Researchers discovered subtle political biases in AI-generated text from ChatGPT and Bard.

ChatGPT's political orientation shifted over time and varied across languages.

Bard consistently exhibited a left-leaning bias.

Undetected biases can lead to manipulation.

Authors developed models to identify biases in text generation.

A multilingual corpus of newspaper articles with political stance annotations was constructed.

The study shows the feasibility of creating meaningful political stance models using distant supervision with a diverse corpus.

Recognizing biases in AI-generated content is vital for a more informed understanding of news.