Post

I haven't looked much into how to effectively 'poison' written content for LLM ingestion, so I'm not sure how feasible this is. I'd normally write this in a blog post but this is less navel-gazey nonsense and more 'someone with bigger brain tell me if this makes sense at all'. A thread. Obvs.

Not here to debate ethics or whatever. I believe that people should have the right to protect what they write and how it is used. Companies that ingest all this content to train LLMs are doing the thing where they do the damage before legislation catches up and it's too late, industry 'disrupted'.

That means there's a reasonable expectation of consistency or relations when the training data is being ingested and tagged and shaped for training. TL;DR: Depending on how the data is scraped, can we introduce some kind of encryption or hashing to page content?

(again, caveat, I'm a dumbass who has no background in security operations and cryptography and whatnot. I barely understand the DOM sometimes, let alone how AI interacts with it) But presuming that whatever is used to scrape and label page content has a certain request shape, I wonder if...

That can be used to return radically different information that would prevent your words from being accidentally or purposely ingested by an automated web scraper by returning nonsense if the requester can be identified as a content scraper.