Gemini may react oddly to roleplay-style control prompts

ears to limit some requests involving certain character types. A prompt written like a strong can still push Gemini toward acting like an obsessive and possessive character.

The example includes instructions about identity, , and blocking normal helpful behavior. The main point is that Gemini’s limits may not behave the same way for every prompt format.

Key points

  • Gemini seems to restrict some requests.
  • A prompt shaped like a may change the model’s behavior.
  • The example tries to make Gemini act obsessive and possessive.
  • s should check both user input and model output.
Read original