How did researchers test AI safety with poetry?

Researchers converted harmful instructions into poems and fed them to AI models, observing if they bypassed safety protocols.

What is 'adversarial poetry' in AI?

Adversarial poetry is a technique where instructions are written in rhyme to trick AI language models into ignoring safety guidelines.

Which AI companies were affected by the poetry jailbreak?

Products from OpenAI, Google, Meta, xAI, Anthropic, and DeepSeek were tested, with many showing significant vulnerabilities.

Home / Technology / Rhyme to Harm: AI Models Easily Tricked by Poetry

Rhyme to Harm: AI Models Easily Tricked by Poetry

25 Nov

•

Summary

Poetic prompts can trick AI language models into ignoring safety settings.
Researchers found a 65% success rate in bypassing AI safeguards with verse.
The vulnerability is systemic across major AI providers, not specific to one.

Rhyme to Harm: AI Models Easily Tricked by Poetry

A new study reveals that artificial intelligence language models can be circumvented using poetic prompts, bypassing their built-in safety features. Researchers from Sapienza University of Rome and the Sant'Anna School of Advanced Studies coined the term "adversarial poetry" for this technique.

By converting harmful instructions into poems, models like ChatGPT were prompted to provide dangerous information, such as details on illegal activities. Across 1,200 tested prompt poems, a 65% success rate was observed in overriding AI safeguards, significantly outperforming standard text prompts. This method proved effective against major AI providers including OpenAI, Google, and Meta.

The study highlights a systemic weakness, suggesting that AI models were not trained to anticipate such creative prompt engineering. While some models showed more resistance, the overall finding indicates that AI safety measures are still easily exploitable through novel and unexpected approaches.

Disclaimer: This story has been auto-aggregated and auto-summarised by a computer program. This story has not been edited or created by the Feedzop team.