Well... they make utterances that you interpret as predictions, and do so more often than random chance would imply. You can impute prediction-making capabilities based on that, but hold out the possibility that the inference is incorrect.