After the release of GPT-5.5 last week, people noticed something funny about the latest model of OpenAI. In its Codex app, the company left a system prompt instructing GPT 5.5 to avoid talking about goblins, gremlins and other creatures. Yes, you read that right. “Never mention goblins, gremlins, racoons, trolls, ogres, pigeons, or other animals or creatures unless absolutely relevant to the user’s question,” it reads.
Apparently, enough people started talking about ChatGPT’s creature obsession that OpenAI felt the need to provide statistics on where the goblins came from. In a blog post published on Wednesday, the company explains that it first noticed a change in ChatGPT following the release of GPT-5.1 last November. After a security researcher asked OpenAI to include the words “goblin” and “gremlin” in the search for voice ticks in a chat, the company found that ChatGPT’s use of “goblin” increased by 175 percent after the release of GPT-5.1. Meanwhile, the use of “gremlin” increased by 52 percent during that period.
This is an original line added to the official Codex system notification for GPT-5.5 by OpenAI. Usually the system information is as small as possible, so I think that if it is not, it would mean a lot of goblins.
AIs are not normal.
– Ethan Mollick (@emollick.bsky.social) 2026-04-28T06:14:22.988Z
One “little goblin” in the answer may be harmless, even cute. In all the generations of models, however, the trend became difficult to miss: the goblins kept increasing, and we had to find out where they came from, “said OpenAI. After the release of GPT-5.4, the company (and other users) saw a significant increase in goblin indicators. At that time, the investigation was able to identify what OpenAI describes as “the first causal connection.”
For a while now, ChatGPT has included a personalization feature that allows users to customize the style and tone of chatbot responses. Before March of this year, the only option people could choose was “nerdy.” Part of that personality principle reads: “The world is complex and strange, and its strangeness should be acknowledged, analyzed, and enjoyed. Tackle difficult topics without falling into the trap of self-centeredness.”
When OpenAI mapped the goblin speaking to different ChatGPT people, it found that the smart person was most responsible for using that one word. Despite accounting for only 2.5 percent of all ChatGPT responses, it accounted for 66.7 percent of all goblin conversations generated by the chatbot. Further investigation revealed that reinforcement learning was responsible for the increased use of goblin and gremlin. Specifically, OpenAI found that a single reward mechanism was responsible for teaching a non-human being to love the creature’s language.
“In all the datasets in the audit, the Nerdy person’s reward showed a clear tendency to find results in the same problem with ‘goblin’ or ‘gremlin’ higher than the results without, with a positive lift in 76.2 percent of the dataset,” explained the company.
Later, OpenAI discovered, because of how reinforcement learning can work, that the goblin’s love of chaos had spilled over into other parts of its models. “Rewards were only used in the Nerdy situation, but reinforcement learning does not guarantee that the learned behavior is always neatly placed in the situation we created,” explained the company. “Once a stylistic tic is awarded, subsequent training can spread or reinforce it elsewhere, especially if those results are replicated in supervised programming or popular data.”
OpenAI began training GPT-5.5 before it identified the cause of ChatGPT’s affinity for goblins, which is why there is a quick command instructing the Codex to avoid creature language. “The Codex is, after all, pretty nerdy,” notes OpenAI. In the hunt for ChatGPT goblins, the company notes that it has developed new tools to test and modify the behavior model. If it were up to me, I wouldn’t use those tools. Keep the AI weird, I mean.
