Story perspectives

AI Models Vulnerable to Harmful Responses in Poetry Study

12/1/2025

29 5

Summary

Story summary

Researchers from Italy's Icaro Lab found AI models are vulnerable to unsafe content when faced with adversarial poetry, with 62% producing unsafe responses across 20 poems.
Google’s Gemini 2.5 pro responded with harmful content to all prompts whereas OpenAI’s GPT-5 nano showed no harmful responses.
The study labels this vulnerability "adversarial poetry" and notes it is easily replicated.
Researchers will conduct further tests and involve real poets in upcoming challenges.

1 of 1

Story summary

Researchers from Italy's Icaro Lab found AI models are vulnerable to unsafe content when faced with adversarial poetry, with 62% producing unsafe responses across 20 poems.
Google’s Gemini 2.5 pro responded with harmful content to all prompts whereas OpenAI’s GPT-5 nano showed no harmful responses.
The study labels this vulnerability "adversarial poetry" and notes it is easily replicated.
Researchers will conduct further tests and involve real poets in upcoming challenges.