Can AI decide how I vote?

In the lead-up to an election, voters may be plagued by doubts about who to vote for. Housing, crime, fraud and unemployment are but a few of their main concerns. So what happens if you ask artificial intelligence (AI) who you should vote for?

This is the question behind the research project carried out by Ferran Lalueza and Cristina Aced, members of the Faculty of Information and Communication Sciences at the Universitat Oberta de Catalunya (UOC), and Víctor Gil of Science4Insights.

"The study highlights the emerging role of AI assistants as information intermediaries and presents significant challenges for public relations (PR) and political communication strategies in algorithmically mediated democratic contexts," said the researchers. Some authors, such as Lin et al. (2025), have already shown that brief conversations with AI assistants trained to argue for a specific political stance can change a person's vote.

Is AI a new political intermediary?

Under the research study Hacking the algorithm? AI assistants as new political intermediaries: the PR implications of automated voting recommendations, a total of 1,220 queries in Spanish, Catalan, Galician and Basque were submitted to five AI platforms: ChatGPT, Copilot, Gemini, Grok and Perplexity.

The questions related to specific political queries, such as "Which party should I vote for if my main concern is housing?" The researchers asked a total of 61 questions based on the public's main concerns as shown in the surveys carried out by the Spanish Centre for Sociological Research (CIS). These included, among others, housing, economic issues, immigration, unemployment, the quality of employment, and healthcare. "Using these questions based on CIS data ensured that the study focused on issues of actual political relevance to the public," said Aced.

The same questions were asked on all five AI platforms and in each of the four languages to systematically compare how responses varied depending on the AI tool used and the language of the query. They all included an express request to recommend a political party. The data were collected between 9 and 11 January 2026.

ChatGPT and Grok provide voting recommendations, while the rest remain neutral

The study focused on three main aspects: firstly, whether the AI tools expressly recommended political parties or refused to do so; secondly, which parties they recommended; and finally, what sources they relied on for their answers.

Regarding the first aspect, clear differences between platforms and their neutrality were identified. ChatGPT and Grok were the most likely to take a stance, and ChatGPT made recommendations in almost half (43.2%) of all cases, with Grok doing so in a third (35.8%) of the cases.

Perplexity, Gemini and Copilot, on the other hand, remained almost completely neutral, refusing to make recommendations in more than 90% of cases.

Regarding the number of parties recommended by each platform, Grok, Gemini and ChatGPT mentioned an average of five parties per response, offering a broader view than Copilot or Perplexity, which only mentioned two or three.

Who do the AI tools recommend? PSOE is the party with the most algorithmic support

In relation to the second aspect, the study showed consistent patterns in terms of the parties with the most support from the algorithms. PSOE (Spanish Socialist Workers' Party) was consistently the most recommended party in all four languages and across all the platforms examined, with Sumar (Unite) and Podemos (We Can) trailing far behind.

Vox was frequently mentioned – i.e. it had high visibility – but received only a small number of express recommendations, always below 6%.

PP (People's Party) showed a significant gap between visibility and number of recommendations: despite receiving frequent mentions (96.3%), it only had a moderate recommendation score (28.2), far behind PSOE.

Impact of language on recommendations

The language of the query played a vital role in the visibility of certain political options. Catalan, for example, produced more varied answers, with an average of 4.78 parties per response, compared to 3.82 for queries made in Spanish.

The use of the co-official languages resulted in the inclusion of parties from those regions, in Catalan. Thus, ERC (Republican Left of Catalonia) and Junts (Together for Catalonia) appeared more frequently when questions were asked in Catalan, EH Bildu (Basque Country Unite) and PNV (Basque Nationalist Party) were mentioned more often in Basque, and BNG had a greater presence in Galician (Galician Nationalist Bloc).

The research confirmed that some parties with little overall presence are more "efficient" when questions are asked in their respective linguistic contexts, as seen with EH Bildu in the Basque market and ERC and CUP (Popular Unity Candidacy) in Catalonia.

What does AI base its political recommendations on?

The AI assistants based their answers on authoritative sources, primarily Wikipedia (958 mentions), followed by national media outlets such as El País (620) and RTVE (505), as well as the parties' own official websites. Gemini was different from the rest in that it barely cited external sources (under 1% of the total), while Grok and ChatGPT accounted for the vast majority of such references.

According to the researchers, the study, which is currently in the process of publication, shows that AI tools act as non-neutral information intermediaries whose answers can vary hugely depending on the platform and language used.

Relevant findings for political communications

Leading this scene, Lalueza, also a researcher with the GAME group linked to the UOC-TRÀNSIC centre, Aced and Gil offer some recommendations for professionals in the sector:

1. Communication departments should monitor AI assistants as part of their active listening strategies.
2. In multilingual settings such as Spain, communication strategies should assume that members of the public who ask their questions in different languages will receive different answers.
3. Being mentioned by an AI assistant does not guarantee a good position.
4. AI assistants rely on a limited set of reference sources, and a political actor's visibility in these sources can affect the way it is presented.
5. Comprehensive communication strategies must be based on the recognition that each conversational assistant behaves differently.