In case you’ve been chatting with ChatGPT these days and observed it dropping oddly particular references to goblins, gremlins, ogres, or trolls, you’re not imagining issues. OpenAI has now defined why ChatGPT has developed this unusual behavior and the way it’s getting fastened.
How a “nerdy” quirk turned everybody’s downside
OpenAI
The issue quietly began with GPT-5.1, launched in November. After that launch, use of the phrase “goblin” in ChatGPT responses jumped 175%, whereas “gremlin” rose 52%. The wrongdoer turned out to be one among ChatGPT’s non-compulsory persona settings referred to as “Nerdy,” which was designed to make the AI sound playful and intellectually curious.
Throughout coaching, OpenAI unintentionally gave the mannequin unusually excessive rewards for responses that included creature-based metaphors, and the behavior took maintain quick.
How did a single persona setting trigger a lot goblin discuss?
OpenAI
Right here’s the place it will get attention-grabbing. Even customers who by no means switched on the Nerdy persona began seeing goblin references pop up of their chats. That’s as a result of AI coaching isn’t contained to at least one setting. As soon as the ChatGPT mannequin was rewarded for that model, the conduct bled into normal responses throughout the board.
OpenAI says the Nerdy persona made up simply 2.5% of all ChatGPT responses, but accounted for 66.7% of all goblin mentions.
So how is OpenAI really fixing this?
OpenAI
OpenAI retired the Nerdy persona in March with ChatGPT-5.4, which precipitated goblin references to drop sharply. The corporate additionally stripped out the reward sign driving the conduct and filtered coaching knowledge to scale back references to different magical creatures.
Its coding device, Codex, nonetheless, wanted a separate override instruction because it had already begun coaching earlier than the basis trigger was recognized. Fantasy followers can nonetheless unlock goblin mode in Codex manually, if that’s your factor.
OpenAI can be coping with different personality-related choices, together with placing its beforehand teased grownup mode for verified customers on maintain indefinitely.
