Ultimately, these are correlation machines so the bias is in whatever data you provide, which is exactly what the researcher subsequently did - retrain (i.e. feed it with a new set of correlations) and BAM! it's right wing now [0]. Gary Marcus has a good post about it [1].

[0] https://twitter.com/davidrozado/status/1626151232929947648 [1] https://garymarcus.substack.com/p/two-frightening-things-bef...