AI summaries of Tripadvisor hotel reviews downplay serious complaints, investigation finds

My job is currently working on creating an “assistant AI editor” for our customers to help them write stuff. One thing I’ve found is that most models tend to be weirdly averse to any kind of violent language; for example I have a document about a bakery ran by rats (I have pet rats, they’re adorable). If I specify in the AI bot instructions to avoid violent references, especially to knives and stabbing and then create a document full of references to knives it will suggest replacements… because that language is “unprofessional”.

It will never specifically cite the previous instructions of “don’t allow violent phrases” in favor of calling it “not business friendly” or “unprofessional” even though the root instructions make no mention of professionalism or the like.

Not to say that TripAdvisor isn’t doing it intentionally, but I could absolutely see this as being a side effect of most LLM models avoiding unpleasant terms.