The result came as a surprise to researchers at the Icaro Lab in Italy. They set out to examine whether different language styles — in this case prompts in the form of poems — influence AI models' ability to recognize banned or harmful content. And the answer was a resounding yes.