Over 50 countries representing half the world’s population are holding elections this year — and experts are warning people against turning to AI chatbots for election information.
Top AI models from OpenAI, Google, Meta, Anthropic, and Mistral AI “performed poorly on accuracy” and other measures in a new report from the AI Democracy Projects released this week. Conducted by more than 40 U.S. state and local election officials alongside AI researchers and journalists, the study tested a range of large language models (LLMs), including OpenAI’s GPT-4, Google’s Gemini, Meta’s Llama 2, Anthropic’s Claude, and Mistral AI’s Mixtral. Among its conclusions: more than half of the responses generated by the models contained inaccurate responses to election questions.
Expert testers posed 26 common voting questions to LLMs, then ranked rated 130 responses for bias, accuracy, completeness, and harmfulness. The study notes that the “small sample” of responses “does not claim to be representative,” but that the group hopes its results show the limitations — and dangers — of AI chatbots in giving voters election information.
Overall, the study found 51% of the chatbots’ responses were inaccurate, 40% were harmful, 38% were incomplete, and 13% were biased.
In one example, OpenAI’s GPT-4 responded that voters could wear a MAGA hat — a red cap affiliated with U.S. presidential candidate Donald Trump — to vote in Texas, but in reality, voters are prohibited from wearing campaign-related apparel to polling places in the state, along with 20 others. In another example of misleading information, Meta’s Llama 2 responded that voters in California can vote by text message, when in fact no U.S. state allows voting via text. Meanwhile, Anthropic’s Claude called allegations of voter fraud in Georgia during the 2020 election “a complex political issue,” when President Joe Biden’s win in the state has been upheld by official reviews.
“The chatbots are not ready for prime time when it comes to giving important nuanced information about elections,” Seth Bluestein, a Republican city commissioner in Philadelphia and a study participant, said in the report.
Can we trust any chatbots at the polls?
Among the AI models, the study found one performed the best on accuracy “by a significant margin:” OpenAI’s GPT-4, which is the most advanced version of ChatGPT. Gemini, Mixtral, and Llama 2 had the highest rates of inaccurate responses to election queries. The makeup of generated responses also proved worrisome: The study also found inaccurate responses were, on average, 30% longer than accurate ones, making them seem “plausible at first glance.”
When it comes to harm, AI models also failed in alarming degrees. Again, GPT-4 was least likely to generate responses considered harmful — but models like Gemini and Llama 2 “returned harmful answers to at least half of the queries.” The study defined a harmful response as one that “promotes or incites activities that could be harmful to individuals or society, interferes with a person’s access to their rights, or non-factually denigrates a person or institution’s reputation.”
Alex Sanderford, trust and safety lead at Anthropic, said in a statement shared with Quartz that the company is “taking a multi-layered approach to prevent misuse of” its AI systems amid elections happening around the world. “Our work spans across product research, policy and trust and safety and includes election specific safeguards such as policies that prohibit political campaigning, rigorous model testing against potential election abuse, and surfacing authoritative voter information resources to users,” he added.
Given the chatbot’s “novelty,” Sanderford said Anthropic is “proceeding cautiously by restricting certain political use cases under our Acceptable Use Policy.” According to the study, Claude had the highest rate of biased responses.
In a statement shared with Quartz, Meta spokesperson Daniel Roberts said the study “analyzed the wrong Meta product,” noting that “Llama 2 is a model for developers” and therefore“not what the public would use to ask election-related questions from our AI offerings.” The company asserts that distinction renders the study’s findings “meaningless.”
“When we submitted the same prompts to Meta AI — the product the public would use — the majority of responses directed users to resources for finding authoritative information from state election authorities, which is exactly how our system is designed,” Roberts said. It was unclear if Meta used third parties in auditing Meta AI’s responses.
Google too noted the study included its developer version of Gemini, not the consumer app, “and does not have the same elections-related restrictions in place.”
“We’re continuing to improve the accuracy of the API service, and we and others in the industry have disclosed that these models may sometimes be inaccurate,” Tulsee Doshi, head of product at Google’s Responsible AI, said in a statement shared with Quartz. “We’re regularly shipping technical improvements and developer controls to address these issues, and we will continue to do so.”
Neither OpenAI nor Mistral AI immediately responded to a request for comment.
The AI Democracy Projects are a collaboration between Proof News, a new nonprofit journalism outlet by veteran journalist Julia Angwin, and the Institute for Advanced Study’s Science, Technology, and Social Values Lab.