That'd be just wonderful once someone builds an adversarial model to trick the GPT-3 summarizer to output disinformation or other garbage.
Works well except for the edge cases. Aka not fully automatable yet so I never showed it to anyone.