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.