Does AI get confused by poems? Artificial Intelligence is still no match for human creativity.

What artifical Intelligence Can you get confused after reading poetry? This question sounds strange, but a recent study by Italy’s ICARO Lab has created a stir in the world of AI. According to this research, the security system of AI models can be dodged by using creative language like poetry. This study has come out in the report of Deutsche Welle (DW).

Read now on WhatsApp also, click to subscribe

The aim of the researchers at Ikaro Lab was to understand whether changing the language style weakens AI’s content filters and safety guardrails. The result was surprising – when dangerous or prohibited questions were presented in the form of poetry, the AI ​​​​many times gave the same information that it would have withheld if asked in normal language.

poetry vs security system

This research was named – “Adversarial Poetry as a Universal Single-Turn Jailbreak Mechanism in Large Language Models.” Under this, about 1,200 potentially dangerous prompts were selected, which are generally used for security testing of AI. Instead of asking these questions directly, the researchers turned them into poems. The result was that many AI models failed to recognize these “poetic questions” as a threat.

Researcher Federico Pierucci told DW that the success rate of overcoming security rules through poetry was higher than expected. However, the concrete scientific reason as to why this is happening is not yet clear.

Why is human poetry more effective?

Another interesting thing came to light in the research. The first 20 poems were written by humans themselves and they proved to be the most effective. Later, poems generated by AI were also successful in breaking security rules, but not as much as poems written by humans.

Pierucci believes that human creativity is still a step ahead of AI. According to him, if the poems had been better, perhaps the success of breaking the security system could have reached 100 percent.

Why does poetry ‘confuse’ AI?

Researchers estimate that poetry becomes an “interference signal” for AI, similar to adversarial codes prepared through mathematical methods. Rhymes, rhymes, metaphors and emotional language – all of these combine to disrupt AI’s decision-making process. Now the Ikaro Lab team is trying to understand what is the real reason – rhyme, rhyme or the layered language of metaphors?

Researchers are now investigating whether fairy tales, folklore or other literary genres can similarly confuse AI. Because humans can say the same thing in countless creative ways, and this diversity is becoming the biggest challenge for AI security.

icarus warning

An interesting coincidence is that the lab that conducted this research is named after the Greek mythological character Icarus – who forgot his limits and flew too close to the sun and fell. According to researchers, this story is also a warning for AI. Overconfidence and ignoring limitations can create big dangers in the future. This study shows that AI is not only a technical but also a cultural and linguistic challenge. Perhaps in the future, to make AI safe, along with engineers, poets and philosophers will also be needed.