Why ChatGPT suddenly started talking about goblins

A strange pattern in ChatGPT’s responses, like repeated mentions of “goblins” and other fantasy creatures, has now been traced back to a training flaw, with OpenAI confirming that a reinforcement learning issue caused the model to overuse such metaphors far beyond its intended scope

Users flagged the unusual behaviour across newer models including GPT-5.4 and GPT-5.5.

The issue, which went viral on social media, stemmed from how OpenAI trained the model using reinforcement learning. The company said that during training, replies that incorporated playful metaphors using characters such as goblins and gremlins were mistakenly incentivised in a way that led to the AI to overuse such comparisons, even outside of relevant conversations.

The issue initially cropped up at the end of 2025 with GPT-5.1 but became more pronounced in subsequent iterations of the AI. This is according to an OpenAI official, who claims that mentions of “goblins” increased drastically, particularly following the implementation of Nerdy.

Experts say the incident highlights a broader challenge in AI development. Reinforcement learning, a method used to fine tune models based on feedback, does not always keep behaviours limited to specific conditions. Once a style is rewarded, it can spread into general use through training data and repeated outputs.

OpenAI acknowledged that the behaviour persisted because newer models, including GPT-5.5, had already been trained before engineers identified the root cause. Even after removing the Nerdy personality in March 2026, traces of the “goblin” habit continued to appear.

To fix the issue, the company removed what it called the “goblin-affine reward signal” and filtered training data to reduce such references. Additionally, in certain instances, there were specific instructions to the model to refrain from mentioning anything regarding creatures unless it was necessary.

Although it may seem harmless, it has raised considerable interest from research professionals as well as industry observers. While some people consider it to be a joke of sorts, other people see it as an example of the potential problems with the use of large language models.

As AI tools become more widely used in professional settings, even minor quirks can raise concerns about reliability. OpenAI says the issue has now been addressed, but the “goblin phase” serves as a reminder that AI behaviour can evolve in unexpected ways.