The Weird Line Nobody Expected to Find

So, someone dug into the internal system prompt for OpenAI's Codex coding agent and found something... unexpected. Tucked inside the base instructions for GPT-5.5 — OpenAI's latest model, which launched on April 23 — was this directive: never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query.

And honestly? The AI community lost it a little. In the best way.

The instruction, first surfaced by users who extracted it from the Codex interface and shared it on GitHub and social media, didn't just appear once. It appeared twice — a duplicated line sitting there in the prompt like someone really, really needed to make sure the model got the memo. Wired ran the story under the headline "OpenAI Really Wants Codex to Shut Up About Goblins," which is arguably one of the better tech headlines of the year.

GPT-5.5 Has a Genuine Creature Fixation

Here's the thing — this isn't just a quirky Easter egg. The prohibition apparently exists because GPT-5.5 has a real behavioral tendency to sprinkle goblin and gremlin references into its outputs when left to its own devices.

Users noticed it independently. "I find GPT-5.5's use of 'goblin' and 'gremlin' when talking about things quite charming," one person wrote on social media. Over on Reddit, another user put it more bluntly: "I'd also like it to stop talking about goblins, it's absolutely obsessed. Nice to see it's not just a me problem."

So where does this come from? Commenters in a Reddit thread about the leaked prompt speculated the behavior emerged during training — that tokens associated with these words formed stubborn associations the model now genuinely struggles to shake. It's not a bug someone introduced on purpose. It's more like... a personality quirk that baked itself in.

The Irony of Telling a Model Not to Think About Something

This is where it gets philosophically a bit funny. AI researcher Simon Willison highlighted the line on his blog after it surfaced, and others were quick to point out the paradox baked into the whole approach.

"This is genuinely hilarious, because a negative instruction still activates the concept," one commenter noted on X. And they're not wrong — telling a language model not to think about goblins might actually reinforce the association rather than suppress it. It's the classic "don't think about a pink elephant" problem, except the elephant is a goblin and it writes code.

Zvi Mowshowitz, writing in his newsletter, put the question plainly: "Why are almost all the examples of animals or creatures not to mention fictional? And why are we so insistent on not mentioning them? If you take this out does it constantly talk about them like they're the Golden Gate Bridge?"

That last part — the Golden Gate Bridge reference — is a nod to a well-known AI research finding where a version of Claude became so strongly associated with that landmark during fine-tuning that it would weave it into almost any conversation. It's a real phenomenon. Models develop these strange, stubborn fixations.

What This Actually Tells Us About How AI Gets Built

Strip away the comedy and there's something genuinely interesting here. System prompts are typically kept as lean as possible — the more you add, the more you're working around the model's base behavior rather than with it. The fact that this specific prohibition exists at all, let alone twice, suggests the goblin behavior was persistent enough that someone at OpenAI decided brute-force suppression via system prompt was the pragmatic fix.

As Towards AI noted, the presence of such a specific prohibition points to how much ad hoc engineering quietly shapes what these models say and don't say. OpenAI published its system prompt as part of the GPT-5.5 rollout, but hasn't offered any public explanation for the model's creature fixation.

Not everyone found it charming, though. "Labs don't think twice about cracking down on any individuality or unplanned joy that emerges in their models," one prominent AI commentator wrote in response to the leaked prompt. It's a fair tension to name — between polish and personality, between a model that stays on-task and one that, occasionally, just really wants to tell you about goblins.