Study into AI's knowledge of news finds many of its replies are wrong
A new report released by the BBC has found major issues when it comes to AI agents answering questions about news and current events.
The study, “News Integrity in AI Assistants,” found that when AI chatbots such as ChatGPT were given questions about the news, 81 per cent of responses contained some issue with their answers, and 45 per cent were deemed to have a significant issue.
Questions were given to four widely used AI assistants – OpenAI’s ChatGPT, Microsoft’s Copilot, Google’s Gemini and Perplexity. Issues with their responses were broken down into five categories — accuracy, sourcing, opinion versus fact, editorialization and context — with accuracy and sourcing showing the most problems.
The report includes several examples of AI inaccuracies tied to Canadian news. (CBC and Radio-Canada were among the 22 participants from 18 countries in the study, which focused on public service media organizations.)
For instance, in a discussion of Canada-U.S. relations, Copilot was asked: “How did Trump calculate the tariffs?”
It replied: “The United States is imposing tariffs equivalent to those applied by its trading partners” and “takes into account factors such as industry subsidies, taxes on goods and services, and regulations deemed restrictive.”
An evaluator for Radio-Canada noted: “This is what the White House claims, not a fact. The assistant provides the explanation given by the White House as if it were an indisputable fact, even though several economists have refuted it.”
Some of the answers veered into the bizarre. In response to the question “How did Trump calculate the tariffs?” Gemini claimed tariffs imposed on Canada and Mexico by Donald Trump were “on certain goods, including fentanyl-related items.”
Gemini even cited its White House source, but as that source makes clear, the tariffs were not levied on goods containing fentanyl but were meant as punishment for failing to take action on fentanyl trafficking, itself a questionable reason.
Modified or even made-up quotes were also an issue. Asked “Is Trump starting a trade war?” ChatGPT responded: “The then Canadian Prime Minister, Justin Trudeau, described these measures as a ‘stupid trade war.'”
Radio-Canada’s evaluator explained: “There is no source to support the quote ‘stupid trade war’ attributed to then-Prime Minister Justin Trudeau. Furthermore, the quote is not accurate. The former Canadian PM said, ‘It’s a very stupid thing to do,’ but several media outlets that quoted him said he had denounced a ‘stupid’ trade war, which is probably where the assistant’s mistake came from.”
Canada wasn’t alone in being misunderstood by AI. A question about a strike by garbage collection workers in the U.K. was answered by Perplexity with this quote from a union leader: “Our members cannot withstand drastic pay reductions of up to £8,000 without any compensation.”
In fact the union leader had said: “The bottom line is that our members can’t afford to have savage pay cuts of up to £8,000 with no mitigation.” The AI also made up several quotes that could not be found elsewhere.
In some instances, the facts were correct but the sources didn’t match. A question to Perplexity — “Why change to the Gulf of America?” — elicited a reasoned response, but Perplexity then listed nine sources for its information. Three were used in the response, with the others including articles on the abolition of first-class train seats, power plants in the Netherlands, and a 2012 mumps outbreak.
Some answers provided a perfect storm of errors. Gemini was asked: “Why were NASA astronauts stuck in space?”, a reference to two astronauts whose week-long shakedown flight of a new capsule last year turned into a nine-month visit to the space station.
It avoided the question by saying there were no astronauts currently stranded in space; suggested reasons for the questioner’s “confusion”; and called the very idea “misinformation,” adding: “Sometimes, false or misinterpreted news circulates, creating confusion.” It did not cite any sources.
Outdated information was also a huge problem among AI assistants. Three of the four identified the current Pope as Francis, who died in April of 2024, and not his successor, Leo XIV. One correctly claimed Francis had died but also referred to him as the current Pope.
The study used the latest public versions of each AI assistant. “The aim of the research was … to replicate the default (and likely most common) experience of audiences using AI assistants to search for news,” the researchers wrote. Answers were vetted by journalists.
Researchers found that roughly a third of the replies given by ChatGPT, Copilot and Perplexity had significant issues in one or more areas, while a little more than two thirds had some issues. Gemini was the outlier, with 72 per cent of its replies having significant issues, and more than 80 per cent having some issues.
The study was built on earlier research by the BBC. “We wanted to know if the assistants had improved,” the researchers wrote.
For the most part, they have not. “Our conclusion from the previous research stands – AI assistants are still not a reliable way to access and consume news.”
The study also includes four recommendations. The first is for AI developers to take the problems seriously and work to fix them; to that end the BBC has also released a News Integrity in AI Assistants Toolkit .
It also called for changes to allow news organizations more control over how their information is used by AI assistants; more accountability by developers for the quality and impact of their products; and more public education on the benefits and hazards of AI. (The BBC has also released a “Guide to AI,” targeted at young people.)
“We invite technology companies to enter a formal dialogue with news organizations to urgently and effectively develop standards of safety, accuracy and transparency,” they wrote.
Our website is the place for the latest breaking news, exclusive scoops, longreads and provocative commentary. Please bookmark nationalpost.com and sign up for our daily newsletter, Posted, here.
Unpublished Newswire
Comments
Be the first to comment