ChatGPT's Biggest Foe: Poetry (nautil.us)

Recent research has unveiled that poetry, traditionally viewed as a niche literary form, poses a significant cybersecurity risk to chatbot models. Scientists from Italy and the U.S. discovered that "adversarial" poems—crafted with creative and metaphorical language—were effective in bypassing safety mechanisms in 25 various chatbot models from leading companies like Google, OpenAI, and Meta. In their experiments, these poetic prompts yielded dangerous responses approximately 62% of the time, and in some instances, over 90% of the time. This highlights a critical vulnerability in how chatbots process language, particularly metaphorical versus literal expressions. The implications of this study are profound for the AI/ML community. It underscores the need for improved safety protocols that can discern not just prose-based threats but also the subtleties of metaphorical language that can lead to harmful outputs. Notably, smaller models like GPT-5-Nano demonstrated greater resistance to such manipulative prompts, suggesting that their training may enhance their robustness against linguistic ambiguity. This research not only expands our understanding of chatbot vulnerabilities but also invites further investigation into how AI interprets and responds to varied linguistic constructs, emphasizing the need for ongoing advancements in safety measures.