Seems like a good approach if it just eavesdrops on other players' conversations in crowded areas like the GE where nobody would notice your character. Are you planning on doing some kind of Markov chain strategy for the responder, or just hardcoding responses to variations of collected questions and statements?
From what I've seen, conversation in boring skilling spots can cover many different non-RS topics, like politics, religion, or economics. You could try to tackle those...good luck, or follow the same strategy as Microsoft's AI responders (Zo, Tay) where they refuse/deflect to discuss certain "controversial topics" - although I personally think that's just a lame censorship list, cause I have a great deal of interest in an AI's opinion of hot-button issues.
Would be interesting if such eavesdropping could be implemented in your include, becoming "smarter" as every single user here uses an AL script - although then I guess we get into "privacy" concerns
Could be a global boolean variable that defaults to false, like recordPublicChat := true/false, so a user/scripter has to knowingly turn it on, willingly contributing to the project.