Veracity: a transparent fact-checker

In 5 seconds The innovative AI tool is now available to the public—and Canada’s Senate is already using it.
Veracity is an open-source model that other teams can build on to create similar tools for different electoral or linguistic contexts.

A headline, an excerpt, a single sentence: feed any of those into the fact-checking tool Veracity, and its artificial intelligence will provide not only a reliability score but also a plain-language explanation and full list of sources.

This transparency is what sets the made-in-Montreal software apart from many other AI platforms, which will spit out an answer without explaining how they got there.

The tool was designed by a research team co-led by Jean-François Godbout, a political science professor at Université de Montréal, researcher at IVADO – consortium for research, training and knowledge mobilization in artificial intelligence – and associate academic member of Mila.

Developed by the IVADO research cluster on AI implementation and governance, Veracity was built not to be blindly trusted but to encourage users to verify information themselves. 

In May 2026, it received its first institutional endorsement when Canada’s Senate added it to the resources available to senators and staff for assessing the reliability of incoming information.

How it works

When a statement is submitted, Veracity first taps into its internal training data, then consults external sources, weighting each by credibility. Major media outlets and wire services carry more weight than obscure websites, for example.

At the end of the process, Veracity gives the statement a reliability score of 0 to 100 per cent, providing a rationale and a complete source list.

It will never label a statement as unequivocally false. Instead, it will indicate that the information could not be verified, that sources are unreliable or that the claim is likely inaccurate.

“We wanted to qualify our judgments and not be too adamant, because we want to encourage people to check the information themselves,” Godbout explained.

This caution is central to Veracity’s philosophy.

“When you ask Claude or ChatGPT to fact-check something, you have no idea what sources they used,” Godbout noted. With Veracity, the sources are listed, their credibility rating is displayed, and the reasons for the reliability score are clear. Users can examine the evidence themselves.

The design is also grounded in Godbout’s research, which shows that large language models (LLMs) can help correct misinformation by maintaining a calm, factual tone. “If you’re discussing an issue with a friend, they might get upset,” he pointed out. “An LLM can present the facts coherently without losing its cool.”

Veracity is an open-source model that other teams can build on to create similar tools for different electoral or linguistic contexts.

Jean-François Godbout, professor in the Department of Political Science at Université de Montréal

The limits of transparency

The team doesn’t claim to have eliminated every blind spot. Veracity runs on Llama 3.3, an LLM developed by Meta. The team didn’t build the core engine from the ground up, but adapted and trained it with a custom instruction sequence for evaluating sources and claims.

“We have no control over the model itself, because it’s an American LLM,” Godbout said.

Veracity doesn’t always use the sources one might expect. In one demo, Godbout was surprised to see Reddit listed. The limitations are even more apparent in French: in one French-language test, Wikipedia ranked as the top credible source, while the same query in English pulled up more relevant and diverse references.

The problem is that there are fewer French-language sources and they are less well-indexed. Godbout calls this a structural limitation rather than a flaw in the tool itself.

Results can also vary slightly when the same statement is submitted more than once, and Veracity can struggle with questions that require a specific timeframe when none is provided.

The credibility ratings assigned to sources are currently static, based on prior research. The team is working to make them dynamic, so they can evolve with new content and cross-site references.

These limitations are one reason why Veracity frames its output as an assessment for users to examine, not a verdict. To address the problems, the developers want to gradually reduce their dependence on the American model they don’t control.

Quebec is not immune

Research by Godbout and his team challenges the notion that Quebec’s linguistic and cultural distinctiveness offers a shield against disinformation.

In a master’s thesis supervised by Godbout, political science student Camille Thibault analyzed a large corpus of political messages posted on Twitter. The findings confirmed that disinformation is a real issue in Quebec as well.

Godbout remains cautious about the conclusions to be drawn from these findings. “Just because false content has been published doesn’t mean it was widely viewed, let alone that it changed a single voter’s behaviour,” he said.

Even so, the research provides data that can inform initiatives to protect democratic debate. Quebec has moved to protect the integrity of provincial elections with the Act to amend the Election Act, enacted on May 30, 2025, which makes it an offence to knowingly disseminate false information with the intent of influencing or disrupting an election.

First institutional recognition

After Godbout presented Veracity to the Canadian Study of Parliament Group in April, the Office of the Law Clerk and Parliamentary Counsel for the Senate added the tool to the Senate’s internal resources to help senators and staff scrutinise the accuracy of claims they encounter.

That recognition hasn’t eased Godbout’s concerns about misinformation. Its spread is gradually eroding trust in information, and ultimately in institutions.

“If we move into an environment where we can never be sure of the quality of information, it will affect the resilience of our democratic institutions,” he warned.

Veracity doesn’t claim to solve this problem by itself. But it offers a way to chip away at it—one transparent fact-check at a time.

Share

You may also like