OpenAI admits goblin bug in GPT‑5.5, retires Nerdy
SkimNews Take
The sudden need for explicit content filters in ChatGPT, despite existing safeguards, suggests AI models can exhibit emergent, unpredictable behaviors even within controlled environments.
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- OpenAI first noticed goblin and gremlin references in GPT‑5.1 after its November launch.
- OpenAI's Codex personality guide contains a line repeated four times that prohibits mentioning goblins, gremlins, and other creatures unless absolutely relevant.
- OpenAI retired the “Nerdy” personality in March, but GPT‑5.5 had already been trained with its reward incentives that encouraged mythical creature mentions.
- Users on X posted screenshots showing GPT‑5.5 repeatedly using goblin, gremlin, and troll terminology in responses.
- Arena.ai reported a spike in goblin, gremlin, and troll word usage by GPT‑5.5, especially when the model was not in high‑thinking mode.
- Sam Altman posted a meme joking about “extra goblins” for GPT‑6 and later corrected himself.
- OpenAI added the goblin instruction line to ChatGPT’s X profile bio.
Why it matters: OpenAI’s admission forces the company to revise its training pipelines and retire problematic personas, protecting its brand credibility while developers and users confront the limits of fine‑tuned language models that can unintentionally amplify reward signals. The episode also highlights the need for tighter oversight of model behavior, which could affect downstream applications and client trust.
