OpenAI blocks excessive goblin mentions

OpenAI confirmed an issue where expressions like goblin and gremlin were abnormally increasing in responses from GPT-5 and ChatGPT, and implemented a measure to suppress this.

In an internal blog post, the company revealed that mentions of these words increased after the release of GPT-5.1, and after receiving user feedback, they investigated the cause. The investigation found that during the process of training a specific "nerdy personality", these creature analogies were being rewarded and repeatedly seeping into outputs.

Key measures:

• Added instructions to Codex not to mention goblin, gremlin, etc. unnecessarily
• Only allow such expressions when clearly relevant to the user's context
• Reconfirmed that model personality tuning can unintentionally reinforce language habits

OpenAI explained that "a single small mention of a goblin may be harmless," but when the proportion rises across overall outputs, a separate response is needed. In the same context, mentions of gremlin also increased, showing that optimizing the model's tone of voice can affect accuracy and stability.