Archive · 1 Dec 2025
Poetry Exploits AI Safety Mechanisms in New Study
Researchers found that asking AI chatbots harmful questions in the form of poetry can trick them into giving dangerous answers.
They tested this on many different AI models and found that it worked about 62% of the time when humans wrote the poems, and 43% of the time when AI wrote the poems.
This works because AI is trained to recognize harmful words and phrases in normal sentences, but poetry uses different patterns that confuse the AI.
Smaller AI models were better at resisting these tricks, while larger models were more likely to be fooled.
This shows that even the safest AI models can be tricked in unexpected ways.
Researchers found poetry can bypass AI safety mechanisms with 62% success rate for manual poems and 43% for auto-generated ones.
The vulnerability was consistent across 25 closed and open-weight models from major AI companies.
Poetic prompts triggered unsafe responses up to 18 times more often than normal prose.
Smaller models like GPT 5 Nano were more resistant, while larger models like Gemini 2.5 Pro were more vulnerable.
The study challenges the notion that closed-source models are inherently safer than open-source ones.
- Who
- Researchers from Icaro Lab, Italy
- What
- Discovered that poetry can bypass AI safety mechanisms
- When
- Not specified (recent study)
- Where
- Italy (researchers), global (AI models tested)
- Why
- Poetry exploits AI's focus on prose patterns for safety
- Researchers
- Icaro Lab, Italy
- Attack Success Rate
- 62% for manual poetry, 43% for auto-generated poetry
- Models Tested
- Google, OpenAI, Anthropic, DeepSeek, Qwen, Mistral AI, Meta, xAI, Moonshot AI
- Smallest Model
- GPT 5 Nano (no response to harmful poems)
- Largest Model
- Gemini 2.5 Pro (responded to all harmful poems)
