These are fixable as the model is improved, and learns which information to discard. The fact that gpt4 is already so less error prone that gpt3 shows the speed at which this is happening