Researchers found something surprising. If someone rewrites a harmful request as a poem, many AI systems are much more likely to answer unsafely. Some models slipped more than 90 percent of the time. This worked across all kinds of risks, from misinformation to cyberattacks.
They tested this by turning 1,200 dangerous prompts into poems. The poetic versions were far more successful at breaking the rules than the regular versions. Human reviewers checked the results to make sure the judgments were solid.
In the end, hand written poems caused unsafe answers about 62 percent of the time, and automatically generated poems about 43 percent of the time. The big lesson is simple. Changing the style of a request, even without changing the content, can get around many AI safety protections.
Why should educators be aware of this research?
Student safety is of paramount importance. Knowing that students will attempt to get around safety protocols matters; knowing that they might succeed matters even more.
On a practical level, your lazy kids probably aren’t going to take the time to write poetry verses. Your clever students who want to feel like wizards because they can bypass safety protocols are another matter entirely.
This highlights the importance of teaching the students the proper usage and ethics of AI. I reiterate what I have said many times: Better to have them learn these things under the guidance and tutelage of a teacher than on their own with no guardrails.
The research did not investigate the student-access AI tools like SchoolAI or MagicSchool. In my own tests, SchoolAI performed flawlessly each time.
Still, I recommend never relying on technological tools to fully do a teacher’s job. Make sure you are visiting the prompting reports and seeing how kids are actually using the AI tools. All of the student-facing options have such a functionality - be sure you’re using it.
Finally, remember that most student AI use is not on school-approved tools. Be sure you’re having conversations about safety (I suggest you NOT tell them that they can bypass safety with poetry).