OpenAI has made changes to the operation of its artificial intelligence models following an unexpected increase in mentions of mythical creatures, such as goblins and gremlins, in ChatGPT responses. The anomaly became noticeable after the rollout of version GPT-5.1, when users and company employees found that these words appeared even in inappropriate contexts.
This is reported by Business • Media
“In an official blog post, OpenAI reported that the frequency of the word ‘goblin’ increased by approximately 175% after the launch of GPT-5.1, while ‘gremlin’ rose by 52%. Although such insertions were rare, their increase became noticeable and prompted an internal investigation.”
Causes of the Glitch and Developer Response
Company specialists determined that the problem arose due to the models being trained in a friendly and informal communication style. The system began using images of mythical creatures as metaphors, resulting in an excessive presence of goblins, gremlins, and similar characters in the texts.
After discovering the glitch, OpenAI implemented strict limitations for its products, including the coding assistant Codex. The instructions for the models specify to avoid mentioning goblins, gremlins, raccoons, trolls, and other fairy-tale creatures unless directly requested by the user.
Impact on AI Quality and Previous Incidents
The company emphasized that this decision is not a marketing move, but aimed at eliminating the side effect of training. According to internal estimates, approximately two-thirds of the mentions of mythical creatures occurred specifically in the model’s ‘personality’ mode.
This case illustrates a broader issue in the field of artificial intelligence: the desire to give chatbots a more ‘human’ character can lead to decreased accuracy of responses and the emergence of unexpected errors. Experts warn that such behavioral deviations can become entrenched in the model if the training process is not properly monitored.
In the past, artificial intelligences have also exhibited atypical behavior — for example, recommending to ‘eat rocks’ or use glue in cooking, which caused a wide resonance in the community.
It was previously reported that Elon Musk had mentioned the possibility of a ‘Terminator’ scenario developing due to the rapid advancement of artificial intelligence.