if you're trying to retrieve information, but you have to do it in natural language, you're likely1 going to be prone to putting a lot of presuppositions into your questions, giving the bot a good signal of what traits you possess (compared to plain queries that could be you being curious about a subject)
for example
- a search query like "husky diet" only tells the model that you might have some interest in what huskies eat, but
- "hey Chatbot, what kind of food is healthy for my husky" gives a much stronger signal that you own a husky and will have continued interest in buying husky-related things
(and the bad for privacy bit being that now the advertiser has a more accurate model of facts about you, which they could potentially sell / share / have stolen)
this post also brought to you by thinking about "Which Linguist Invented the Lightbulb? Presupposition Verification for Question-Answering" by Kim et al 2021
-
I haven't gone and found data to support this, but I'll betcha posing something in the form of a natural language conversation will make it more likely for an interlocutor pose something in the form of a question with presuppositions that will end up implying facts about themselves, compared to search queries
