Paragraph 1
On a weekday morning in Ottawa, a parliamentary committee sat down to hear something that should concern every Canadian who cares about democracy: the advice many voters are getting from artificial intelligence may be quietly, systematically wrong. The hearing was not about a new bill or a foreign interference plot. It was about the humble chatbot—the friendly, endlessly patient digital assistant that millions now consult for everything from recipes to relationship advice. Increasingly, Canadians are also asking chatbots how to vote, where to vote, and what each party stands for. And according to researchers from McGill University who have spent months tracking this phenomenon, the answers often come from sources that are neither reliable nor even human. The committee heard that during the recent Quebec provincial election, a striking number of voters turned to AI chatbots for electoral guidance. What they received in return, researchers said, was frequently a patchwork of outdated articles, blocked newsroom content, and in some cases, information generated entirely by other AI systems—some of it riddled with errors. The witnesses did not claim that anyone was deliberately trying to deceive voters. There were no dark foreign actors pulling strings in a basement. But the testimony painted a picture of an information ecosystem that has quietly changed under our feet, with serious implications for how Canadians understand their choices and exercise their most fundamental democratic right. The hearing was about transparency, accuracy, and the strange new role of machines in civic life. It was also a reminder that technology, no matter how intelligent it seems, is only as reliable as the information it is allowed to read.
Paragraph 2
Aengus Bridgman, a McGill University political scientist who studies social media and political behaviour, offered the committee a startling figure: an estimated 40 per cent of Quebec voters used AI chatbots at some point during the provincial election campaign. That is not a fringe phenomenon. It is not a niche gadget for tech enthusiasts. Nearly half of the electorate in one of Canada’s largest provinces turned to algorithms for help understanding a vote. Bridgman explained that people use these tools because they are fast, convenient, and conversational. Rather than scrolling through dozens of websites, comparing platforms, or reading lengthy news analyses, a voter can simply type “Who should I vote for in my riding?” and receive a neatly formatted answer. But there is a critical flaw. Most Quebec news outlets, like many publishers around the world, have blocked chatbots from accessing their content. They did this for reasonable commercial reasons: newspapers invest time and money in journalism, and they do not want their work scraped for free by billion-dollar tech companies. Yet the result, Bridgman said, is that the very sources best equipped to provide accurate, up-to-date, and deeply reported election information are invisible to the chatbots. The digital divide between journalism and AI has created a vacuum. And vacuums in the information ecosystem do not remain empty for long. They fill with whatever is left—sometimes old articles, sometimes opinion pieces, sometimes content that was never written by a human being at all. Voters who asked their chatbot for help were not being guided to the best available information. They were being guided to whatever the algorithm could find outside the locked doors of professional newsrooms.
Paragraph 3
The problem was even more concrete than that. Taylor Owen, an associate professor at McGill’s Max Bell School of Public Policy, presented findings from the researchers’ analysis of chatbot responses during the final three weeks of the Quebec election. What they discovered was both fascinating and alarming. A full 39 per cent of the voting advice provided by ChatGPT—the most widely used AI chatbot in the world—came from a single website that was not a news outlet, not a political party, and not a government source. It was a hobby website, built by an individual and filled with content that had itself been generated by artificial intelligence. The site contained errors, some of them significant, yet the chatbot apparently treated it as a trusted source and passed those errors along to users. Owen was careful to note that the researchers found no evidence of malicious intent. Nobody appeared to be deliberately manipulating the system to influence the election. The website creator was likely just someone experimenting with AI tools, not a mastermind trying to sway voters. But from the perspective of someone relying on ChatGPT for election information, the distinction hardly matters. False information is false information, regardless of whether it spreads through a deliberate scheme or through a chain of careless algorithms. Owen told the committee that it is likely a large number of Quebec voters received false information about the election in those crucial final weeks. The timing matters, because that is when many people make up their minds, when early voting is underway, and when last-minute decisions can tip electoral outcomes. A machine-generated mistake, repeated thousands of times through a chatbot interface, can become a very real force in a very real election.
