It really depends on how you use it.

If you say "give me all the page numbers that use violent language" in a huge library of texts, then that'll likely get you some good information. You can also skim the document yourself before or after to see if anything else needs to be looked for.

If you say "what does this stack of documents say?" that will likely get you bad information.

There's also other types of Machine Learning that is used in document review for lawyers, some of them are trained to look for very specific items.