If the model was trained proper to the conversation with the researcher took place, there'd be no question of tainting the results. But if any amount of training on the model took place afterward, then yes, everything is thrown into doubt (a core problem with considering anything "original" from a model because of how >a % of everything ever written has been used a corpus for the training).