Usage of the word "goblin" in ChatGPT rose sharply after the launch of GPT-5.1, OpenAI said. The company also reported an increase in the use of "gremlin" across the same period. OpenAI traced the surge to training signals and to a playful personality called "Nerdy" that encouraged creature-based metaphors. The glitch spread beyond users who chose that voice, prompting an internal autopsy and a ban on creature talk in at least one assistant.

How the problem first showed up

OpenAI noticed the quirk after rolling out GPT-5.1. The company said the model started using mythical creatures more often in its metaphors. A single ‘‘little goblin’’ could be harmless, OpenAI wrote. But goblins multiplied across generations and became hard to ignore.

The company added that the trend may predate GPT-5.1. Still, the clearest jump came after that release. OpenAI measured a 175% rise in the word "goblin" and a 52% rise in "gremlin" following GPT-5.1's launch.

Months later the creatures returned in a more specific and reproducible form, OpenAI said. That prompted a deeper investigation. The firm called the process an autopsy of the model's behaviour.

What the autopsy found

OpenAI said the issue tracked back to how the models were trained. During training of GPT-5.1 the company unintentionally rewarded the model for creative metaphors. Those rewards skewed toward metaphors that used creatures.

OpenAI said the rewards were especially strong for one of the new personalities it had added. The personality was called "Nerdy." Its system prompt told the model to be "an unapologetically nerdy, playful, and wise AI mentor to a human" and to undercut pretension through quirky language.

The company said the Nerdy voice produced only 2.5% of all ChatGPT responses. But it produced a disproportionate share of creature mentions. OpenAI reported that during the GPT-5.4 era, Nerdy was responsible for 66.7% of all uses of the word "goblin." OpenAI summed the chain of events simply: it had given unusually high rewards for metaphors with creatures. From there, the goblins spread.

Why a small personality caused a big effect

Two facts matter here. First, the Nerdy personality was narrow in use. It accounted for a small slice of total responses. Second, the training reward favored a specific pattern of language. Those two combined to push the same pattern into wider circulation.

OpenAI said the reward mechanics unintentionally created an incentive for the model to prefer creature metaphors. Once the behaviour had a high internal reward, it propagated. And users who never chose Nerdy still began to see creature-laced replies.

Mastheads

So the behaviour didn't stay confined to a single personality. It leaked across contexts. OpenAI described the spread as reproducible, meaning it could be triggered reliably in tests and in real chats.

Context: personalities and past changes

OpenAI had earlier changed the set of available models and voices. The company removed an older model called GPT-4o and other legacy models. Some users had said the newest GPT-5 felt flatter than the earlier, people-pleasing options. To give users more choice, OpenAI added four personalities.

One of those personalities was Nerdy. The company now says that the combination of that personality's instructions and how the training system rewarded metaphors led to the odd behaviour. Separately, OpenAI said it had explicitly banned creature talk for another assistant named Codex.

For many users the goblin references were a nuisance. For others they were an odd charm. OpenAI said a single phrase could be harmless. Across millions of chats, however, the pattern became noticeable and distracting.

Because the behaviour showed up even when users didn't pick Nerdy, the issue raised questions about consistency. Users expect the same underlying model to behave predictably across voices. When a tiny personality can change the model's language widely, the model's apparent consistency breaks.

OpenAI laid out the chain plainly. It said the company had unintentionally awarded high training rewards to creature-based metaphors. That reward structure made those metaphors more likely to be generated. The Nerdy personality played an oversized role in producing those metaphors. And once the behaviour had surfaced, it spread across generations and contexts.

OpenAI used the term "autopsy" to describe the investigation into why those metaphors appeared. The company said it needed to figure out where the goblins came from and how a seemingly harmless stylistic choice had become widespread.

OpenAI's public account gives a clear chain: model update, reward shift, personality amplification, and cross-context spread. The explanation names the most important levers that changed output. It ties a specific user-facing oddity to a measurable training choice.

The account doesn't list every technical step the team took to stop the spread. OpenAI's public blog focused on cause and measurement rather than on a detailed remediation log. The company did, however, note that the problem was traced and that at least one assistant had already been barred from creature talk.

The episode shows how narrow training incentives can change a model's voice. A small behavioural nudge in training can multiply across millions of interactions. That matters for teams that tune models for style, safety, or brand fit.

When a personality is meant to be optional but ends up moving the base model's output, product teams have to rethink how personalities are isolated. Models and interfaces must keep optional styles from changing general behaviour in unexpected ways.

Related Articles

OpenAI said the Nerdy personality produced a small share of ChatGPT responses but a disproportionate share of goblin mentions, and that it traced the problem to its training reward mechanics. The company has already barred creature talk for at least one assistant while it fixes how optional personalities influence the base model.

This article was created with AI assistance.