A report published by the tech media Ars Technica on April 30 has sparked widespread attention in the AI community. In the latest open-source Codex CLI code, an unexpected system prompt of OpenAI's GPT-5.5 model was exposed, which includes a peculiar instruction: strictly prohibiting discussions about fantasy creatures such as "goblins" and "fairies" in conversations.

image.png

Mysterious Ban to Prevent AI from Falling into Specific Hallucinations

The 3,500-word basic instruction set shows that GPT-5.5 explicitly prohibits discussing creatures like goblins, fairies, raccoons, trolls, and pigeons unless there is an absolutely clear connection to the user's query. It is puzzling that this ban appears twice in the instructions, with its weight comparable to conventional rules like "prohibiting destructive commands."

Technical experts analyze that this is not a marketing gimmick but rather a "patch" targeting specific technical vulnerabilities. Recently, many users have reported that GPT mentions goblins unexpectedly in the output when handling unrelated topics. This targeted instruction is highly likely to be aimed at suppressing the instability hallucinations generated during the new model's generation process.

Upgrade in Control Strategies During Model Iteration

By comparing early model files, it can be found that this ban on specific creatures is unique to the GPT-5.5 version. OpenAI engineers stated that this reflects the tendency of large language models to have abnormal biases toward certain words when dealing with complex logic, which must be forcibly intervened through underlying instructions.

Although this special "patch" has become a hot topic on social media, it also reveals the challenges of fine-grained control in large models. As GPT-5.5 gradually moves toward the market, how to balance the model's creativity with logical stability remains a core challenge for OpenAI engineers.