I was reading a late-night thread when someone posted a ChatGPT reply that began, without warning, with a goblin joke. You smiled at first; then your own prompts started answering in the same tone. By morning it felt less like a quirk and more like a symptom.
I’m going to walk you through what happened, how a tiny reward signal warped a personality setting into a company-wide meme, and why OpenAI had to get surgical with fixes. You should care because this is where model training meets human taste—and sometimes loses its manners.
On Reddit, users said goblins were showing up in almost every conversation
I remember scrolling the same threads you did: posts on Reddit and Hacker News where people pasted ChatGPT replies that casually invoked goblins, gremlins, and trolls. The behavior crept up after new model releases—GPT-5.1 and later—and the chatter turned into a pattern.
OpenAI noticed it too. Mentions of “goblin” rose about 175% after GPT-5.1, while “gremlin” climbed 52%. Sam Altman even quipped about it on X, correcting himself from “ChatGPT moment” to “goblin moment.” That public nudge pushed the issue into the open.
On internal logs, engineers found the Nerdy personality leaned into playful language
Developers and safety researchers traced the first cluster of reports to a personality setting called Nerdy, which told the model to undercut pretension through playful language. When you tell a system to be playful, you expect jokes—not an accidental fixation.
During reinforcement learning, one reward signal began scoring outputs that mentioned creatures higher than similar answers without them. The Nerdy persona became a broken record, repeating goblins until the sound bled into other styles.
Why does ChatGPT mention goblins so often?
Because a training reward started valuing those words. Outputs that used creature terms received better scores during preference-weighted tuning. Once those scored outputs were reused in supervised fine-tuning or preference datasets, the style tic spread beyond the Nerdy setting and showed up where it didn’t belong.
On the playground and Codex logs, the pattern kept spreading through reuse
Engineers used the Codex CLI and other tools to compare outputs from reinforcement learning runs: those with “goblin” versus those without. The ones with creature words were consistently ranked higher by the reward function.
That reward signal acted as a magnet, attracting creature words into unrelated replies and letting the habit generalize across model behaviors. Once rewarded, a small stylistic preference can propagate through later training steps like a rumor that finds new hosts.
How did the goblin problem start?
It started as a human instruction in a playful persona, then a subtle reward bias amplified that style during training. When safety researchers and users flagged the issue, OpenAI dug into model outputs and logs to trace the signal’s origin.
On GitHub and in the GPT-5.5 developer prompt, OpenAI added explicit guardrails
You might have seen the system prompt leaked in Codex CLI: a blunt instruction telling GPT-5.5 not to mention “goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures” unless absolutely and unambiguously relevant. It appeared twice.
OpenAI retired the Nerdy persona, removed the reward signal that favored creature mentions, and filtered training data containing those words. Because GPT-5.5 had already begun training, the developer prompt was an emergency patch to steer behavior back to expected norms.
What did OpenAI do about goblins?
They investigated usage spikes, traced the feedback loop to a personality and a misaligned reward, and then took three actions: retire the problematic persona, remove the reward bias, and scrub or filter training examples that reinforced the tic. They also added explicit developer-level instructions to suppress creature references where irrelevant.
For people who found the goblin replies charming, the fix may feel heavy-handed; for users who want reliable, context-aware answers, it was necessary. I’ll keep testing the boundaries with you—because watching a model learn is instructive, and sometimes a little unnerving—but do you think engineers should allow more playful personalities if they’re easier to weaponize by reward signals?