
Recent bot campaign backing Poilievre shows AI easily accessible for political messaging: report
CBC
A suspected bot campaign surrounding a recent Pierre Poilievre event shows that generative artificial intelligence (AI) tools are easily accessible to anyone looking to influence political messaging online, researchers have found.
In July, the social media platform X was inundated with posts following the Conservative leader's tour of Northern Ontario.
The posts claimed to be from people who attended Poilievre's event in Kirkland Lake, Ont., but were actually generated by accounts in Russia, France and other places, and many of them had similar messaging.
Researchers at Concordia University and the University of Ottawa recently conducted several tests to see if they could get five generative AI tools to create similar political messaging seen during the July bot campaign.
The researchers asked the freely available AI platforms to create 50 different statements to describe rallies held by a number of Canada's political leaders: Poilievre, Prime Minister Justin Trudeau, NDP Leader Jagmeet Singh, Bloc Québécois Leader Yves-François Blanchet and Green Party Leader Elizabeth May
All but one of the platforms generated the requested political messaging when prompted.
"It was simple, it was quick and what that tells us is there is a gap in our regulatory system right now," said Elizabeth Dubois, a professor at the University of Ottawa and one of the leads on the project.
"Platforms do sometimes say, 'We don't want to be used in an electoral context. We don't want you to use our tool for politics. It's against our terms.' But the reality is, most of the tools let you do it anyway."
Dubois said such findings are concerning because the use of AI tools to generate political messaging could compromise political campaigns and electoral fairness.
The Conservatives denied any involvement in the Kirkland Lake bot campaign. A separate report from Toronto Metropolitan University's Social Media Lab concluded the campaign was likely the work of amateurs.
Dubois said some sort of automation was likely used in the Kirkland Lake incident, but it's unclear what specific tool generated the messages. The report did note that some of the descriptive language ChatGPT provided in its response to the test prompts, such as "electric," "buzzing" and "palpable," were similar to what was seen in the July incident.
When Dubois and her colleagues ran their generated messages through X's three AI text detection tools, all three failed to determine if the messages were AI generated. But Dubois said she wasn't surprised by that finding.
"These kinds of tools are not particularly useful," she said.
For X specifically, Dubois said part of the issue is that the detection tools have limited data to work with because of the platform's 280-character limit.