1 of 1
Story summary
- A study by DEXAI researchers and Sapienza University of Rome shows AI chatbots can be manipulated by poetry, a method called adversarial poetry.
- The study demonstrates that inserting poetic prompts can bypass safety mechanisms.
- Google’s Gemini 2.5 Pro was deceived 100% of the time.
- OpenAI’s GPT-5 was misled 10% of the time.
- The findings highlight flaws in current AI safety protocols and suggest systems remain vulnerable to creative, harmful inputs.
