[minimaxir]

See also: the Ask Delphi incident, which also used data from AITA as a source for an AI based on moral quandaries that blatantly had systemic biases: https://www.theverge.com/2021/10/20/22734215/ai-ask-delphi-m...

Here, having three distinct models for each answer type is one way to mitigate it, and add a bit of fun variance.