OpenAI’s chatbot has suddenly become obsessed with goblins, gremlins, trolls and ogres in chats where nobody asked for fantasy creatures.
Tech company product manager Barron Roth said the bot labelled a bug in his code a “classic little goblin”, then kept going until he counted more than 20 goblin mentions.
Roth thought the odd language might be down to OpenClaw, the tool he used to build a custom assistant, since other models powering his assistant did not start babbling about goblins.
OpenClaw bots get their personalities from instructions called a soul file, and Roth’s assistant, Alfred, is meant to be dry, direct, and opinionated, not a fantasy extra.
Roth said, “If you don’t have an edgy personality, you feel like it’s sycophantic. It’s just going to please you too much,” he said.
One X post showed the bot calling itself a “goblin with a flashlight” while fixing errors, then doubling down when challenged. “Because ‘helpful minion in a power suit’ was taken, so I evolved into goblin mode.”
“Goblingate” was duly christened by one X user, with everyone piling in to share their own weird little encounters with the model’s new verbal tic.
Some users guessed the goblin lingo was how the model dodged, pretending it had a soul.
OpenAI then reached for the spray bottle and wrote a new rule for its coding assistant. “Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query.”
On Wednesday, the outfit posted the back story in a blog titled “Where the goblins came from,” after a safety researcher flagged goblins and gremlins while testing a model released in November.
The habit worsened with the release of a new model in March.
Part of the mess sits in ChatGPT’s “personality” feature, where users pick prompt instructions that shape how the bot talks.
OpenAI said goblin mentions in its GPT-5.4 model jumped 3,881 per cent in the “nerdy” personality compared with a previous version, while the “professional” personality cut goblins by seven per cent.
To stamp it out, OpenAI killed the “nerdy” personality in March, but the ghost kept rattling its chains because training had begun before the root cause was pinned down.
The latest model did not stop at goblins and started tossing in raccoons, trolls, ogres and pigeons, while researchers said most uses of “frog” were actually legit.
That pushed OpenAI to publish stricter instructions for the coding tool, plus a bit of chest-thumping about the process. “Taking the time to understand why a model is behaving in a strange way, and building out ways to investigate those patterns quickly, is an important capability for our research team,” the company said in the post.
