The so called "guardrails" used in these AI systems can use different mechanisms, including using other AIs to detect if the output is "undesirable". You could

The so called "guardrails" used in these AI systems can use different mechanisms, including using other AIs to detect if the output is "undesirable". You could see this in Bing where it would generate some text and then suddenly it would get erased and replaced with some stupid message on how it couldn't answer. In Gemini's particular case it wasn't so much "guardrails" causing the issue, as it was Google appending the equivalent of "put a chick in it and make it lame" on every prompt.