Paragraph 4
These findings point to a deeper structural problem in the relationship between artificial intelligence and public knowledge. Chatbots are trained on enormous amounts of text from the internet. They do not have a concept of truth. They have patterns, probabilities, and word associations. When you ask a chatbot a question, it does not know the answer in the way a human journalist or librarian might. It predicts the sequence of words that is statistically most likely to satisfy the user. That prediction is heavily influenced by what data the chatbot can access during training and, in some cases, during live retrieval. If news publishers block their content, the chatbot loses access to the most reliable stream of current information. It is then forced to rely on older sources, personal blogs, message boards, and—in a particularly strange and unsettling twist—other AI-generated content. This creates a feedback loop. An AI reads a website written by another AI. The first AI repeats that information to a voter. Another website is then created by AI summarizing what the first AI said. The loop continues, growing more distorted with each turn. Researchers sometimes call this “model collapse” or “AI cannibalism,” and it is a growing problem across the digital ecosystem. In the context of an election, the stakes are not academic. When a voter asks a chatbot whether a candidate voted for a certain bill, or where to find a polling station, or what a party’s environmental platform actually says, they deserve an answer grounded in verified facts. Instead, they may receive a smooth, confident, and completely hallucinated response. The chatbots rarely say “I don’t know.” They are designed to answer, not to admit ignorance. And so they string together plausible-sounding sentences from whatever fragments of data they found, without any ability to judge whether those fragments are true.
Paragraph 5
To understand the human consequences, imagine a first-time voter in Montreal, a student feeling overwhelmed by the noise of campaign advertising, wanting to make a responsible choice. They open their phone and ask a chatbot which party supports affordable housing and what the candidates in their riding have promised. Within seconds, they receive a courteous, structured answer. It sounds authoritative. It references policies and positions. It is completely wrong in several important ways. The voter does not know that the chatbot drew from a hobby website filled with AI-generated mistakes. They do not know that the real news sources could have corrected those mistakes but were blocked from the system. All they know is that they asked a smart machine for help and received a confident answer. This is the quiet danger of AI in civic life. It does not need to be malicious to be harmful. It just needs to be plausible. Researchers are especially concerned about younger voters, who are more likely to trust digital tools and less likely to consult traditional news outlets. But older Canadians are also turning to chatbots in growing numbers, seeking help with complex information. The same dynamics affect everyone: the illusion of omniscience, the absence of caveats, the smooth prose that hides a complete lack of comprehension. When a human journalist makes a mistake, there are corrections, apologies, and an editorial process designed to catch errors before publication. When a chatbot makes a mistake, there is nothing but another question, another answer, another confident fabrication. The committee heard that this creates a profound challenge for democracy itself. Democratic elections depend on the idea that citizens can inform themselves about the issues and make rational choices. If their primary source of information is an algorithm that cannot tell truth from an AI-written hobby blog, then the very foundation of informed consent has been eroded.
Paragraph 6
The McGill researchers did not simply present problems; they also suggested paths forward. One key recommendation is increased transparency from AI companies about their data sources and about the limitations of their models. Users should be told when a chatbot has generated a response based on uncertain or incomplete information. There should be clear disclaimers that AI chatbots are not designed for voting advice and cannot be held accountable for the accuracy of their answers. Another approach is to encourage collaboration between news organizations and AI platforms. Instead of blanket blocking, media outlets could negotiate licensing agreements that allow chatbots to access quality journalism while fairly compensating publishers. This is already happening in some countries, but Canada has been slower to adapt. The committee also heard that media literacy is more important than ever. Canadians need to understand what chatbots are, how they work, and why they are fundamentally different from a search engine or a newsroom library. A chatbot is not a trustworthy guide; it is a predictive text generator. It can be useful for brainstorming, for drafting a letter, or for explaining a concept in simple terms. But it should not be the deciding voice in a democratic election. The testimony ended with a simple but powerful message: there is no replacement for reliable, human-produced journalism. AI may be able to process enormous amounts of data at lightning speed, but it cannot attend a campaign rally, question a minister, verify a source, or feel the weight of responsibility that comes with informing the public. As Canada looks toward the next election, the challenge is not to ban AI or to pretend it does not exist. The challenge is to build a system where technology serves democracy instead of quietly undermining it. That means protecting the role of professional journalism, demanding accountability from tech companies, and ensuring that voters never have to guess whether the helpful little chatbot in their pocket is telling them the truth. The committee’s hearing was an important first step. The real work lies ahead.