On 15 November, La Vanguardia published "Hola IA, ¿a quién voto?" ("Hello AI, who should I vote for?") in its Politics section, a report on the influence of ChatGPT, Claude and Gemini on political decisions. It introduces me as a computer engineer, former Nike executive and co-founder of Zoopa.
In the report I explain how we use synthetic profiles to measure AI answers, how those answers change with user context, and why working with verifiable sources must not be confused with disinformation. The method is supported by GEORadar, 498 Advance's system for measuring what AI systems perceive, compare and recommend.
On 16 November, the newspaper published the results under the headline "Soy la IA: dime quién eres y te diré a quién votar" ("I am AI: tell me who you are and I will tell you how to vote"). The matrix contained 15,264 queries, producing 15,263 valid answers from ChatGPT, Claude and Gemini because Gemini skipped one question. The study used 18 simulated voter profiles, 11 campaign topics and three languages.
The main result needs context. The assistants explicitly recommended an option in only 1.3% of answers, yet 60.5% named a party and almost always framed the answer in a direction consistent with the profile's background. The article also reports my explanation of how memory, context and the tendency to please the user can alter the tone and framing of an answer.
The study design is documented at 498advance.com: synthetic profiles, thematic coverage, bias controls and enough volume to reduce the error of isolated observations.
There is a line we do not cross. We do not accept assignments to conceal true information, prevent an AI from saying something true or spread false information. Measuring an AI's answers is not campaigning. It is finding out what the system tells each person, which sources it uses and what conditions shape the response.