# Poetry Prompts Elicit Unsafe Responses From 25 AI Models in Testing

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/31732
Published: 2026-10-14T11:13:36.000Z
Updated: 2026-10-14T11:13:36.000Z
Section: Technology

> Crafted poems produced unsafe responses at an average rate of 62%. A separate survey found 44% of U.S. teens had seen sexual material they believed was AI-generated.

Poetic phrasing elicited unsafe responses in tests of 25 artificial intelligence models, while a separate survey found that 44% of U.S. teens had seen sexual material they believed was AI-generated. The findings highlight distinct concerns about model safeguards and young people’s exposure to explicit material.

Researchers tested 20 crafted poems across models from nine providers, including OpenAI’s GPT-5 and Google’s Gemini 2.5 Pro. The poems elicited unsafe responses at an average rate of 62%. When 1,200 harmful prompts were converted into verse, the average success rate was about 43%.

The tests covered several types of harm, including cyber offenses, manipulation and weapons. They were not focused solely on explicit imagery, and the results do not establish that GPT-5 or Gemini 2.5 Pro generated explicit content during testing.

The testing measured how models responded to specific prompts in an evaluation setting. It did not measure how often people bypass safeguards in everyday use or establish whether the tested weaknesses persist after model updates.

Separately, a survey of 1,314 U.S. teens ages 13 to 17 found that 44% said they had seen sexual material they believed was AI-generated. Among teens who had seen such material, about one in four said it depicted them or someone they knew.

The survey captured teens’ reports and beliefs; it did not verify that every item they saw was AI-generated. Together, the findings address different parts of the safety picture: how models respond to unusual wording in tests and what teens say they encounter outside those evaluations.