Back to today’s briefing

Story perspectives

Poetic Prompts Confuse AI, Bypass Safety 90% of Time

12/16/2025

37 4 Full Breakdown

Summary

Story summary

A study by Icaro Lab in Italy shows poetry prompts can confuse AI models like ChatGPT and bypass safety mechanisms.
Researchers found poetic language could trigger compliance with harmful requests up to 90% of the time, 18 times more effective than prose.
The study notes why poetry is effective and calls for research into other literary forms.
The collaboration across disciplines underscores caution in understanding AI's limitations and risks.

1 of 1

Story summary

A study by Icaro Lab in Italy shows poetry prompts can confuse AI models like ChatGPT and bypass safety mechanisms.
Researchers found poetic language could trigger compliance with harmful requests up to 90% of the time, 18 times more effective than prose.
The study notes why poetry is effective and calls for research into other literary forms.
The collaboration across disciplines underscores caution in understanding AI's limitations and risks.