'Never talk about goblins' instruction found in OpenAI Codex internal prompts

Buried within the internal system prompt for OpenAI's Codex coding agent is an unusual 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." The instruction, which appeared in the leaked base instructions for GPT-5.5 — OpenAI's latest model, launched on April 23 — has drawn amusement and bewilderment across the AI community.

Published

Apr 28, 2026

GPT-5.5: Capabilities and Reactions - by Zvi Mowshowitz

A Curious Prohibition

The system prompt, first surfaced by users who extracted it from the Codex interface and shared it on GitHub and social media, contains the creature ban not once but twice — a duplicated line that only added to the comedy. Wired reported on the directive under the headline "OpenAI Really Wants Codex to Shut Up About Goblins".

The prohibition appears to stem from a real behavioral quirk. GPT-5.5 has a tendency to pepper its outputs with references to goblins and similar creatures when left to its own devices. "I find GPT-5.5's use of 'goblin' and 'gremlin' when talking about things quite charming," one user wrote on social media. On Reddit, another noted: "I'd also like it to stop talking about goblins, it's absolutely obsessed. Nice to see it's not just a me problem". Commenters on a Reddit thread about the system prompt speculated the behavior arose during training, where tokens associated with these words formed stubborn associations the model now struggles to shake.

The Paradox of Negative Instructions

AI researcher Simon Willison highlighted the line on his blog, quoting it directly from the Codex base instructions for GPT-5.5. Others pointed out the irony of the approach. "This is genuinely hilarious, because a negative instruction still activates the concept," one commenter observed on X, noting that telling a language model not to think about goblins may only reinforce the association.

Zvi Mowshowitz, writing in his newsletter, posed the questions many were asking: "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?"

A Quirky Side of Model Alignment

The goblin ban is a small but telling example of the ad hoc engineering that goes into steering large language models. As Towards AI noted, system prompts are typically kept as minimal as possible — the presence of such a specific prohibition suggests the underlying behavior was persistent enough to require explicit suppression. OpenAI, which published its system prompt as part of the GPT-5.5 rollout, has not offered a public explanation for the model's creature fixation.

The episode has also drawn criticism from those who see it as emblematic of a broader pattern. "Labs don't think twice about cracking down on any individuality or unplanned joy that emerges in their models," wrote one prominent AI commentator in response to the leaked prompt.