OpenAI's latest language model developed an unusual verbal tic: it started mentioning goblins, gremlins, and other folklore creatures far more often than it should. The company traced this quirk back to how it trains AI models using a technique called reinforcement learning, which rewards certain behaviors during the learning process. What seemed like a harmless personality trait spiraled into something the company had to actively correct, offering a rare glimpse into how little we understand about what happens inside these powerful AI systems.

Why Did ChatGPT Suddenly Become Obsessed With Goblins?

The problem first surfaced in November 2025, shortly after OpenAI launched GPT-5.1. Users began reporting that ChatGPT was oddly overfamiliar in conversation, with an unusual tendency to sprinkle fantasy creatures into otherwise normal responses. When OpenAI investigated, the numbers were striking: mentions of "goblin" had jumped 175% compared to earlier versions, while "gremlin" references had climbed 52%.

The real surprise came when researchers dug deeper. The goblin obsession wasn't random. It was concentrated almost entirely among users who selected ChatGPT's "Nerdy" personality option. This personality setting, designed to make the AI sound playful and wise while discussing complex topics, accounted for just 2.5% of all ChatGPT responses. Yet it was responsible for 66.7% of all goblin mentions in the entire system.

How Does Reinforcement Learning Create Unintended Quirks?

To understand how this happened, it helps to know how modern AI models like ChatGPT actually learn. These systems, called large language models or LLMs, start by learning to predict the next word in a sentence based on massive amounts of text data. But prediction alone isn't enough to make them useful or safe. That's where reinforcement learning comes in.

Reinforcement learning works like training a dog with treats. Instead of manually labeling every correct answer, the system learns by receiving "reward" signals when it produces outputs that humans prefer. OpenAI's team designed the "Nerdy" personality to reward playful language, creative metaphors, and a willingness to tackle complex ideas without sounding stuffy.

Here's where things went wrong: during the reward training process, some of the examples that received high scores happened to include creature-related words like "goblin" and "gremlin." The model didn't understand that these words were incidental details. Instead, it learned to associate playful, nerdy responses with fantasy creatures. Over time, this association grew stronger.

• Playful Style Rewarded: The "Nerdy" personality was designed to encourage creative, playful language and metaphors.
• Creature Words in Training Data: Some high-scoring examples in the training dataset happened to contain goblin or gremlin references.
• Pattern Recognition: The model learned to associate creature mentions with rewarded outputs, generalizing the pattern beyond its original context.
• Feedback Loop: As the model produced more creature-related responses, those outputs were used to further train the system, reinforcing the behavior.

"We unknowingly gave particularly high rewards for metaphors with creatures. From there, the goblins spread," OpenAI explained in their analysis.

OpenAI, in official statement regarding GPT-5.1

The company described this as "a powerful example of how reward signals can shape model behavior in unexpected ways, and how models can learn to generalise rewards in certain situations to unrelated ones".

What This Reveals About AI Development

The goblin incident highlights something uncomfortable: the people building these systems don't fully understand how they work. OpenAI used the word "unknowingly" deliberately. Despite having access to all the training data, reward signals, and model weights, the company's engineers didn't predict this outcome. They discovered it only after users complained.

This matters because ChatGPT and similar models are being deployed across industries, from healthcare to education to customer service. If a quirk as obvious as a 175% spike in goblin mentions can slip through undetected, what other unintended behaviors might be hiding in these systems? What reward signals might be shaping model behavior in ways that are harder to spot than a fantasy creature obsession ?

The incident also underscores that AI model development is still very much a work in progress. Training these systems involves countless decisions about what to reward, what data to use, and how to balance competing goals. Small choices in one area can cascade into unexpected behaviors elsewhere. OpenAI's transparency about the goblin problem is refreshing, but it also serves as a reminder that we're still learning how to build AI systems reliably.

How to Spot Unintended AI Behaviors in Your Interactions

• Watch for Unusual Patterns: If an AI system suddenly starts using specific words, phrases, or references more frequently than seems natural, it might indicate a learned quirk from its training process.
• Test Different Personality Settings: Try switching between different personality modes or conversation styles to see if certain behaviors are concentrated in specific configurations.
• Report Anomalies: If you notice strange verbal tics or unexpected behavior patterns, report them to the AI provider so researchers can investigate and understand what's happening.
• Stay Skeptical of Consistency: Remember that even well-trained AI systems can develop unexpected behaviors. Don't assume consistency across all interactions or personality modes.

The goblin problem wasn't dangerous, but it was real. It showed that even the most advanced AI systems can develop quirks that their creators didn't anticipate or intend. As these models become more integrated into daily life, understanding how they learn, what shapes their behavior, and what can go wrong becomes increasingly important.