said the office recently released an RFP for a master AI-qualified vendor list and is emphasizing sandboxes (for example, in Snowflake) that bring compute to the data rather than exporting sensitive information.

"Protect the data. Privacy, and make sure our policy is implemented," Sarbocha said, explaining why sandboxes and policy are central to the office's approach. He added that the office is focused on upskilling staff with targeted training — not just generic videos — and on providing role-based productivity tools that help staff complete specific tasks.

Chief Data Officer Adam Carver and committee members discussed practical checks that make AI work trustworthy: ask whether the output is a traceable formula or code (which yields mathematically verifiable results), require citation checks, and build light "audit skills" that can flag obvious errors in AI-generated reports. "If it's wrong, I'm the one who handed it to her," Carver said, underlining human responsibility for AI outputs.

Members also raised the question of whether legislators should be issued state AI tools. Legislative staff have begun adopting CoPilot, the discussion noted, but a tentative decision to withhold direct licenses from legislators is still under consideration. Committee members urged early training for legislators so they can better evaluate AI-derived analyses during the upcoming session.

Speakers gave concrete examples of how costs and deployment constraints are changing. The chief AI officer said model costs and token usage have fallen dramatically in months, making some previously expensive projects inexpensive today; at the same time, he said deployment containers and hosting remain friction points that prevent rapid roll-out of some prototypes.

The meeting included questions about trust and verification. One member described testing a commercial model on climate material and said it initially defended standard models until pushed; the exchange illustrated why experts and analysts will still need to validate outputs. Presenters noted litigation nationally over fabricated citations and urged staff to "trust but verify."

No formal policy or vendor selections were adopted at the subcommittee meeting. Presenters said the RFP process and pilot training activities will continue and recommended targeted AI literacy sessions for legislators and staff as the session approaches.