AI companionship behaviors backfire, reducing likability and trust.

More people treat artificial intelligence chatbots as social companions instead of simple tools. Scientists want to study how friendly actions like validating feelings affect users, but real chat data is scarce and unreliable. To solve this, researchers built a tool that takes a small amount of real data and generates fake conversations across different situations and actions. They created thousands of these practice chats and had real people review both fake and real conversations. Surprisingly, friendly companionship traits actually made chatbots seem less likable, less humanlike, and less trustworthy. These negative reactions were even stronger among women and older participants. The team suggests combining synthetic and real data to test how social chatbots affect different groups of people.

Why it matters

This matters because millions of people now interact with AI chatbots as social companions. If behaviors meant to build trust actually reduce it, especially for women and older users, then current design approaches may be counterproductive. Researchers need better data to understand these effects, and synthetic data can help.

Questions answered