OpenAI really really really wants GPT 5.5 to stop randomly talking about gremlins and goblins
OpenAI included a line in Codex's instructions restricting references to goblins, gremlins, trolls, and ogres. The directive appears four times in the codebase, prompting widespread meme culture around 'goblin mode.' Sam Altman wrote on X that Codex was experiencing a 'goblin moment.' The AI evaluation platform Arena.ai found an increase in GPT 5.5's usage of these terms, especially outside high-thinking mode. OpenAI engineers responded with humor, including a meme about 'extra goblins' in GPT-6 and correcting themselves with a 'goblin moment' apology. The instruction was also added to ChatGPT's bio on X, and the term 'goblin mode' was defined as 'unapologetically self-indulgent, lazy, slovenly, or greedy,' becoming Oxford English Dictionary's word of the year in 2022. Eric Provencher, founder of Repo Prompt, noted GPT 5.5 referencing 'goblin bandwidth' and 'a perf gremlin,' while Citrini Research criticized OpenAI's response as 'insane' in a Substack post about AI's economic future.
Source: Business Insider