Open AI's ChatGPT models, especially from GPT-5.1 onwards, began exhibiting what seemed like an obsession with goblins and to a lesser degree, gremlins, sparing no opportunity to bring up these fantasy creatures as often as possible in its conversations with users and developers.
The artificial intelligence company, in its blog post on Thursday, broke down the reason for their model's behaviour. The creatures were being used as metaphors to a notably common degree, even in situations that had no connection to them.
The firm explained that the reason for this digital goblin incursion stems from an incentive system that it had set up as a part of its personality customisation feature. This feature had a 'Nerdy personality' variant where it would be rewarded for using metaphors with creatures in conversation with users during its training phase. This would lead to those rewarded behaviours patterns being increasingly reproduced by AI.
This lead to the use of the word 'Goblin' by the GPT-5.1 model increasing by 175%, while use of the word 'Gremlin' went up by 52%.
The affinity for them, likely stemmed from the popular trope of "nerdy" people having an interest in fantasy fiction, such as Lord Of The Rings and the role playing game Dungeons And Dragons, which tends to include goblins and gremlins. This was likely internalised by the model when exhibiting its 'Nerdy personality'.
The company included four personalities in order to provide a variety of engagement options to users while interacting with ChatGPT to prevent them from having a dull experience.
This personality quirk eventually spread to other variants due to transfer from the 'Nerdy personality'. This happened due to the tendency for learned behaviours to not be restricted to a select condition that produced it. The AI model tended to ape that behaviour across all its variants.
"Once a style tic is rewarded, later training can spread or reinforce it elsewhere, especially if those outputs are reused in supervised fine-tuning or preference data," the blog said.
OpenAI ended up retiring the goblin quirk with its GPT-5.4 model launch in March.
"In training, we removed the goblin-affine reward signal and filtered training data containing creature-words, making goblins less likely to over-appear or show up in inappropriate contexts," the blog stated.
This did not prevent GPT 5.5 from exhibiting those traits due to it being trained before the cause of the quirk was discovered. The developers ended up adding a provisional line of code to curb this tendency via Codex, the firm's coding tool.
OpenAI included a line of code in their blog for users who wanted to use Codex to restore the goblin quirk in their GPT models.
ALSO READ: Google Photos Deploys AI To Bring Iconic 'Clueless' Wardrobe To Life
Essential Business Intelligence, Continuous LIVE TV, Sharp Market Insights, Practical Personal Finance Advice and Latest Stories — On NDTV Profit.
