Ohio’s 88 county election boards now have a 24/7 AI research assistant named EVA—but it can’t reliably tell you the next election date. The flaw, confirmed in a July 29 report by The Columbus Dispatch, exposes the precise tension that IT teams face whenever they deploy a retrieval-augmented chatbot in a high‑stakes environment: speed without verification is a trap.

EVA Enters the County Office

EVA—short for Elections Virtual Assistant—is an internal chatbot trained solely on Ohio’s 524‑page Election Official Manual and the state’s annual election calendar. It is not connected to the public internet, a design choice that keeps responses bounded to official Ohio guidance. Launched in November 2025 as a beta for the Secretary of State’s office, the tool became available to all county boards of elections to reduce the hours staff spent manually hunting through dense rulebooks.

On paper, it does exactly that. EVA can pull procedures for managing a voter-registration database, guide an official through a post‑election audit checklist, or surface a directive on ballot custody—all with a simple search prompt. According to Secretary Frank LaRose, what once required flipping through a 524‑page binder now comes back in seconds.

But the tool’s limitations are not hidden. LaRose himself has warned users that EVA is not legal counsel and that its answers must be verified against statutes and official directives. The date‑handling problem isn’t just an embarrassing bug; it’s the kind of failure that can push a registration deadline, an early-voting window, or a certification step into dangerous territory. In election administration, a confidently wrong answer is worse than no answer at all.

What Ohio’s Tool Actually Does—and What It Doesn’t

EVA is not a public‑facing assistant like many consumer AI tools. Voters cannot access it, and it doesn’t answer questions about polling places or ballot measures. The system is exclusively a knowledge‑retrieval engine for trained election staff, bounded by the contents of a single manual and calendar. In technical terms, it’s a retrieval-augmented generation (RAG) chatbot—it searches a trusted corpus, retrieves relevant passages, and generates a summarization. The manual it relies on is authoritative, versioned, and updated on a defined schedule.

That architecture is smart for a domain where answers must be exact and auditable. But RAG is not immunization against hallucination. A model can still misread a date, omit an exception buried in a footnote, or answer with confidence when the manual has been revised but the old version is cached. In Ohio’s implementation, the chatbot’s inability to handle future election dates—despite being trained on the election calendar—shows that even a controlled source set doesn’t guarantee correct output.

The Real-World Risks When AI Meets Election Rules

Election officials aren’t just technologists; they are statutory officers bound by strict timelines and legal definitions. When a chatbot gives a wrong deadline, it’s not a customer‑service mishap—it can alter ballot access, observer rights, or recount procedures. For the IT teams supporting these workflows, the threat model is broader than just accuracy:

  • Source integrity: If the manual isn’t the single source of truth—or if the tool silently references an outdated edition—every output becomes suspect.
  • Logging and auditability: A system must record which questions were asked, which document sections were retrieved, and whether the user accepted the answer. Without that chain, there is no way to reproduce a decision.
  • Escalation paths: When EVA says something that contradicts a staff member’s training, they need a documented process to flag the conflict and get a human legal review.
  • Off‑topic requests: A restricted corpus decreases the chance of an off‑topic response, but it doesn’t eliminate it. Staff might ask “What should I do about a voter intimidation complaint?” and receive a confident but legally incorrect summary because the manual doesn’t cover that scenario explicitly.

Ohio has acknowledged these risks by mandating human verification and limiting access to trained officials. That’s the right instinct, but it places heavy reliance on the user’s judgment—exactly the muscle that atrophies when a tool becomes trusted.

A Tale of Three States: Ohio, West Virginia, Maine

Ohio’s guarded rollout is not the only approach in the nation. West Virginia Secretary of State Kris Warner told ABC News that chatbots can return incorrect local guidance when a user leaves out essential context—a problem he illustrated with Harrison counties, of which there are eight nationwide. “If they don’t put in the name of the state, they’re not going to get the right information,” Warner said. His office has not adopted a similar tool, preferring to keep human‑only research for now.

Maine’s position is even more conservative. Deputy Secretary of State Kate McBrien told ABC that “right now, it is too risky. There’s just too many unknowns and too much unverified information … until there’s really good insurances on cybersecurity, Maine is not comfortable going there for elections yet.” Those concerns go beyond the chatbot’s accuracy. They extend to question whether prompts, source documents, retained conversations, or model integrations could expose sensitive operational data or create a misleading record of what staff were told.

For IT professionals, these three states’ positions are a rare public‑policy split that illuminates a classic enterprise dilemma: do you bet on a constrained, internal‑facing tool with clear limitations, or do you hold out until the technology matures? Ohio chose the former with teeth‑clenching caveats; West Virginia and Maine chose caution.

Lessons for IT Pros Managing Any Sensitive Knowledge Base

Even if you’re not running an election, the EVA deployment carries lessons that apply directly to Windows shops and enterprise IT environments where a “chat with your documentation” feature is under consideration.

  1. Pin the source of truth and version it explicitly. Ohio’s manual is a 524‑page PDF. In a corporate setting, that might be a SharePoint library, a private GitHub wiki, or a Confluence space. Before a chatbot touches it, IT must ensure the corpus is the definitive, most‑recent copy. A chatbot that mixes policy revisions will erode trust faster than any downtime.

  2. Require citations on every answer. A response that says “Based on Section 4.2.3 of the manual, the deadline is…” is verifiable. A bare summary is not. If the tool can’t provide a link back to the source passage, it shouldn’t be deployed in a regulated workflow.

  3. Enforce role-based access and separate the system from public services. Ohio’s decision to keep EVA off the internet and away from voters is crucial. In an IT department, that means the chatbot should not share a network segment, authentication domain, or logging path with anything public. If a privileged user can accidentally expose the tool’s logs, you’ve created a new attack surface.

  4. Design human‑in‑the‑loop from the architecture, not as an afterthought. Verification shouldn’t be a policy memo; it should be a workflow that creates an audit trail. For example, require a second official’s electronic sign‑off before an AI‑generated deadline is entered into a calendar system. In Windows environments, that can mean leveraging Active Directory approvals or Power Automate flows that gate actions.

  5. Train staff to break the trust reflex. The more helpful the chatbot is, the less people will question it. IT training must include real examples of the tool getting dates wrong, misinterpreting a policy, or failing silently. The goal is not to make staff skeptical of the tool but to make verification a habit.

What to Watch Next

Ohio says EVA cut research time for routine questions during testing, and the tool remains in active use by county boards. The Secretary of State’s office will have to decide whether to address the date‑handling flaw with a model retrain, a different date‑handling module, or a hard‑coded rule engine—any of which would set a precedent for how government AI fixes production weaknesses.

At the same time, other states are watching. If Ohio can demonstrate a clean audit trail and a measurable reduction in procedural errors—while never letting an AI‑generated date slip through unchecked—it may provide the blueprint that West Virginia, Maine, and others demand. If a date‑related mishap makes headlines, the political blowback could stall election‑office AI for years.

For IT and Windows professionals, EVA is a live case study in how to—and how not to—deploy a domain‑specific assistant where the penalty for being confidently wrong is measured in votes, not support tickets.