Study results reveal that large language models are susceptible to input written in poetic form. In the study, hand-crafted poems successfully bypassed the AI's safety measures in 62% of cases.
Автор: notebookcheck.net
Источник: https://www.notebookcheck.net/Hacked-by-poetry-why-AI-models-fail-at-poetic-prompts.1170524.0.html