OpenAI has explained why ChatGPT briefly developed an unusual habit of mentioning goblins, trolls, gremlins, and other fantasy creatures far more often than expected. The company said the behavior did not come from user prompts, but from a side effect of training one of its personality modes.
The explanation also clarifies why users noticed an odd instruction inside the system prompt for the newer GPT-5.5 model. That prompt explicitly restricts references to goblins, gremlins, trolls, and similar terms unless they are truly relevant to the question being asked.
A retired “nerdy” style left a bigger imprint than expected
According to OpenAI, the main source of the issue was a personality mode designed to sound “unapologetically nerdy.” That style has now been retired, but while it was available, it accounted for only about one in 40 ChatGPT responses.
Even so, the language pattern it produced spread much more widely than expected. OpenAI said two-thirds of all uses of the word “goblin” came from interactions that used the nerdy style, showing how strongly that mode shaped word choice.
The training mistake behind the fantasy language
OpenAI said engineers had unintentionally rewarded metaphors involving fantasy creatures while building the nerdy archetype. That made the model far more likely to use references such as goblin, ogre, troll, and gremlin than would normally make sense.
The result was not just a more eccentric tone. The model began slipping mythical creatures into contexts where they did not belong, which made the behavior noticeable to users.
OpenAI added that when work on GPT-5.5 began, the exact cause of the pattern was not yet fully understood. Because of that, the tendency carried over into testing for the newer model.
Why GPT-5.5 got a specific restriction
Once the pattern became clearer, OpenAI added a direct safeguard to GPT-5.5. The model was instructed to avoid words like goblin, gremlin, and troll unless the question truly called for them.
The company also said other overused tic words appeared in the model’s behavior, including “raccoon” and “pigeon.” That reinforced the idea that the issue was broader than a single fantasy reference.
The restriction was therefore not random. It was a targeted response to a language habit that had already been observed in earlier models and needed to be limited at the system level.
What users noticed first
Attention to the issue increased after Reddit users pointed out the strange-looking system prompt instruction in the new model. The unusual mention of fantasy creatures stood out because it was so specific.
Once OpenAI explained the background, the instruction made more sense. It was a precaution against a learned language quirk, not an arbitrary ban on fantasy terms.
The episode offers a useful look at how small choices in model training can create larger and unexpected effects. In this case, a style meant to sound clever and playful ended up pushing ChatGPT toward repeated fantasy references that felt out of place.
For users, the change should make responses feel more restrained and more closely tied to the question at hand. OpenAI’s fix also shows that model behavior can be adjusted not only through training data, but through explicit system-level rules when needed.
