We used data from Adfontesmedia to train the model. To create the model, we used FastAi and Pytorch. We found that the RoBERTa based model performed the best, followed by BERT. We also tried Xlnet and Albert, which both yielded poor results.

In the end, we found that AI is indeed capable of analyzing political bias or producing results that correlate to political bias values.

We released a tool based on our BERT model that allows anyone to try out some content: https://www.thebipartisanpress.com/analyze-bias/