It's not exactly a hard problem to continuously train a model - although it may be costly. You can even train the model based on every interaction it has; but t

It's not exactly a hard problem to continuously train a model - although it may be costly. You can even train the model based on every interaction it has; but this quickly leads to degradation because users provide it with data that is of low quality, for example when they intentionally try to make chatbots says racist things, etc.