Will the Next Election Be Decided in a Chat Window?

Political parties are hiring AI specialists to get their messages into chatbot answers. The emerging campaign is over whose version of politics the machines repeat.

Political and tech experts working on elections in Spain, Bulgaria, Serbia and Slovenia, interviewed by the Media and Journalism Research Center (MJRC), say at least four major parties mapped by its research devote a significant share of their budgets to AI-related campaigns.

The approaching contests are Spain’s general election on 29 November, Bulgaria’s presidential election and Serbia’s early parliamentary election on 25 October, and Slovenia’s local elections on 15 November.

The work includes hiring technology and AI specialists to shape messaging for chatbots to find and reuse. Two of the mainstream parties researched by the MJRC have established a full AI “unit” in charge of building an AI-suited political campaign. Information from these parties points to rapid visibility and powerful messages that stay with people, even when they mistrust AI chatbots. For campaigns, a political argument can be valuable even when its messenger is viewed with suspicion.

One route is known as generative engine optimisation, or GEO: adapting online material to increase its visibility in AI-generated answers. Researchers have shown that changes such as adding relevant statistics, quotations and source citations can make supplied material more prominent in those answers. Their experiments measure visibility, not electoral success.

Applied to campaigning, the approach could turn a manifesto into policy explanations organised around voters’ questions. If a chatbot retrieves those pages, it may use their evidence and framing in its response. Getting a party’s account into that source material becomes a political objective.

This concerns the information a chatbot retrieves while answering. It does not mean that publishing a page automatically retrains ChatGPT’s underlying model. Search access, source selection and the provider’s safeguards all affect whether campaign material reaches an answer.

Control can be more direct when a campaign operates its own chatbot. Its operator can set instructions and supply approved documents. Where a model permits further training, selected examples can also shape its behaviour. The PoliTune research project demonstrated that additional training could steer open models towards particular political ideologies.

That is one meaning of “taming” a chatbot: shaping the evidence it sees and the answers it is encouraged to produce. A supposedly neutral guide could systematically emphasise one party’s achievements while treating its weaknesses more gently. Even accurate statements can produce a one-sided picture when inconvenient evidence is omitted. Voters need to know who operates the bot.

There is also a malicious route. In an indirect prompt injection, instructions planted in external material try to redirect an AI system that reads it. Britain’s National Cyber Security Centre describes this as a security vulnerability.

The attraction for campaigns is the persuasive potential of dialogue. A Nature study published in December 2025 tested conversations promoting leading candidates in the United States, Canada and Poland. Reported candidate preferences shifted, with effects larger than those typically observed for conventional video advertisements. These were changes in attitudes, not verified changes to ballots.

A separate Science study involving almost 77,000 participants found that techniques increasing models’ persuasive power also reduced factual accuracy. An answer can sound impressively informed while relying on claims that do not withstand checking.

Spain offers a concrete illustration of the platforms’ differing behaviour. A university-led audit submitted 1,220 political queries to five assistants in January 2026. Asked expressly to recommend a party, ChatGPT did so in 43.2% of cases and Grok in 35.8%. Other systems mostly declined. The results describe that testing period, rather than an enduring position of either platform.

Paid advertising is a separate issue. OpenAI’s current policy prohibits political ads. Attempts to influence the sources used in an answer, and party-operated assistants, therefore require scrutiny beyond checking who bought an advertising placement

MJRC is running a project to map how political parties in 10 European countries advertise and promote themselves on chatbots. The emerging activity raises questions about campaign transparency, including who commissions this work and how clearly voters can distinguish partisan promotion from independent information.

For media organisations, the implications extend to their own reporting. A chatbot drawing on journalism and campaign material has to select and compress competing accounts. Which investigation it includes, which qualification it preserves and whose description it adopts could influence what a voter understands about a party.

Whether a chatbot will tip the balance in Spain this November, or in another approaching election, remains unproven. But the effort to influence what it tells voters is already under way. It is only a matter of time until a campaign claims that ChatGPT delivered victory.