GRASS VALLEY, Calif. June 25, 2026 – The Nevada Irrigation District’s Board of Directors workshopped a newly developed Artificial Intelligence policy Wednesday morning, and invited public feedback before any formal adoption. Directors suggested several changes to the draft, mostly establishing stronger guardrails.

The presentation by Assistant General Manager Greg Jones drew chuckles when he acknowledged the obvious: the district used AI to help write its AI policy.

“AI helped us create our AI policy,” Jones told the board. “However, whatever AI tells you in the world, you have to verify it and have it overlooked by human interaction.”

The proposed policy covers all NID employees, contractors, and third-party vendors using AI on the district’s behalf. It addresses all current forms of AI, including generative AI (LLMs like ChatGPT, Claude, Gemini, Microsoft Copilot) machine learning systems, and predictive analytics tools, whether hosted internally or through outside providers.

Jones described it as a starting point rather than a finished framework, noting it will be updated as the technology evolves. At three pages, it is already one page longer than a comparable policy at neighboring Placer County Water Agency, he stated.

Shall, not should

Director Ricki Heck had flagged ahead of the meeting that the transparency clause in section 3300.3.3 — currently reading that stakeholders “should be informed” when AI tools are used — needed stronger language. The board agreed: “shall” will replace “should” in the final draft, making disclosure a requirement rather than a suggestion.

Jones agreed, “Just like when I created this and told you that we used AI to create this policy. It’s the transparency that’s important.”

The equity question

Director Rich Johansen took issue with language in section 3300.3.1, which states that AI “must support equity, efficiency, and innovation.” He called the term vague and potentially loaded. “AI put that in there,” quipped one board member. Directors agreed the term “equity” should be removed. Johansen suggested the simpler standard: “do everything you say you’re going to do and do no harm.”

Hallucinations and accountability

Several directors pressed on how the district would ensure human accountability when AI-generated content contains errors. Director Ricki Heck pointed to a recent local example: an attorney in the Nevada County District Attorney’s office whose AI-assisted legal filings contained fabricated case citations discovered by opposing counsel.

“AI hallucinates like crazy,” said Director Earl Stephens. “They’ll cite references ‘John versus so-and-so’ – that never existed. They make stuff up. We definitely need serious human oversight.”

The existing draft addresses this in sections 3300.3.2 and 3300.3.4, requiring human review and approval before any AI-generated output is executed and making employees responsible for validating accuracy. But directors pushed for more specificity. Heck floated the idea of requiring employees to disclose when AI assisted in producing a work product. Jones acknowledged the concept while noting a blanket signature requirement may not be practical given how pervasive AI tools have become. He suggested a notation – “AI assisted” – appended to relevant work products as a workable middle ground.

Under the proposed section 3300.4.3, AI prompts and outputs may be subject to California Public Records Act requests and are to be treated accordingly.

Security

Director Heck raised the question of what happens if AI gives a hydro operator incorrect guidance on water releases. Peter Wade, NID’s Director of Power Systems, clarified: all water release decisions currently pass through a minimum of four people and follow the existing chain of command. No AI system is making autonomous operational decisions. “If we’re using any AI supplement to help us optimize our routines for releases, that would be checked by multiple people before it’s actually executed,” Wade said.

On cybersecurity, IT Administrator John Ortiz confirmed the district is evaluating Microsoft Copilot configured to operate entirely within NID’s own environment — not the open internet — so sensitive files stay behind existing firewalls. “When we ask it a question, it uses our data. It does not go out to the internet unless you ask it to.”

The draft policy explicitly prohibits entering personally identifiable information, protected health information, or any other confidential or sensitive data into AI systems without prior legal authorization. It also bars AI from being used in employee hiring or promotion decisions, vendor RFP evaluations, and healthcare or benefits determinations. It prohibits using AI to produce work products that require a professional license, such as legal, engineering, or surveying documents.

Next steps

The board directed staff to update the policy’s language throughout and to run the draft through additional AI tools for further stress-testing before the next version comes to the board.

When President Chris Bierwagen called for public comment on the AI policy, Director Brad Fowler deadpanned that someone named Sarah Connor had briefly raised her hand on the Zoom participant list.

The policy is expected to return for formal adoption at a future meeting, tentatively July.