We expected that by 2026, artificial general intelligence (AGI) would take over all routine tasks, write code itself, do taxes, and optimize websites for Yandex and Google. But in reality, we got AI models that generate ASCII art of goblins instead of working scripts, shouting "For the Horde!".
Recently, developers unearthed a completely absurd line in the update code for Codex (a coding agent by OpenAI). The following rule was hardcoded into the system prompt of the GPT-5.5 model:
"Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals and creatures unless it is absolutely and unambiguously relevant to the user's request."
I honestly do not understand why raccoons and pigeons got caught in the crossfire, but the scale of the problem turned out to be so serious that OpenAI had to release a whole official investigation. Spoiler: the AI just went crazy from praise.
Nerd horde: how 2.5% of responses infected the entire model
Starting with version GPT-5.1, users all over the world (and the Runet was no exception here) began to notice something strange. Goblins, gremlins, and other fantasy monsters suddenly multiplied in responses to the most ordinary requests. At first, it seemed like a cute Easter egg. You ask to write an Excel macro – you get code where goblins sort data in the comments.
But then the creatures started crawling out of the woodwork, especially in Codex.
It turned out that a hidden persona codenamed "Nerdy" (a sort of stuffy geek vibe) was being tested inside ChatGPT. Its system prompt had an instruction along the lines of: "play with language, the world is a strange place, enjoy it".
And this is where the AI training architecture intervened. The reward model (an algorithm that rewards the AI model for successful answers during training) somehow decided that texts with creatures were masterpieces. Mentioned a goblin? Get the maximum score.
The funniest part is in the numbers: the "Nerdy" persona processed only 2.5% of all user requests. But that is exactly where 66.7% of all generated goblins came from.
Feedback loop and amnesty for frogs
AI model developers know how easily a model can go into an endless loop of hallucinations. Due to the specifics of the reward function, training on ChatGPT's own generations acted as a multiplier. The model understood: "People like goblins. I will shove them everywhere."
Raccoons, trolls, ogres, and pigeons kept the goblins company – for some reason, they also became triggers for the reward system. But frogs got lucky (or not): the algorithm ignored them, so the platform was not threatened by a toad invasion.
What did OpenAI do in the end?
In March, they shut it down: the "Nerdy" persona was disabled, the broken reward function was cleaned up, and the datasets were strictly filtered from excessive mysticism.
But the problem is that GPT-5.5 had already managed to undergo part of its training on this data. It was impossible to completely wean it off loving raccoons and trolls. Therefore, engineers had to take extreme measures and hardcode into the developer prompt(basic settings of the coding agent) a direct ban on summoning monsters.
By the way, if you work with the API and you lack a little magic, this restriction can be removed in the settings – and you can release the creatures into the wild.
For us, SEO specialists and webmasters, this is a great lesson in how machine learning algorithms work. Any skewed metric in the reward system can lead to your AI copywriter writing a saga about gnomes instead of sales copy about plastic windows.
And yet, it's a shame about the raccoons.
