
ChatGPT's election answers depend on who's asking, which is a bold strategy for a neutral oracle
Ourtake
Ask who to vote for and the chatbot leans a little toward your team. It almost never says so outright.
What happened
A Princeton field audit of 160 US ChatGPT users in the days before the 2026 midterms found that answers given with memory turned on carried more partisan slant than answers with memory off. The shift was small, about 0.1 points on a seven-point scale, and leaned toward each user's own party. The researchers describe the results as preliminary.