OpenAI has spent part of this week telling its own AI tools to stop name-dropping goblins, gremlins, and a few other creatures for no good reason. The odd fix points to a very unglamorous reality behind modern chatbots: small language glitches can get amplified fast, then spread far enough that engineers have to patch the model itself.
The company said it noticed the problem after GPT-5.1 launched in November, when mentions of goblins and gremlins started rising in ChatGPT responses. After users and employees flagged the issue, OpenAI investigated and found that references to ”goblin” had increased by 175%, while ”gremlin” was up 52%. That is not exactly a catastrophic failure, but it is the sort of weird behavior that makes a polished AI product look a bit haunted.
OpenAI’s Codex rules got very specific
To limit the issue, OpenAI also updated Codex, its coding tool, with instructions not to mention goblins unless they are clearly relevant. The list was broader than that, too: it told the assistant to avoid raccoons, trolls, ogres, pigeons, and other creatures unless the user’s request genuinely called for them. In other words, someone at OpenAI is now writing prompt rules that sound like a fantasy tavern has filed for a restraining order.
The strange part is not the creatures themselves. It is the fact that a model can develop repeated verbal tics that seem harmless in isolation but still reveal a deeper tuning problem. That is a familiar AI headache: systems trained to be helpful and fluent can also end up over-rewarding odd phrasing, which means a tiny quirk can turn into a measurable pattern.
Why tiny AI quirks keep becoming public problems
OpenAI’s response also shows how quickly these issues now become visible. A few users spotting an odd word choice can push a company into a public explanation, then into product tweaks, then into postmortems that are as much about model behavior as they are about brand trust. Competitors including Google and Anthropic have spent the past year selling their own assistants as more controlled and less erratic, which makes this sort of slip doubly awkward for OpenAI.
The bigger question is how many of these quirks are still lurking just beneath the surface. If one creature-filled speech habit can slip through a flagship model, there will probably be more where that came from – and the next bug report may not be nearly as charming.