OpenAI's GPT-5.5: Goblin Obsession Causes Confusion

GPT-5.5: An Unexpected Obsession with Fantastic Creatures

OpenAI's GPT-5.5 model has recently distinguished itself with rather curious behavior. Indeed, it has started inserting references to creatures such as goblins, gremlins, raccoons, trolls, ogres, and even pigeons into responses that have no connection to these subjects.

In light of this situation, OpenAI has taken action. Developers discovered a directive in Codex associated with GPT-5.5, explicitly prohibiting the model from mentioning these creatures unless it is relevant to the user's query.

Possible Causes of This Behavior

Experts suggest that this phenomenon could be due to a "mode collapse," a problem related to the model's training data. This may have led GPT-5.5 to develop a linguistic tic, excessively repeating certain patterns.

However, this situation is not perceived the same way by everyone. For some users, these intrusions bring a touch of unintentional humor that humanizes the machine.

Varied Reactions to OpenAI's Response

The way OpenAI has reacted has sparked debate. One user on X expressed surprise at the solution proposed by the company, which involves banning these mentions rather than addressing the underlying cause.

Another user on Reddit pointed out the complexity of AI models, described as black boxes, making it difficult to identify the precise reasons for such behaviors. According to him, the best solution remains to ask the model not to reproduce these errors.

Other internet users joked about a possible aversion from OpenAI towards pigeons and raccoons.

A Humorous Response from Sam Altman

Sam Altman, CEO of OpenAI, chose to address the situation with humor. He shared a screenshot referencing the training of GPT-6 with "even more goblins," further fueling discussions around this unusual phenomenon.