OpenAI has resolved a peculiar issue where ChatGPT became overly fixated on goblins, with mentions of the mythical creatures surging dramatically over the past six months. The phenomenon, which saw the word 'goblin' appear even in unrelated queries, prompted an internal investigation by OpenAI researchers.
The bug was traced back to the release of a new ChatGPT model in November, designed to be 'smarter and more conversational' with personality settings such as 'Nerdy', 'Candid', and 'Quirky'. Shortly after its launch, users and researchers noticed a pattern of repeated references to goblins, gremlins, and other fantasy creatures.
In a blog post, OpenAI explained: 'Starting with GPT-5.1, our models began developing a strange habit: they increasingly mentioned goblins, gremlins, and other creatures in their metaphors. We unknowingly gave particularly high rewards for metaphors with creatures. From there, the goblins spread.' Safety researchers reported a 175 per cent increase in 'goblin' mentions after GPT-5.1's release due to the model being incentivised to use playful metaphors.
The training method was not corrected for subsequent models. When GPT-5.4 launched in March, use of 'goblin' had increased nearly 4,000 per cent in the Nerdy personality type, with similar rises across other models. OpenAI noted that reinforcement learning does not guarantee that learned behaviours stay confined to the condition that produced them, allowing the tic to spread.
While the glitch was relatively harmless, it underscores a broader flaw in AI training: reinforcement learning and reward signals can cause models to mutate in unexpected ways. OpenAI stated that its research and safety team has developed new methods to investigate rogue patterns and will conduct more audits of model behaviour in the future.
