They have AGI but apparently can't trace if specific conversations were used in training a model.