If there was enough Base64 in the training set then it would pick it up. It's just more text to predict. The best prediction is the correct prediction. Maybe t

If there was enough Base64 in the training set then it would pick it up. It's just more text to predict. The best prediction is the correct prediction. Maybe this answer seems a bit lacking but no one, not even the people training these models know exactly what it is these artificial neurons are learning or doing during inference so anything more is more speculation than not.