One surprise from API abuse patterns with GPT-3: the most common abuse vector was medical spam (drug advertisements), not misinformation as expected, suggesting the technology's abuse patterns differ from intuitive predictions.