OpenAI has had to instruct some ChatGPT models to stop talking about goblins after mentions of the mythical creatures bizarrely increased following an update implemented last year.
OpenAI has had to instruct ChatGPT to stop talking about goblins.
In a blog post on Thursday (30.04.26), the tech company said that it had spotted increased mentions of the mythological creatures – as well as gremlins – in metaphors used by ChatGPTand other tools powered by its latest model GPT-5.
After users and employees noted problems being described as “little goblins”, OpenAI revealed that it would take steps to mitigate the issue – such as telling its coding agent Codex not to refer to them unless it is necessary.
It was found that a “nerdy personality” it had developed for ChatGPT had unwittingly been incentivised to reward mentions of goblins.
The incident showcases the challenges that AI companies face in tackling the possibility of the systems and their training to reward and reinforce errors such as language quirks.
OpenAI explained that it had first noticed increased mentions of goblins, gremlins and other fantasy creatures following the launch of GPT-5.1 last November.
The company wrote: “Users complained about the model being oddly overfamiliar in conversation, which prompted an investigation into specific verbal tics.”
It explained that after a researcher who had noticed “goblin” mentions asked for it to be checked out, developers discovered that the appearance of the term in ChatGPT responses had increased by 175 per cent since GPT-5.1’s launch.
According to OpenAI, “a single little goblin could be harmless, even charming”, but the uptick in their appearance across the firm’s output required further investigation.
The revelation comes after some social media users flagged a bizarre detail amongst lines of code instructing Codex how to behave in certain instances.
As well as being told to avoid platitudes, it said that Codex should “never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it’s absolutely and unambiguously relevant to the user’s query”.
A Reddit user described it as “genuinely insane”.
They wrote: “Why does GPT 5.5 have a restraining order against ‘Raccoons,’ ‘Goblins,’ and ‘Pigeons’?”
Although some speculated that it was a publicity stunt, an OpenAI researcher responded on X by saying: “It really isn’t a marketing gimmick.”
ChatGPT told to stop mentioning goblins
