State officials describe expanding AI pilots, risks and election threats as agency use grows

State technology managers and policy experts briefed the Government Operations Interim Committee on June 18 about expanding artificial intelligence use inside state government, pilot projects for agency call centers and tax services, and risks AI poses to election systems and public information.

“Artificial intelligence is a technology with both tremendous promise and tremendous peril,” Alan Fuller said, summarizing the challenge as agency leaders described pilots, governance and procurement cautions. Officials and outside experts urged careful oversight even as Utah agencies roll out tools intended to boost productivity.

Why it matters: Utah agencies have begun integrating large language and generative AI tools into daily work and customer service. Committee members heard both benefits (efficiency gains, improved drafting and faster retrieval of policy text) and risks (incorrect or hallucinatory outputs, data exposure, disinformation and deepfakes affecting elections). Lawmakers and staff noted recent legislation and audit work aimed at improving signature verification and election integrity.

State programs and use: Marvin Dodge said the state launched Google’s Gemini for executive‑branch employees this year and reported about 146,000 Gemini transactions in the previous four weeks among the state’s 7,500 active Google Workspace users (out of roughly 21,000 executive‑branch employees). DTS reported 42 generative AI projects under consideration and 10 in production; pilots include automated chat‑assistance for call centers and a tax‑commission project that loaded tax code and training materials into candidate chatbots. Dodge described a pilot “bake off” of four vendors where initial correct‑answer rates were about 75% and vendors proposed further refinement toward a 99% target.

Security, procurement and governance: Officials advised caution about vendor claims and data handling. Marvin Dodge and others told the committee that the state checks for FedRAMP or GovRAMP certifications when purchasing AI services and is concerned about startups that do not build in adequate data controls. Committee members asked about workforce effects and possible staff redeployments; Dodge said AI could make employees more productive but acknowledged potential for workforce changes.

Elections and deepfakes: Zachary Boyd, director of Utah’s Office of AI Policy, and Adam Cook‑Hook of the National Conference of State Legislatures reviewed how AI is already used in elections. Boyd said, “by default, we should be very cautious about deploying AI in any election operations because this is such critical social infrastructure.” He described common current uses—signature verification and duplicate‑voter detection—and warned that deepfakes and AI‑generated disinformation could cause rapid, high‑impact harms in campaigns and voting.

Policy responses and standards: Panelists urged two complementary steps: (1) strengthen provenance and content‑credential systems so legitimate publishers can cryptographically label content, and (2) build reliable ways for trustworthy actors (campaigns, election officials, news organizations) to verify themselves. Adam Cook‑Hook outlined options used by states—disclosure requirements, short pre‑election prohibitions and metadata or content‑credential mandates—and noted that 25 states have enacted laws addressing election deepfakes. He also said a proposed federal moratorium in a budget bill could affect state authority for a period.

Next steps: State officials said agencies will continue pilots, refine training for staff, require FedRAMP/GovRAMP‑type assurances for vendor contracts and work with legislative partners on any statutory changes. The Office of AI Policy and other state offices offered to consult on further legislative work.

No formal committee action was taken; members later discussed opening committee files on signature verification and other AI‑related items.