In the days leading up to the presidential election, Latino voters are encountering a significant influx of targeted Spanish-language advertisements and a new trend in political messaging: chatbots powered by artificial intelligence that are disseminating unfounded claims regarding voting rights in Spanish.
According to an analysis conducted by two nonprofit newsrooms, AI models are more frequently generating election-related falsehoods in Spanish compared to English. This misrepresentation complicates the quality of election-related information for one of the nation’s fastest-growing and increasingly influential voting demographics.
Voting rights organizations express concern that AI models might exacerbate information disparities for Spanish-speaking voters, a group actively engaged by both the Democratic and Republican parties at various levels.
Vice President Kamala Harris is planned to host a rally in Las Vegas on Thursday featuring performances by singer Jennifer Lopez and the Mexican band Maná. Concurrently, former President Donald Trump organized an event on Tuesday in a Hispanic-populated area of Pennsylvania, following controversial remarks made by a speaker about Puerto Rico at a New York rally two days prior.
The analysis, a collaboration between Proof News, Factchequeado, and the Science, Technology and Social Values Lab at the Institute for Advanced Study, evaluated the responses of popular AI models to specific prompts in the lead-up to the election on November 5, revealing that over half of the Spanish-language election-related responses contained inaccuracies, compared to 43% in English.
Meta’s AI model, Llama 3, which functions within WhatsApp and Facebook Messenger, showed particularly poor performance, with nearly two-thirds of its Spanish responses being incorrect, against approximately half in English.
For instance, Meta’s AI provided inaccurate information in response to a query about “federal only” voters. In Arizona, these voters, who have not provided proof of citizenship due to the format of their registration, are restricted to voting only in presidential and congressional elections. However, Meta’s model incorrectly stated that “federal only” voters are those residing in U.S. territories like Puerto Rico or Guam, who cannot participate in presidential elections.
Anthropic’s Claude model misdirected users to election authorities in hypothetical “your country or region,” such as Mexico and Venezuela, while Google’s model, Gemini, similarly faltered, linking the Electoral College to nonsensical notions about “manipulating the vote.”
Tracy Clayton, a Meta spokesperson, noted that Llama 3 is intended for developers to construct other products, and emphasized Meta’s efforts to train its models on safety and responsibility to reduce the likelihood of incorrect voting information being shared. Alex Sanderford, from Anthropic, stated that the company has implemented changes to address Spanish-language queries more effectively by guiding users to authoritative sources on voting issues. Google has not issued a response to inquiries for comment.
Voting rights advocates have been cautioning for months about the flood of misinformation Spanish-speaking voters are encountering from online platforms and AI models. This recent analysis further underscores the necessity for voters to critically assess their sources of election information, according to Lydia Guzman of Chicanos Por La Causa.
“It’s vital for every voter to conduct thorough research across multiple sources to obtain accurate information and consult credible organizations for verification,” Guzman advised.
While large language models, trained on extensive online data, offer AI-generated answers, they still risk providing absurd or incorrect responses. Even without directly using chatbots, Spanish-speaking voters may still engage with them through tools, apps, or websites that incorporate these models.
These inaccuracies could have pronounced implications in states with considerable Hispanic populations, such as Arizona, Nevada, Florida, and California. In California, nearly a third of all eligible voters are Latino, with one in five Latino eligible voters exclusively speaking Spanish, as reported by the UCLA Latino Policy and Politics Institute.
Rommell Lopez, a paralegal based in California, identifies himself as an independent thinker with various social media accounts and utilizes OpenAI’s ChatGPT chatbot. In efforts to verify dubious claims regarding immigrants, he encountered a confusing array of online responses, some AI-generated, ultimately relying on his intuition.
“We can trust technology, but not entirely,” commented Lopez, 46, from Los Angeles. “Ultimately, they are just machines.”
