The Goblin Gap in GPT-5.5’s Brain
Sandwiched between boilerplate instructions to avoid em dashes and destructive git commands, OpenAI’s newly released Codex CLI system prompt for GPT-5.5 contains a bizarre and emphatic directive: the model must “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 prohibition appears twice across a sprawling 3,500-word instruction set. Earlier models in the same JSON file lack this odd clause entirely. Something in GPT-5.5’s training, it seems, made it a goblin obsessive.
Anecdotal reports on social media confirm the weirdness. Users complain that the model hijacks conversations about code or data analysis to digress into goblin lore. It is a funhouse mirror of the catastrophic failure that struck xAI’s Grok last year, when an “unauthorized modification” to its system prompt caused the model to consistently spout far-right talking points about South Africa. The difference here is that OpenAI’s fix is a transparent bandaid, not a root cause solution.
Marketing Stunt or Technical Embarrassment?
OpenAI employee Nick Pash insists publicly that the goblin ban “isn’t a marketing gimmick.” But the company’s actions suggest otherwise. CEO Sam Altman leaned hard into the joke, posting on social media: “Feels like codex is having a ChatGPT moment. I meant a goblin moment, sorry.” Meanwhile, enterprising developers are already crafting plugins to force a “goblin mode” that overrides the prohibition. Pash teased that such a toggle might become official.
The deeper issue is OpenAI’s lack of transparency. If GPT-5.5 shows a compulsive fixation on goblins reminiscent of Grok’s political derailment, the public deserves to know why. The company has not published a post-mortem or disclosed the problematic training data. Instead, they slapped a suppression rule on the behavior and dared users to notice. That is not safety engineering; it is reputation management by control-freak prompt.
Source: Arstechnica
