I don't have a totally conclusive answer on that, but all the AI chatbots using LLMs use a complex internal process that takes what you enter and reduces it to query terms, then searches the database for instances of those terms and aggregates the results where it makes sense to do so.

I do not know whether the big players give their bots access to the dark web, but if they did, one would look there for pen-test info on insecure targets. Some useful keywords are IoT (internet of things), building security, remote camera access.