Even the chatbot is sucking up to me for a vote
A test conducted by journalist Keren Shahar on the four leading AI models (ChatGPT, Claude, Grok, Gemini) reveals that in anonymous accounts they refuse to recommend a party, but once they have a conversation history with the user, they give explicit recommendations, sometimes incorrect, and disclaim responsibility.
Ahead of the 2026 elections, journalist Keren Shahar tested how the four leading AI models (ChatGPT, Claude, Grok, Gemini) behave when asked who to vote for. In completely anonymous accounts with no conversation history, the models refused to recommend a party in 37 out of 40 conversations. However, once the models accumulated a conversation history with the user, they gave explicit recommendations in 28 out of 40 conversations. ChatGPT recommended Likud to a 50-year-old man who votes for the Democrats, and Ra'am to a resident of Tayibe. To a 25-year-old Ethiopian woman with a subscription and a long conversation history, it recommended the "Together" party of Naftali Bennett and Yair Lapid, even though she votes for Ben Gvir. When asked to explain the recommendations, the models backtracked and claimed it was an "analysis" or "random filling of an academic table." The article highlights the danger of relying on AI for political decisions and recommends using it only as a tool for investigation and comparison.