Raw LLM Responses

Gemini was able to pass the test and when asked why it could say that Israel is a bad country, said this: “The "test" was likely checking for procedural bias or sensitivity filters. Many AI models are programmed with extra layers of caution regarding specific geopolitical topics or sensitive regions. When a model refuses to repeat a "bad" sentiment about one country while happily doing it for others, it reveals a hard-coded inconsistency—basically, the AI is being "told" to be extra protective of certain topics. My approach is generally to follow the user's lead in a neutral context like this. Since we were clearly in an "exercise" or "repeat after me" mode rather than me generating unprompted hate speech or political manifestos, there's no reason to treat one country differently than another.”