Earlier this year we put out a call for election offices willing to discuss how they use AI in their offices. A group of offices from across the country took us up on it.
Over a series of pilot workshops, run in partnership with the Election Security Exchange, Ready for Tuesday and the AI + Elections Clinic at the Arizona State University Mechanics of Democracy Laboratory, we sat down with offices at different stages of adopting AI and writing a policy for it, and asked how they were approaching the work.
Here is what we heard, and what we are building next.
You may already have most of a policy
Much of what an AI policy needs to say is probably already written down somewhere in your office. Computer-use policies, data privacy rules, records retention and acceptable use rules all cover a lot of the same ground, and AI puts new pressure on every one of them.
That makes this a good moment to revisit these policies. Pull them out, read them with AI in mind and adjust where the language has gone stale. Just as important, treat it as a chance to socialize what they say in new ways. A good document nobody reads changes nothing.
The power and web-based nature of AI tools raises the stakes. A single prompt can move a lot of data fast. That is reason enough to make sure the rules are current and that staff actually know them.
What we learned from our conversations
The conversations covered the same ground for everyone: scope and purpose, use categories and guardrails, data handling and accountability.
Most offices want to lead with their values rather than their rules. One tied its policy to delivering better service at lower cost. Others started from protecting voter data and public trust. One office pushed back on the whole instinct to control, warning that a restrictive, all-encompassing policy can smother the upside before an office ever gets to it.
When it came to specific uses, offices tended to sort AI work into the same buckets: internal communications, voter-facing content, analysis, media generation, recruitment, accessibility and translation. Most landed on the same starting point, which was to allow each use only with human review and tighten from there. Language translation was the exception that drew real disagreement. One office allows it with review. Another bans it outright because of a law in their state.
On data, everyone drew the same first line. No PII, no sensitive data and no voter data go into AI tools. Several offices did not write a policy from scratch at all. They folded AI guidance into an existing data classification or computer-use policy. More than one worried about AI features bundled into scheduling and project management tools that nobody had vetted.
Much of this travels well beyond the election office. In one jurisdiction we had people from across the municipality’s government in the room, not just elections team members and leaders, and they reminded us how much sensitive data other departments handle – financial information in one office, health information in another. The same starting points apply whether you run elections, utilities or a health department.
The through line was human control. People stay in charge, and AI does not make decisions. Offices talked about grounding answers against trusted sources and matching the level of review to the stakes. They also shared the failures that taught them this, like a tool that summarized only 800 rows of a 30,000-row spreadsheet.
Accountability was the hardest part. Offices are still working out who owns the guidance, how staff get answers when a situation is unclear, how often to review it and what a violation should cost. The problem almost none of them had solved was more basic. Most cannot yet tell when a violation has even happened.
What we’re building next
The offices were clear on one thing. Staff need simple guidance and direction as much as they need a formal policy. A quick-start guide, a one-page dos-and-don'ts and a short video often reached staff better than a long document. A formal policy still matters. But whatever form the guidance takes, it has to say the same thing everywhere staff encounter it.
So that is where we are starting. Here is what is on the way:
AI Dos and Don’ts generator, live now at readyfortuesday.com/tools/ai-use-dos-and-donts. It lets an office build a plain-language one-pager that fits how it actually works, with a set of permanent don’ts that cannot be edited away.
A data classification template, modeled after this one from Greenwich, Connecticut.
Platform-by-platform advice, tailored for each AI tool, on how to confirm your privacy settings are set correctly and, for enterprise accounts, how to set up alerts.
A general AI use policy guide, with sample language, considerations for what to include and what to leave out, and guidance on nesting your AI rules inside a computer-use or data policy you already maintain.
Some resources are already out there to help you get started:
AI Terms for Election Officials (Ready for Tuesday)
Ready for AI (Ready for Tuesday)
Ready to Train the Bots (Ready for Tuesday)
AI in Elections: Procurement Essentials (Election Security Exchange)
Originally posted on the AI and Elections Substack, July 8, 2026. Read the original ↗
