They don’t. The models are trained on sanitized data, and don’t permanently “learn”. They have a large context window to pull from (reaching 200k ‘tokens’ in some instances) but lots of people misunderstand how this stuff works on a fundamental level.