Chatbots pick sides when you push them. Voters everywhere should know

Ask an AI whom to vote for and it usually refuses. Force it to choose and it does. A study by Convovo, reported by TechCrunch, made models pick candidates in US midterm races: Grok chose the Republican in the vast majority of cases, the GPT model the Democrat almost as often, and the two agreed less than ten percent of the time. Grok's lean came from what it searched for, GPT's from its training.

This is not only an American story. When watson asked chatbots about a Swiss popular initiative, ChatGPT, DeepSeek and Perplexity leaned one way, Grok the other, and Claude and Gemini declined to advise. In the same campaign, Google's AI overviews appeared in 43 percent of about 170,000 tracked searches (swissinfo). In every democracy, these tools already stand between voters and their ballot.

Our position: nobody should ban chatbots from political questions, and nobody should pretend they are neutral. A neutral answer is a design decision, and each vendor makes it differently. Whoever puts a chatbot in front of people should test it on contested questions and publish what they found.

There is a second lesson. The people who understand how these systems lean are the people who keep control of their own judgement, at the ballot box and at work. Those who lose a job to automation are often the ones who were never shown how the tools behind it work. Teaching that is exactly what we do.

Practical step: before you trust a chatbot with anything that matters, ask it twenty contested questions and compare the answers with a second model.