A growing chorus of ChatGPT users is hitting the same wall: the model refuses creative writing requests that have nothing to do with harmful content. The latest example making rounds involves ChatGPT blocking a roleplay prompt that simply included the word "Shazam" - the DC Comics character's magic word - in a fictional spell.
This isn't a case of someone trying to get around safety guardrails. It's a person writing fantasy fiction who got stonewalled by a filter that apparently couldn't tell the difference between a comic book reference and something dangerous.
The pattern has been building for months. Users writing fiction, game scenarios, and creative projects report that ChatGPT increasingly refuses prompts involving conflict, magic, mild tension, or even historical events. The model sometimes won't generate villain dialogue, describe a fantasy battle, or write a scene where a character is angry. Each individual refusal might seem like a reasonable edge case. Taken together, they suggest the filters are calibrated for maximum caution at the expense of basic usefulness.
OpenAI faces a real tension here. Every jailbreak disclosure (including one today involving a competitor) puts pressure on AI companies to tighten restrictions. But filters that block "write me a spell with the word Shazam" aren't preventing misuse - they're just annoying paying customers. A $20/month subscription that can't write a fantasy scene is a hard sell against competitors like Claude or Gemini that handle creative writing with fewer false refusals.
OpenAI has not publicly addressed the recent wave of filter complaints or indicated whether recalibration is planned.