BBC Article
OpenAI tells ChatGPT models to stop talking about goblins
https://www.bbc.com/news/articles/c5y9wen5z8ro
This is a partial GPT-5.5 Thinking translation of the content.
OpenAI has had to issue instructions to some of its AI tools to stop talking about goblins. This is because the company discovered the word randomly mixing into responses.
In a blog post published Thursday (local time), the company revealed that it had detected an increase in references to not only goblins, mythical creatures, but also gremlins in metaphorical expressions used by ChatGPT based on GPT-5, its flagship model, and other tools.
https://openai.com/index/where-the-goblins-came-from/
When users and employees pointed out instances of describing issues as "little goblins," OpenAI stated that it had taken measures to mitigate this problem. This included instructing Codex, the coding agent, not to mention goblins unless relevant.
OpenAI confirmed that the nerdy personality created for ChatGPT had inadvertently been learning to value mentions of goblins positively.
This issue demonstrates the difficulty AI companies face in dealing with the possibility that AI systems and their training methods can encourage and entrench habitual errors.
The company revealed that it first noticed an increase in references to goblins, gremlins, and other creatures following the release of GPT-5.1 last November.
In the same blog post, it explained: "Users complained about the model behaving in an unusually overly friendly manner in conversations, which led to an investigation into specific speech patterns."
It added that after a researcher who had noticed goblin mentions several times requested an investigation, developers discovered that the frequency of the word appearing in ChatGPT responses had increased by 175% since the release of GPT-5.1.
Meanwhile, mentions of gremlins were found to have increased by 52%.
While the increase rate was significant, its proportion in overall responses may not be large. OpenAI stated that while "a single appearance of 'little goblins' in a response is harmless and could even be charming," the increasing frequency of this expression across outputs needed investigation.
[Remaining content omitted]
https://x.com/sama/status/2049692014586048973
