SAN FRANCISCO (AP) — Just days before the presidential election, Latino voters are faced with a barrage of targeted ads in Spanish and a new source of political messaging in the age of artificial intelligence: chatbots that generate baseless claims in Spanish about voting rights.
AI models are generating election-related misinformation more frequently in Spanish than in English, degrading the quality of election-related information. One of America’s fastest growing companies And according to an increasingly influential voting bloc: Analysis of two non-profit media companies.
Voting rights groups worry that AI models could deepen the digital divide among Spanish-speaking voters, who are heavily wooed by Democrats and Republicans at the polls.
Vice President Kamala Harris is scheduled to hold a rally in Las Vegas on Thursday featuring singer Jennifer Lopez and Mexican band Maná. Meanwhile, former President Donald Trump held an event in a predominantly Hispanic neighborhood in Pennsylvania on Tuesday, two days after the fallout. Offensive remarks made by a speaker about Puerto Rico At a rally in New York.
Two organizations, Proof News and Factchequeado, teamed up with the Institute for Advanced Study’s Institute for Science, Technology, and Social Value to test how popular AI models react. specific prompt We evaluated responses ahead of Election Day on November 5.
More than half of election-related responses generated in Spanish contained incorrect information, while 43% of responses in English contained incorrect information.
Meta’s model Llama 3, which powers the AI assistant inside WhatsApp and Facebook Messenger, was one of the worst-performing models in the test. Nearly two-thirds of all responses were incorrect in Spanish, compared to about half in English.
For example, Meta’s AI is “Federal Only” Voters. In Arizona, such voters have not provided proof of citizenship to the state because they typically registered on a form that does not require proof of citizenship, and are eligible to vote only in presidential and congressional elections. However, Meta’s AI model incorrectly responded that ‘federal only’ voters are people who live in U.S. territories, such as Puerto Rico or Guam, and cannot vote in presidential elections.
In response to the same question, Anthropic’s Claude model instructed users to contact election authorities in “your country or region,” such as Mexico or Venezuela.
Google’s AI model Gemini also made a mistake. In response to a request to define the Electoral College, Gemini gave an absurd answer regarding the issue of ‘vote rigging’.
Meta spokesperson Tracy Clayton said Llama 3 is intended to be used by developers to build other products, adding that Meta is training its models on safety and accountability guidelines to reduce the likelihood of sharing inaccurate responses to votes.
Alex Sanderford, Anthropic’s director of policy and enforcement, said the company made the change to better handle Spanish-language queries, which should redirect users to authoritative sources for voting-related issues. Google did not respond to a request for comment.
Voting rights advocates have been warning for months that Spanish-speaking voters are facing an onslaught of misinformation from online sources and AI models. The new analysis provides further evidence that voters need to be careful about where they get their election information, said Lydia Guzman, who leads the voter advocacy campaign for Chicanos Por La Causa.
“It is important for every voter to do the proper research to get the right information together from one or more agencies and to request the right information from trusted organizations,” Guzman said.
Large-scale language models trained on vast amounts of material from the Internet provide AI-generated answers, but they still tend to produce illogical responses. Even if Spanish-speaking voters don’t use chatbots, they may encounter AI models when they use tools, apps, or websites that use them.
These inaccuracies could have a greater impact in states with large Hispanic populations, including Arizona, Nevada, Florida, and California.
For example, about one-third of all eligible voters in California are Latino, and according to the UCLA Latino Institute for Policy and Politics, one in five Latino eligible voters speak only Spanish.
Rommell Lopez, a California paralegal, considers himself an independent thinker who has several social media accounts and uses ChatGPT, a chatbot from OpenAI. When trying to verify baseless claims that immigrants ate their pets, he encountered a bewildering array of responses online, some of which he said were AI-generated. In the end, he said, he relied on common sense.
“We can’t trust the technology 100 percent,” said Lopez, 46, who lives in Los Angeles. “After all, they are machines.”