On October 8, McGill political scientist Aengus Bridgman told a House of Commons committee that an estimated 40 per cent of Quebec voters asked AI chatbots for election information during the October 5 provincial vote. Many, he said, received false answers. His co-witness Taylor Owen added that in the campaign's final three weeks, about 39 per cent of ChatGPT's voting advice traced to an error-filled hobby website that had itself been built with AI, according to the Canadian Press report. Owen stressed that no one acted with malicious intent.
That last point is the policy problem. Canada has just rewritten its election-falsehood rules, and they were built for people who lie on purpose.
The case for tougher rules
The strongest argument for regulation is simple: voting information is not ordinary content. A wrong polling date or eligibility rule can cost someone their vote, and a chatbot answers each person privately, so no journalist, rival campaign or fact-checker sees the error to correct it. The Canadian Press report also describes a chatbot that sent prospective volunteers in a federal byelection to the wrong campaign office, and ChatGPT reversing itself on whether to recommend a party. Bridgman suggested election queries should face standards similar to those applied to safety-related ones. Calkins, a Conservative MP, said that if this can happen by accident, deliberate manipulation could be worse. Those are fair concerns.
What C-25 does and does not reach
The Strong and Free Elections Act (Bill C-25) received royal assent on June 18, 2026. According to the enacted text, new section 482.01 covers any person or entity that makes or publishes a statement it knows is false or misleading about who may vote, how to register, where, when or how to vote, how votes are counted, or results, and does so "with the intention of affecting the results or disrupting the conduct of an election." Section 480.1 separately bars image or voice imitations of election figures made with intent to mislead, with a parody and satire exception, as BetaKit summarized on June 19.
Both provisions turn on knowledge and intent. A chatbot that summarizes a stale web page and gets a polling rule wrong has neither. Nor does the company that deployed it, absent evidence it meant to disrupt the vote. In the excerpt of the Act I reviewed, I found no platform duties such as correction, labelling or reporting obligations. This is my reading of the statute, not a court ruling; no tribunal has tested the point. But the hearing suggests the gap is real, and the Canadian Press report notes that Owen said elections law should address how chatbots present election information.
The upstream cause is access
Bridgman identified a specific mechanism: almost all Quebec newsrooms have blocked AI chatbots from collecting their content. Publishers have understandable reasons, since they object to uncompensated use of their work. But the consequence, in a campaign, is that chatbots had little current, professionally edited reporting to draw on and fell back on whatever was crawlable, including an AI-built hobby site. The Centre for Media, Technology and Democracy, which monitored the campaign through the Canadian Digital Media Research Network, had flagged generative AI as a distinguishing factor of this election before it began.
This suggests the failure was largely a sourcing problem, not a speech problem. This analysis rests on the ChatGPT findings reported to the committee.
A proportionate response
The tempting reaction is a new offence for inaccurate AI election answers. That would be a mistake. Strict liability for statements about elections would push providers toward blanket refusals on civic questions, which hurts the voters who most need a quick answer, and it would hand regulators a lever over what lawful, if imperfect, speech software may produce. Canada has also just legislated; stacking rules before measuring C-25's effect is premature.
Better options are narrower and mostly non-coercive:
- Authoritative-source defaults. Providers should route voting-logistics questions to Elections Quebec or Elections Canada pages and show them prominently. This is voluntary-code territory, and the committee's study is a natural venue for pressing firms to commit.
- Licensed access for election periods. Publishers and AI firms could negotiate time-limited, low-friction access to election reporting, so the best-sourced material is available when stakes are highest.
- Independent audits. The McGill method, repeated at each election, gives Parliament evidence rather than anecdote. Transparency reporting is cheaper and less speech-restrictive than liability.
- Targeted review of C-25 after the next federal election. If audits show persistent errors on core voting facts, a narrow duty limited to voting logistics, with a notice-and-fix safe harbour, would be defensible.
Bottom line
Quebec offers a rare, measured case of AI mediating a real election. The evidence points to a design and access failure that existing law, aimed at knowing deceivers, does not capture. The right response is to fix the information supply and measure results before writing a new speech offence.