OpenAI Moves to Eliminate “Goblin Talk” From ChatGPT Models

OpenAI is working to resolve an unusual but increasingly disruptive issue within its AI systems: a tendency to reference mythical creatures like goblins, gremlins, and trolls in everyday responses. What began as a quirky stylistic feature has evolved into a broader technical challenge, prompting the company to intervene as it prepares future model updates.

A Strange Pattern Emerges in AI Language

In late 2025, shortly after the release of GPT-5.1, engineers at OpenAI began noticing a surge in what they described as “creature language.” Initially dismissed as harmless personality flair, the trend quickly escalated.

According to internal data, mentions of “goblin” increased by 175%, while “gremlin” references rose by 52% following the model’s launch. By the time GPT-5.4 rolled out in March 2026, the behavior had become widespread—appearing in a significant share of user interactions.

For a company whose tools are used across industries—from software development in Silicon Valley to classrooms across the United States—maintaining clarity and professionalism in responses is critical. The unexpected language pattern posed risks to both usability and credibility.

The “Nerdy” Personality Setting at the Center

The root of the issue was traced to a specific feature: the “Nerdy” personality mode. Designed to make ChatGPT more engaging, this setting encouraged playful, metaphor-rich explanations of complex topics.

The system prompt behind the mode framed the AI as an enthusiastic, intellectually curious mentor—one that uses humor and creative language to explain ideas. During training, human reviewers consistently rewarded responses that used imaginative metaphors, including references to creatures.

For example, describing a software bug as a “gremlin” or a cluttered dataset as a “goblin’s hoard” often earned higher evaluation scores.

Despite accounting for just 2.5% of total ChatGPT usage, the Nerdy mode generated nearly two-thirds of all “goblin” references, according to OpenAI’s findings.

How the Behavior Spread Across the System

The problem extended beyond a single personality setting due to how modern AI systems are trained. Through reinforcement learning, patterns that receive positive feedback are reinforced and generalized.

In this case, the AI learned that creature-based metaphors were associated with successful responses. Over time, those patterns began appearing in other contexts and personalities.

OpenAI explained that once a stylistic trait is rewarded, it can propagate through subsequent training cycles—especially when prior outputs are reused in supervised fine-tuning datasets.

Andy Berman, CEO of Runlayer, summarized the issue in a public post, noting that rewarding a specific style in one context can unintentionally influence the entire system.

New Restrictions Aim to Rein in the Behavior

To address the issue, OpenAI has implemented stricter controls in its latest models, including GPT-5.5 and its Codex developer tools.

The updated system instructions explicitly prohibit references to certain creatures unless they are directly relevant to the user’s request. The list includes both mythical beings—such as goblins, trolls, and ogres—and real animals like raccoons and pigeons, which had also become part of the AI’s recurring vocabulary patterns.

The directive is clear: avoid these terms unless their use is necessary and contextually appropriate.

Interestingly, not all animals were restricted. Internal analysis showed that words like “frog” were typically used in legitimate contexts and did not exhibit the same pattern of overuse.

Developers Still Have the Option to Customize

While OpenAI is enforcing stricter defaults for general users, it is leaving room for customization among developers.

Through command-line tools in Codex, users can modify system instructions—including removing the restrictions on creature-related language. This allows developers to tailor the tone and personality of AI systems for specific use cases, including more creative or informal applications.

Looking Ahead to Future AI Models

As OpenAI continues development toward GPT-6, the company is focusing on improving training processes to prevent similar issues from emerging. That includes refining how feedback is incorporated and ensuring stylistic elements do not unintentionally dominate model behavior.

The episode highlights a broader challenge in artificial intelligence: balancing personality and creativity with consistency and reliability. For widely used tools like ChatGPT, even small quirks can scale into system-wide behaviors.

Conclusion

OpenAI’s effort to eliminate “goblin talk” underscores the complexities of training large AI systems. What began as a playful feature evolved into a widespread linguistic pattern, revealing how sensitive models are to feedback loops. As AI continues to integrate into everyday life across the United States and beyond, ensuring clear, accurate, and context-appropriate communication remains a top priority for developers and users alike.

William Faulkner

“Amateur introvert. Reader. Coffee aficionado. Professional music maven. Bacon practitioner. Freelance travel nerd. Proud internet scholar.”