Hacker News Reader

Instead of deleting posts or garbling them, one could replace them and add new posts that portray a false persona.

Check out an open source browser extension called track me not that does this with search engine queries

Agree. If it is deep learning do not know any way to get from weights back to data.

There is some research being done on this topic. At least in ways to sign models so that you can test copyright based on certain input keys.

You can't un-train the model. Your contribution has already been made.

Either

#1 You can train the model continously

#2 the model stops learning

#3 it has a submodel which detects you are feeding it false data.

Given that people change over time, #2 seems unlikely. The question to me is "how advanced is #3?". Another thing I wonder about is how relative are these to eaxh other. At odds?

#1 is a well-established technique called "online learning", and I'd bet money that it's how many of these "industrial scale" ML algorithms are trained.

#2 makes no sense from a business perspective.

#3 is also well-established. This is how Google's Captchas work, for example.

There are ways to be reclassified as a undesirable persona