OpenAI fixes ChatGPT bug that made it overuse goblin and gremlin metaphors
OpenAI has said its ChatGPT models picked up an unusual habit of referencing goblins and similar creatures in responses, a quirk that began with GPT-5.1 and became more visible in later versions. The company said it has now taken steps to curb the behaviour by removing the training signals that encouraged it, effectively telling the models to stop using such recurring metaphors after the pattern spread across generations.
What are goblins and gremlins
OpenAI described these references as a stylistic issue rather than a functional bug. Terms like goblin and gremlin started appearing in responses as part of metaphorical or playful language. While occasional usage was not a concern, the company said the frequency increased over time, making it a noticeable pattern.
The behaviour was not limited to just those terms. OpenAI said it also observed other creature words such as raccoons, trolls, ogres, and pigeons appearing in responses, though not all instances were considered inappropriate.
Where it originated from and affected models
OpenAI traced the behaviour back to training signals tied to its personality customisation feature, particularly the Nerdy personality. The company said this mode was designed to encourage a playful and expressive tone, but in doing so, it unintentionally rewarded responses that used metaphorical creature-based language.
The first clear signs emerged after the launch of GPT-5.1, when OpenAI observed a sharp rise in such terms, including a 175 per cent increase in goblin mentions and a 52 per cent rise in gremlin references. The pattern became more pronounced with later models such as GPT-5.4, where a majority of goblin-related outputs were linked to the Nerdy personality, despite it accounting for a small share of overall usage.
OpenAI said further analysis showed that once rewarded, these stylistic patterns spread beyond their original context. Even responses generated without the Nerdy personality began exhibiting similar language, suggesting that the behaviour transferred during training and was reinforced over time through feedback loops in model development.
OpenAI’s fix for the problem
To address the issue, OpenAI said it retired the Nerdy personality after the GPT-5.4 release and removed the reward signals that favoured such language. It also filtered training data containing these terms to reduce their prevalence in future outputs.
However, the company noted that GPT-5.5 had already begun training before the root cause was identified. As a result, additional mitigation steps, including developer-level instructions, were introduced to limit the behaviour during testing and deployment.
What it means for users
For users, the change mainly affects tone rather than functionality. ChatGPT responses are expected to be more neutral and less likely to include repeated or out-of-context creature metaphors.
OpenAI’s intervention also suggests tighter control over stylistic quirks, especially those that can unintentionally scale across different use cases. However, if you liked having goblin and gremlin references, OpenAI has also released a fix for you.
How users can bring it back
OpenAI said users who prefer the style can still enable it in certain environments, such as Codex, by removing the suppression instructions applied to reduce such references. It has shared a command to launch Codex with the goblin-suppressing instructions removed, which users can find in its blog.
The company added that the episode underscores how small training incentives can influence model behaviour over time and has led to improvements in how it identifies and addresses such patterns.
