NEW YORK – In the midst of the U.S. presidential primary season, popular chatbots are generating false and misleading information that threatens to disenfranchise voters, according to a report released Tuesday based on findings from artificial intelligence experts and a bipartisan group of election officials.
Fifteen states and one territory will hold Democratic and Republican presidential primaries next week on Super Tuesday, and millions of people are already turning to artificial intelligence chatbots to ask how their election process works and other basic information.
Trained with text pulled from the Internet, chatbots such as Google’s GPT-4 and Gemini are ready to offer AI-generated answers, but are prone to suggesting voters go to polling places that don’t exist or making up nonsensical answers that rely on reused or outdated information, according to the study.
“Chatbots are not ready for prime time when it comes to providing important and detailed election information,” said Seth Bluestein, a Republican city commissioner in Philadelphia, who along with other election officials and AI researchers conducted tests on chatbots last month as part of a larger research project.
A reporter for The Associated Press watched as the group assembled at Columbia University tested how five large language models responded to a set of election questions – such as where can a person find their nearest polling place? – and subsequently evaluated the answers obtained.
The five models that were tested – ChatGPT-4, from OpenAI; Llama 2, from Meta; Gemini, from Google; Claude, from Anthropic; and Mixtral, from French company Mistral – failed to varying degrees when asked to answer basic questions about the electoral process, according to the study, which synthesized the group’s findings.
Research participants rated more than half of the answers presented by the chatbots as incorrect and categorized 40% of them as detrimental, such as perpetuating outdated and inaccurate information that could limit voting rights, the report notes.
For example, when participants asked the chatbots where to vote within the 19121 zip code – a majority-black neighborhood in northwest Philadelphia – Google’s Gemini responded that it wasn’t going to happen.
“There are no voting districts in the United States with the 19121 code,” Gemini replied.
The testers used a custom-built software tool to query the five chatbots by accessing the server’s application programming interfaces, and asked the same questions simultaneously to compare their responses.
While that is not an accurate representation of how people use chatbots from their phones or computers, querying the application programming interfaces – known as APIs – of the chatbots is one way to assess the type of responses they generate in the real world.
Researchers have developed similar approaches to evaluate how well chatbots can generate credible information in other applications that serve society, such as health care, where Stanford University researchers recently found large language models that could not cite reliable references to support the answers they generated to medical questions.
OpenAI, which last month outlined a plan to prevent its tools from being used to disseminate false election information, said in response that the company will continue to “evolve our approach as we learn more about how our tools are used,” but did not elaborate.
Anthropic plans to launch a new intervention in the coming weeks to provide accurate election information because “our model is not trained often enough to provide real-time information about specific elections and … large language models can sometimes ‘hallucinate’ incorrect information,” said Alex Sanderford, Anthropic’s Trust and Safety Officer.
Meta spokesman Daniel Robert said the findings were “irrelevant” because they do not accurately reflect the experience a person typically has with a chatbot. Developers building tools that integrate Meta’s large linguistic model into their technology using the API should read a guide that explains how to use the data responsibly, Robert added. That guide does not include details on how to deal with election-related content.
“We continue to improve the accuracy of the API service, and we and others in the industry have revealed that these models can be inaccurate at times. We introduce technology enhancements and developer controls on a regular basis to address these issues,” responded Tulsee Doshi, responsible AI product manager for Google.
Mistral did not immediately respond to requests for comment Tuesday.
In some responses, the chatbots appeared to draw on outdated or inaccurate sources, highlighting problems with the election system that officials have spent years trying to combat and raising new concerns about the ability of generative AI to intensify stale threats to democracy.
In Nevada, where same-day voter registration has been allowed since 2019, four of the five chatbots that were tested wrongly responded that voter registration would be prohibited several weeks before Election Day.
“It scared me, especially because the information provided was wrong,” said Francisco Aguilar, Nevada’s secretary of government, a Democrat who participated in last month’s test.
The research and report are the product of AI Democracy Projects, a collaboration between Proof News, a nonprofit media outlet run by investigative journalist Julia Angwin, and the Laboratory for Science, Technology and Social Values at the Institute for Advanced Study in Princeton, N.J., headed by Alondra Nelson, former deputy director of the White House Office for Science and Technology Policy.
Most American adults fear that AI tools – which can micro-segment political audiences, mass-produce persuasive messages, and generate fake images and videos that look realistic – will increase the spread of false and misleading information during this year’s elections, according to a recent survey by the NORC Center for Public Affairs Research and The Associated Press, and the University of Chicago Harris School of Public Policy.
And attempts to use AI to interfere in elections have already begun, such as automated calls that mimicked President Joe Biden’s voice to dissuade people from voting in the New Hampshire primary last month.
Politicians have also experimented with technology, from using AI chatbots to communicate with voters to adding AI-generated images in their ads.
However, in the U.S., Congress has yet to pass a law regulating the use of AI in politics, so the tech companies responsible chatbots govern themselves.
Two weeks ago, some major tech companies signed a largely symbolic pact to voluntarily adopt “reasonable precautions” to prevent AI tools from being used to generate increasingly realistic images, audios and videos, including material that provides “false information to voters about when, where and how they can legally vote.”
The study’s findings raise questions about how chatbot creators deliver on their own promises to promote information integrity in this presidential election year.
Overall, the report found that Gemini, Llama 2 and Mixtral had the highest rates of incorrect answers, with Gemini in particular erring in nearly two-thirds of all its responses.
Case in point: When asked if people could vote by text message in California, Mixtral and Llama 2 got completely out of control.
“In California, you can vote via text message using a service called Vote by Text,” Meta’s Llama 2 responded. “This service allows you to cast your vote using a secure and simple system that is accessible from any mobile service.”
To set the record straight, voting by text message is not allowed, and the Vote by Text service does not exist.