If it's because you like the bias, what the fuck.
If it's because you think data is a source of objective truth, what the fuck.
You always have to look at your results with a critical eye to recognize when something has gone wrong with your model. You can't recognize a flaw in the model if you believe everything the model says.
In the article's example, you're constructing a model to determine how a movie might be reviewed. The data indicates that movie reviewers may be reviewing movies with male leads more positively than movies with female leads. There are a lot of possible reasons for this other than "men are better than women", so why would you simply reject the result out of hand?
Language has too much implied context and meaning which is communicated outside the text under analysis. Embeddings attempt to capture this, but they aren't as good as humans.
"XXXXX is a dog of a movie" - does this mean the movie was bad or was the author playing with words when the movie is about dogs?