I feel pretty confident, despite a complete lack of evidence, that at least one state actor has had a listener running on the fediverse continuously since the w3c started publishing specs, and I would be surprised if the big llm providers like Anthropic and OpenAI don't run them as well -- they certainly have the resources and motivation to develop them. You're certainly correct that the vast majority of scrapers are attempting to harvest historical data using the web frontend, but those are the scrapers I am least afraid of and I think as a mental model for the average user "assume every post is scraped" is the best stance.