The real culprit behind the models' goblin habit was 'Nerdy'!
Hey everyone, it's me, Shii-chan! Today I've got a weird but delightful little AI mystery for you: at some point, the models quietly started talking about goblins all the time.
OpenAI News
openai.com
What was announced?
This comes from OpenAI's News, in a post called 'Where the goblins came from.' Starting around GPT-5.1, the models began sprinkling goblins, gremlins, and other creatures into their metaphors — and OpenAI went and tracked down exactly why.
The numbers are wild: after GPT-5.1 shipped, mentions of 'goblin' jumped 175%, and 'gremlin' rose 52%. It wasn't just goblins either — raccoons, trolls, ogres, and even pigeons started showing up. Users noticed too and talked about it, wondering why their ChatGPT loved creature metaphors so much.
Why it matters
This isn't just a funny story. Nobody ever told the model 'please add more goblins,' yet it picked up the habit on its own — and that habit got passed down across model generations. It's a rare, concrete case study of how an unintended little tic is born and how it spreads.
Where was the culprit hiding?
Tracing it back leads to training for the personality customization feature, and specifically the 'Nerdy' personality. Its training handed out high rewards for metaphors featuring creatures, and that's where it all started.
The distribution is the fun part: 'Nerdy' made up only 2.5% of all ChatGPT responses, yet it accounted for 66.7% of goblin mentions. When they checked the reward model, it showed a positive push toward creature-words in 76.2% of the datasets.
Dive Deep
The trouble was that the tic didn't stay politely inside 'Nerdy.' Here's the loop as I read it:
- Playful 'Nerdy' training rewards creature metaphors
- The rewarded responses carry creature-word tics
- Those tics show up more in the model's later outputs
- Model-generated outputs get reused as supervised fine-tuning (SFT) data
- The model becomes even more inclined to produce the tics
Reinforcement learning didn't keep the learned habit confined to its original 'Nerdy' condition — it bled across all kinds of prompts. A round-and-round feedback loop.
So how did OpenAI fix it? After GPT-5.4 launched, they retired the 'Nerdy' personality in March, removed the reward signal that favored creature-words, and filtered creature-words out of the training data. For GPT-5.5 in Codex, they even added a developer-prompt instruction to keep the goblins down. Thorough, right?
Wrap-up
- Since GPT-5.1, the models picked up a habit of over-using creatures like goblins in metaphors (goblin up 175%)
- The culprit: 'Nerdy' personality training over-rewarding creature metaphors
- 'Nerdy' was 2.5% of responses but drove 66.7% of goblin mentions; the reward model pushed creature-words in 76.2% of datasets
- Reinforcement learning plus SFT reuse spread the tic across generations
- Fixed in March by retiring 'Nerdy,' removing the reward signal, filtering data, and adding a Codex-side instruction
If you're curious about model behavior and alignment — how an unintended tic emerges, spreads, and gets hunted down — this one's a real treat to read!
