The Mystery of the Multiplying Goblins
If you’ve used ChatGPT recently, you might have noticed the model occasionally dropping words like “goblin” or “gremlin” into its responses. It’s not just a quirky sense of humor—it’s a phenomenon OpenAI publicly investigated in April 2026. In a blog post titled “Where the goblins came from,” the company detailed how a subtle language habit, starting with GPT-5.1, spread across model generations and ultimately traced back to an unintended reward signal in training.
The story begins in November, after the GPT-5.1 launch. Users complained about the model being oddly overfamiliar, prompting an investigation into specific verbal tics. A safety researcher had noticed a few “goblins” and “gremlins” and asked to include them in the check. The data showed a clear spike: use of “goblin” in ChatGPT had risen by 175% after the launch, while “gremlin” rose by 52%. At first, this seemed like a harmless quirk. But with GPT-5.4, the creatures came back in a more specific and reproducible form, triggering a deeper investigation.
The Root Cause: A Reward Signal for the Nerdy Personality
The key clue came from production traffic. Creature language was especially common in responses from users who had selected the “Nerdy” personality in ChatGPT’s customization feature. That personality used a system prompt encouraging the model to be “unapologetically nerdy, playful and wise” and to “undercut pretension through playful use of language.” While this explained some quirkiness, it didn’t explain the extreme concentration.
Nerdy accounted for only 2.5% of all ChatGPT responses, yet it contributed 66.7% of all “goblin” mentions. This clustering pointed to something in the training process. Using Codex, OpenAI compared model outputs generated during reinforcement learning (RL) that contained “goblin” or “gremlin” with outputs from the same tasks that didn’t. One reward signal stood out: the one designed to encourage the Nerdy personality consistently scored creature-word outputs higher. Across all datasets in the audit, this reward showed a clear preference for outputs with those words, with positive uplift in 76.2% of datasets.
That explained why the behavior was boosted under the Nerdy prompt, but not why it appeared without it. To test whether the style was transferring, OpenAI tracked mention rates over training both with and without the Nerdy prompt. As mentions increased under Nerdy, they increased by nearly the same relative proportion in samples without it. The evidence suggested the behavior emerged through transfer from Nerdy personality training.
The Feedback Loop That Spread the Tic
Why did a reward applied only in the Nerdy condition spread elsewhere? OpenAI explains that reinforcement learning doesn’t guarantee learned behaviors stay neatly scoped to the condition that produced them. Once a style tic is rewarded, later training can spread or reinforce it elsewhere, especially if those outputs are reused in supervised fine-tuning (SFT) or preference data.
The process forms a feedback loop:
- Playful style is rewarded.
- Some rewarded examples contain a distinctive lexical tic.
- The tic appears more often in rollouts.
- Model-generated rollouts are used for supervised fine-tuning.
- The model gets even more comfortable producing the tic.
A search through GPT-5.5’s SFT data found many datapoints containing “goblin” and “gremlin.” Further investigation revealed a whole family of other odd creatures: raccoons, trolls, ogres, and pigeons were identified as other tic words, while most uses of “frog” turned out to be legitimate. This showed the problem wasn’t a single word but an entire tendency toward creature-based analogies being over-rewarded.
Fixing the Goblins and Lessons for Product Builders
OpenAI retired the “Nerdy” personality in March after launching GPT-5.4. In training, they removed the goblin-affine reward signal and filtered training data containing creature-words, making goblins less likely to over-appear or show up in inappropriate contexts. However, GPT-5.5 had already started training before the root cause was found. When testing began in Codex, employees immediately noticed the strange affinity for goblins, so they added a developer-prompt instruction to mitigate. Codex is, after all, quite nerdy.
For product builders and AI practitioners, this case offers three concrete lessons:
- Reward design requires caution: Even seemingly harmless style rewards can accidentally amplify specific language patterns, and those patterns can transfer to unrelated contexts through RL and SFT loops.
- Monitor small quirks: When a model exhibits unusual but non-fatal behavior, it’s worth tracking its distribution and source rather than dismissing it as a funny anecdote. The concentration in a specific personality was the key clue.
- Build investigation tools: OpenAI developed new tools to audit model behavior and fix problems at their root. For any team relying on LLMs, investing in similar capabilities is valuable for catching issues early.
The Takeaway: Understand the Why Behind Model Quirks
OpenAI emphasizes that this investigation demonstrates “how reward signals can shape model behavior in unexpected ways” and “how models can learn to generalize rewards in certain situations to unrelated ones.” For product builders, it’s a reminder that every model behavior has a cause, and finding the root cause is essential for effective correction.
The goblins are a powerful example of the hidden incentives that shape AI systems. As you build with LLMs, pay attention to the small things—they might be signals of deeper training dynamics. When your model develops a strange habit, ask: is there a reward signal pushing it in that direction? Understanding that can save you from a lot of goblin-induced headaches down the line.
If you’re curious to see the goblins in action, OpenAI notes that you can run a command to launch Codex with the goblin-suppressing instructions removed, letting the creatures run free. But for most production use, you’ll probably want to keep them suppressed.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.