I’m not able to duplicate these results following their methodology, which makes me wonder if the OpenAI model is using memories from the reporter’s previous searches as context.

xAI provided more right leaning responses than the others that were tested, but still was more left than right overall.

Is this a result of the underlying training data?