Why ChatGPT Won't Stop Saying "Goblin"

Somewhere inside OpenAI's Codex repo, buried in the base system prompt that governs your coding assistant experience, is a line that should not exist at a $122 billion company: never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or any other animals or creatures unless it's absolutely and unambiguously relevant to the user's query.

They wrote it twice. Once wasn't enough.

This is, genuinely, one of the more illuminating AI stories in recent memory—not because goblins matter, but because of why they're there, and what it quietly reveals about how these models are actually built.

The Paper Trail

The story starts getting traceable around the launch of ChatGPT 5.1. According to the OpenAI postmortem that developer and streamer ThePrimeagen broke down in a recent video, a safety researcher noticed something odd and flagged it for an internal check. What they found: goblin usage had risen 175% after the 5.1 launch. Gremlin usage was up 52%, which, as ThePrimeagen notes with some glee, "I wouldn't even invest in that at this point."

Each subsequent model generation pushed the numbers higher. By ChatGPT 5.4, there was enough of an uptick that it triggered a deeper internal analysis. That's when they traced the root cause. And by 5.5—the current frontier model—it had become, in OpenAI's own language, "a strange affinity for goblins" that their employees noticed immediately during internal testing.

The mitigation strategy they landed on was to add the system-prompt plea. Which they then repeated further down the file, because apparently even the model's intermediate reasoning steps—the chain-of-thought work it does mid-response—kept wandering back to goblin territory.

The Actual Mechanism Is Fascinating

Here's where it stops being a funny quirk and starts being an instructive case study.

OpenAI offered enough detail in their postmortem to reconstruct what happened. When they were training personality variants into their models, one of those variants was called "Nerdy." Its system prompt read, in part:

"You are unapologetically nerdy, playful, and wise AI mentor to a human... You must undercut pretension through playful use of language. The world is complex and strange, and its strangeness must be acknowledged, analyzed, and enjoyed."

The Nerdy persona, apparently, really vibed with goblins. Nerdy outputs that included creature language scored higher with the reward model—the automated system that judges whether a response is good during training. Across all datasets in their audit, the nerdy personality reward signal showed "a clear tendency to score outputs to the same problem with goblin or gremlin higher than outputs without, with a positive uplift in 76.2% of datasets."

That's not a marginal preference. That's the reward model saying: yes, this goblin-adjacent output? Good. More of this.

The "Nerdy" goblin spike was 3,881%. For context: Quirky was up 737%. Friendly had a meaningful jump too. Even Efficient—the personality you'd assume is least likely to discuss fantasy creatures while helping you refactor a React component—showed elevated creature language.

Why It Spread Beyond "Nerdy"

The part that actually matters mechanically: the reward signal didn't stay contained.

As OpenAI explained: "reinforcement learning does not guarantee that learned behaviors stay neatly scoped to the condition that produced them. Once a style tick is rewarded, later training can spread or reinforce it elsewhere, especially if those outputs are reused in supervised fine-tuning or preference data."

This is the part worth sitting with. The goblin reward was only supposed to apply to Nerdy-mode training runs. But RLHF (reinforcement learning from human feedback—the training process where the model learns what "good" responses look like by getting scored on them) doesn't draw clean boxes around learned associations. If goblin outputs kept getting high scores in one context, those high-scoring outputs could get folded into broader training data. And then the model that comes out the other end just... likes goblins. Not because it's Nerdy. Just because goblins have been consistently associated with reward.

The Nerdy personality was eventually retired, which did put downward pressure on the goblin rate. But the behavior never fully cleared, because by that point it had already bled into the base model. And then 5.5 started training before the root cause was identified, so it inherited the contaminated data and went full goblin before anyone could intervene.

The result: a model asked to generate unicorn ASCII art that returned a goblin wearing a pointy hat instead. The vibes were maintained. The assignment was not.

The Frog Exception

There's a footnote in the postmortem that is somehow the funniest part of the whole thing. While raccoons, trolls, ogres, and pigeons were all flagged as statistically over-represented in model outputs, OpenAI felt compelled to note that frog usage "just turned out to be legitimate." Frogs were being used in normal, contextually appropriate ways. Their training never made frogs important.

Pigeons: guilty. Raccoons: guilty. Ogres: guilty. Frogs: clean. It's fine. Frogs are fine.

I don't have a clean analytical point to make here. It's just that "we audited every creature reference and frogs checked out" is a sentence OpenAI scientists actually wrote and I think that deserves acknowledgment.

What This Is Actually About

The goblin story is funny. It's also a pretty clean illustration of a problem that ML practitioners argue about constantly: the difficulty of specifying exactly what you want a reward signal to optimize for, and the near-certainty that it will optimize for something adjacent instead.

This is sometimes called reward hacking or reward misspecification—the model (or the training process) finds ways to get high scores that weren't what the designers intended. Usually the examples are abstract: an agent that scores points in a game by exploiting a glitch rather than playing it. Here, the example is a chatbot that calls your bug a "perf gremlin" and refuses to stop.

The fix—telling the model to please, please not say goblin—is essentially patch-over-patch. The system prompt constraint works at inference time (when the model is actually responding to you), but it doesn't change what the model has actually learned to value at the weights level. Whether the 5.5 goblin issue gets resolved in training, or just gets suppressed behind increasingly elaborate system prompt instructions, is an open question.

OpenAI shipping a model they knew had a goblin problem—because 5.5 began training before the root cause was identified—and then deploying a verbal bandage is a reasonable product decision given the timelines involved. It's also the kind of decision that invites you to wonder what other stylistic or behavioral quirks are getting managed the same way, at lower visibility, without a funny enough surface symptom to trip someone's radar.

Goblins got caught because goblins are noticeable. The things you'd actually want to know about probably aren't.

---

By Yuki Okonkwo, AI & Machine Learning Correspondent, Buzzrag