What Really Happened When Quebec Voters Trusted an AI Chatbot for Election Info

What Really Happened When Quebec Voters Trusted an AI Chatbot for Election Info

Voters in Quebec turned to an automated chatbot for basic voting instructions during the 2022 provincial election, and the bot flatly lied to them. Not out of malice. It just couldn't handle simple questions without making things up.

That's the core takeaway from testimony given to Canadian Members of Parliament during hearings into artificial intelligence and election integrity. When people asked the tool where to cast their ballots or what identification they needed to bring, they received fabricated rules, wrong locations, and flat-out bad legal advice.

If you've ever built or tested a customer-facing AI tool, this shouldn't shock you. Large language models and older rule-based hybrids aren't databases. They're prediction machines. When you put one between a citizen and a ballot box, every single hallucination becomes a potential barrier to voting.

Why the Quebec Election Chatbot Failed Voters

Election agencies love the idea of automated chat tools because call centers get crushed during advance polling days and on election night. Thousands of people ask the exact same questions at the exact same time. Where is my polling station? Can I vote if I moved last week? Do I need my voter information card?

On paper, an automated assistant solves that bottleneck instantly. In practice, election rules are full of edge cases that trip up software.

During the parliamentary committee hearings in Ottawa, MPs heard directly about how an election chatbot deployed in Quebec fed inaccurate details to residents trying to exercise their democratic rights. Instead of admitting it didn't know an answer or routing the user to a human clerk at Élections Québec, the system confidently delivered wrong information.

Think about what happens when a voter gets told the wrong hours for a polling place. Most people don't double-check an official-looking answer on a government or civic portal. They just show up at the wrong time, see locked doors, and go home. Voter suppression doesn't require a conspiracy when bad software does the job for free.

The Structural Problem With AI in Civic Spaces

Here is what most technology vendors won't tell government procurement officers. A chatbot that is accurate 92 percent of the time is a massive success in retail customer support. If an e-commerce bot gives the wrong return window to eight out of every hundred shoppers, a manager issues a coupon and moves on.

In an election, a 92 percent accuracy rate is a disaster.

Elections run on razor-thin margins. In Quebec's 2022 general election, several ridings were decided by just a few hundred votes—and in at least one Montreal riding, by a handful of ballots after a judicial recount. When thousands of voters interact with a faulty automated tool, even a small error rate affects enough people to alter local outcomes or spark legal challenges.

I've watched organizations rush to slap conversational interfaces onto complex regulatory frameworks for years, and they almost always make the same three mistakes:

  1. Treating probabilistic text as verified truth. Generative models predict the next likely word. They don't verify facts against a legal statute unless they're strictly constrained by a retrieval system—and even then, they still drift.
  2. Failing to program hard refusal boundaries. A civic bot needs to say "I don't know, call this phone number" the second a query steps outside a pre-approved script. Most deployments are tuned to be helpful, which makes them guess.
  3. Ignoring regional and linguistic nuance. Quebec's political and administrative vocabulary has specific legal meanings in French and English. Translating concepts loosely between languages or mixing municipal, provincial, and federal rules is a classic failure mode.

What MPs Heard About Election Interference and AI Risks

The testimony regarding Quebec voters didn't happen in a vacuum. Canadian lawmakers on the Standing Committee on Access to Information, Privacy and Ethics—alongside related parliamentary reviews of foreign interference and emerging tech—have been grilling experts on how artificial intelligence threatens democratic processes.

Most media coverage obsesses over flashy deepfakes. Everyone worries about fake audio clips of party leaders or fabricated videos circulating on social media the night before a vote.

Those threats are real, sure. Look at the robocall scandal in New Hampshire during the 2024 U.S. primaries, where an AI-cloned voice impersonated Joe Biden to tell Democrats not to vote. Lawmakers everywhere took notice.

Yet the mundane, boring failures are arguably more dangerous because they wear an official badge. Voters expect trolls on social media to lie to them. They don't expect a tool designed to help them navigate a provincial election to hand them bogus rules about residency requirements or mail-in ballots.

Witnesses before Parliament pointed out that bad election data doesn't even require a foreign adversary. Poorly tested civic tech, third-party news bots scraping outdated web pages, or search engine AI summaries pulling old rules from the 2018 election cycle can confuse tens of thousands of citizens without a single hacker lifting a finger.

How Search Engines and Third-Party Bots Compound the Mess

Even when an official election body builds a rock-solid website, voters rarely start there. They type a question into a search bar or ask a general-purpose AI assistant on their phone.

That's where things get ugly fast.

General-purpose chatbots are trained on massive snapshots of the open internet. In that training data, rules from a 2014 municipal election sit right next to rules from a 2021 federal election and a 2022 provincial vote. Without strict live grounding, an AI assistant routinely mashes those rulebooks together.

For example, federal elections in Canada allow vouching under specific conditions, while provincial rules in Quebec operate under the Quebec Election Act with distinct identification and domicile requirements. Ask a generic chatbot whether your neighbor can vouch for you at a polling station in Montreal, and there's a very real chance it'll cite federal Elections Canada rules instead of Quebec's provincial law.

You show up at the polling station expecting to vote. The deputy returning officer tells you no. You leave angry, convinced the system is rigged or broken, when really a language model just blended two jurisdictions together.

Why Guardrails Alone Won't Fix Election Chatbots

Whenever a civic AI project blows up in the headlines, the standard corporate response is to promise "better guardrails" and "more human oversight."

Honestly, that usually translates to adding a tiny disclaimer at the bottom of the chat window that nobody reads. Telling users "AI can make mistakes, please verify important information" is a legal liability shield for the developer, not a solution for the voter. If a citizen has to cross-check every single sentence the chatbot outputs against the official Election Act, what was the point of building the chatbot in the first place?

To make automated tools safe for public administration, agencies have to ditch open-ended generative chat for high-stakes queries. Deterministic decision trees—boring, old-school software where a human lawyer writes every single response branch—beat generative AI every time when accuracy is non-negotiable.

If a government agency insists on using modern language models, the architecture has to lock the model down completely:

  • Zero open-web generation. The system must only pull text verbatim from verified, current-cycle statutory documents.
  • Mandatory citation links. Every answer must link directly to the exact page on the official election authority's domain where the rule lives.
  • Instant escalation triggers. Any question involving disputed eligibility, moved addresses, accessibility accommodations, or lost ID needs to bypass the bot and connect straight to a trained human operator.

How Voters Can Protect Themselves From Bad AI Election Info

Until tech companies and public agencies stop treating voters like beta testers, you have to assume any conversational AI tool is guessing when it answers questions about voting rules.

Skip the chatbot summary entirely when checking how, where, or when to cast a ballot. Go straight to the primary domain for your jurisdiction—such as Elections Canada for federal votes or Élections Québec for provincial ones—and use their direct postal code lookup tools rather than conversational search widgets. Look up the physical address of your returning office and call their local phone line if your living situation, ID, or registration status has any quirks at all.

IL

Isabella Liu

Isabella Liu is a meticulous researcher and eloquent writer, recognized for delivering accurate, insightful content that keeps readers coming back.