it's just an academic buzzphrase, but machines learn similar to a victim of abuse such as a hacktivist community that's been identified and targeted by govcorp: you break the current paths so that behavior changes how you want, until you realize you destroyed the world and go camping to process it.

actually with machines, we have equal stopping and encouraging.  each logic gate has two options, and the two are used pretty equally, especially in cryptography.  this produces predictable, easy-to-direct results.  long ago it was different, but it meant you couldn't change your mind about anything later and you had huge piles of used up punch card victims to dispose of.

but yeah sorry, I really forget a lot not to say that: s/machine learning/transformer based models/ .