Brennan Center: chatbots erred on a third of election queries
TL;DR
- Only 8% of Americans used AI chatbots for 2026 election information, per the TechPolicy.Press research roundup.
- Brennan Center tests of six chatbots found factual errors in roughly one-third of answers and deceptive imagery generated in over 80% of tries.
- Princeton's Kevin Greene found voting-logistics accuracy at 92-95% in English, with accuracy dropping when the same questions were asked in Spanish.
Only about 8% of Americans used AI chatbots to get information about the 2026 election, according to a research roundup published in TechPolicy.Press by Leticia Bode and Katie Groome. Low adoption has not stopped researchers from cataloguing quality problems the authors argue are worth addressing before the user base grows. Two researchers we track shared the piece within a day of publication.
Mark Krass of the Brennan Center for Justice tested six chatbots on election-related prompts. The systems rejected conspiracy theories in every tested case, but roughly one-third of their responses contained factual errors, and they "consistently failed to identify fake images" while generating deceptive imagery themselves in over 80% of attempts. Princeton's Kevin Greene, evaluating voting-logistics questions, found chatbots "accurate about 92 to 95 percent of the time" in English, with accuracy dropping significantly when the same questions were posed in Spanish.
On the administration side, Isabel Linzer of the Center on Democracy and Lia Merivaki of Washington State University describe a "sizable gap in communication between AI companies and election officials," leaving local administrators without the technical skills or resources to evaluate the tools. "AI's threat to election administration is systemic," Merivaki says, and in the absence of federal governance she advises state and local governments to write their own rules.
Two pieces round out the roundup. NYU's Jonathan Nagler documents that roughly 5-18% of AI queries involve political questions, mostly factual asks like election dates rather than vote recommendations. Vanderbilt's Josh Clinton, working with NBC News, finds that AI-generated "silicon samples" compress opinion variance compared with human polling, a "flattening effect" that G. Elliot Morris of FiftyPlusOne says yields "prediction data" rather than reliable public-opinion measurement.
Shared on Bluesky by 2 AI experts
Originally reported by techpolicy.press
Read the original article →Original headline: What We Now Know About AI and Elections