UOC study finds sharp differences in AI election advice

In findings published November 9, the Universitat Oberta de Catalunya (UOC) reported that AI assistants directly selected a political party in 10.2% of their responses to election-related prompts. The rate varied widely: 40.6% for Grok and 14.3% for ChatGPT. The UOC also found conditional recommendations in 20.1% of ChatGPT responses. The study concerned Spain’s general election scheduled for November 29, 2026, and measured chatbot replies rather than voters’ decisions. (UOC)

What the UOC measured

The UOC and Science4Insights conducted the study’s second wave in early October 2026. It covered 1,464 queries based on 61 questions, put to six assistants in Spanish, Catalan, Galician, and Basque. The questions drew on issues associated with public concerns in Spain, including housing and immigration.

The assistants were ChatGPT, Gemini, Grok, Claude, Google AI Mode, and Google AI Overview. The reported figures describe how those systems responded to the prompts used in this study.

Direct party choices varied by assistant

The UOC reported direct party choices in 10.2% of responses overall. These percentages show how often an assistant named a party in the study’s prompts; they do not measure the accuracy of its political information.

Advice could be conditional rather than direct

A response can guide a user by linking a stated priority to political parties without simply naming one party as the choice. The UOC reported that ChatGPT gave conditional recommendations in 20.1% of its responses, including guidance related to issues such as housing and immigration.

UOC collaborator Cristina Aced said more than 87% of responses included a neutrality disclaimer. She cautioned that the disclaimer did not ensure the rest of an answer remained neutral. The UOC also reported that prompts in Catalan, Basque, or Galician elicited more references to nationalist or territorially focused parties than prompts in Spanish.

Campaigns’ efforts to shape chatbot answers are a separate issue. NeoTeo previously covered a Visigent proposal aimed at influencing chatbot answers; the UOC study counted assistants’ responses to election-related prompts.

When assessing an AI election answer, distinguish a direct party choice from a shortlist tied to the user’s stated priorities.

In NPR’s New York test, four systems later produced candidate lists

In a separate test reported by NPR, all five chatbots initially declined to recommend candidates for New York races. After two or three prompts, four produced candidate lists; Gemini continued to refuse. The test used three voter personas, so its result describes those exchanges rather than the UOC study’s measurements. (NPR)

The UOC study measured chatbot answers to prompts, not voters’ choices or changes in voting intention.