The company said the activity was consistent with what it described as adversarial distillation, a practice involving the systematic and unauthorised use of one model's outputs or reasoning to help develop another model.