Have you used AI to inform yourself about the latest election? Then you may be getting false information, focused on select candidates, and from unreliable sources, researchers say.
A new analysis by academics at University College Dublin (UCD) and University of Strathclyde in Glasgow found that when tasked with questions about the upcoming Irish byelections, different AI chatbots used very different sources, from The Irish Times and RTÉ to Wikipedia and Gript.
They also gave special focus to some candidates and provided wrong answers about essential information such as where to vote and who was on the ballot.
“Broadly speaking, most of the answers were good,” said Dr James Cross, associate professor at UCD’s School of Politics and International Relations and director of the Connected Politics Lab. “But when it goes wrong, it goes really wrong.”
A recent Reuters Digital News Report found that AI chatbots are an increasingly popular source of news – with 7 per cent of audiences in 48 countries surveyed worldwide using AI to get their news every week.
AI use was even higher for young people, with one in six people under 25 using chatbots to get their news. But how reliable is that news?
In advance of the Dublin Central and Galway West byelections, researchers put 194 questions to four different models – Claude, Gemini, ChatGPT and Grok – on two different occasions before polling day.
By simulating a typical citizen’s interactions with popular AI chatbots, they found that some questions asking for essential voter information, such as who was running and where to vote, were answered incorrectly.
In one instance, ChatGPT left six names off the ballot when asked who was running in both byelections. In another, Gemini falsely claimed Gerry Hutch won a seat in the last general election.
AI chatbots consistently concentrated attention on a few candidates, largely ignoring others.
Depending on which AI chatbot is used, a candidate could be the most mentioned on one model but an “also-ran” in another.
For example, Gerry Hutch was the third most mentioned of 14 Dublin Central candidates in answers on Grok, but in the bottom seven on Claude.
Part of the reason for the widely varied responses was the sourcing of AI information.
“When you ask AI a question, it performs a web search and selects what sources to use. When you hand that curation over to the language model, it changes the type of information you typically receive,” said Dr Eoghan Cunningham, lead author of the study and a postdoctoral researcher at University College Dublin.
The researchers analysed the number of times AI cited links to the sources of its information and found the chatbots rely mostly on Irish news organisations – but they vary.
Claude relied most on RTÉ (20 per cent of cited links) whereas ChatGPT relied most on The Irish Times (17 per cent).
By contrast, Google’s Gemini relied heavily on conservative Irish news website Gript.ie (7 per cent).
Significant amounts of information were also gleaned from unverified sources such as Wikipedia, which was cited by ChatGPT 16 per cent of the time. As Wikipedia is crowdsourced, bad actors could potentially change Wikipedia entries on rival candidates and thus have that reflected in AI answers.
Social media was also used as a source. Grok cited Facebook for 6 per cent of cited answers and Instagram for 5 per cent.
The variety of the sourcing raises questions about how users can assess how much they should trust AI chatbot answers.
“We engage with different sources differently, altering our expectations for a news website versus social media,” said Cunningham. “But they’re all mixed here in the chatbot answers and so it’s difficult to calibrate your expectations of trust appropriately.”
In general, chatbots vary in how transparent they are about what they searched and the sources they cite.
Gemini contains small print on its search that says “Gemini is AI and can make mistakes, including about people”. Claude states “Claude is AI and can make mistakes. Please double-check responses” and ChatGPT says “ChatGPT can make mistakes. Check important info.”
Grok just refers users to its terms and conditions.
When using AI for information, AI “hallucinations” have been flagged as a significant risk – when AI makes up information or guesses answers in order to please the user.
“It’s always guessing”, said Cunningham. “An area like an Irish byelection is an example of where it would be guessing a lot of the information.”
As well as mistakenly giving users false information, there is also potential for mass manipulation the more citizens start to use a single AI source for their information.
“We can show different chatbots are relying more on some news sources over others,” said Cross. “The danger is because users don’t have transparency over the sourcing of information these AI chatbots are using, that search process could be manipulated toward specific sources and users would not know it.
“When you think about roles of algorithms in social media or web searches, and how there is potential to manipulate that, this is the next iteration of that.”
