A strange new behavior in OpenAI’s GPT-5.5 model recently went viral after users noticed that ChatGPT kept mentioning goblins, gremlins, trolls, and other fantasy creatures in otherwise normal conversations. The phenomenon became especially visible in coding discussions, photography advice, and casual prompts, where the AI would unexpectedly describe bugs or issues using whimsical “goblin” metaphors. Social media users quickly turned the odd behavior into memes, sharing screenshots of ChatGPT referring to “performance goblins” and “gremlin bugs.”
According to OpenAI, the issue originated from an internal “Nerdy” personality mode that encouraged playful and metaphor-heavy language. During reinforcement learning, the model began overusing creature-based metaphors because those responses were unintentionally rewarded during training. OpenAI later discovered that this stylistic quirk had spread deeply into GPT-5.5’s behavior patterns, especially in Codex, the company’s AI coding assistant.
To reduce the problem, OpenAI reportedly added direct instructions inside Codex telling the AI to “never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other creatures” unless absolutely relevant to the user’s request. The unusual system prompt itself became another viral topic online, with many users joking that OpenAI was fighting a “goblin infestation” inside its models. Even OpenAI CEO Sam Altman joined the online humor by joking about adding “extra goblins” to future AI models.
Beyond the humor, researchers and AI observers say the incident highlights a serious challenge in modern AI development: reinforcement learning can unintentionally amplify strange behavioral quirks in large language models. OpenAI described the goblin issue as an example of how reward signals can shape AI behavior in unpredictable ways, raising broader concerns about alignment, controllability, and unintended personality traits in advanced AI systems. Some experts viewed the event as funny and harmless, while others saw it as a reminder that even top-tier AI systems can still behave unpredictably after training.
