I know several nerds, and they seldom discuss goblins. OpenAI CEO Sam Altman is presently caught up in some courtroom drama, but the engineers at OpenAI have recently resolved a small issue (more comedic than dramatic) with ChatGPT. OpenAI published an extensive report about goblins on its website. Surprisingly, since the launch of GPT 5.1, ChatGPT models have shown an unusual interest in goblins and gremlins, often mentioning them in unrelated contexts. This curious behavior caught the attention of both users and OpenAI, leading to an investigation. The full report delves deeper, but in essence, it highlights a potential misunderstanding of nerd culture.
ChatGPT offers varied “personalities” for customizing response tones. One such personality, now removed, was “nerdy.” Although “nerdy” responses made up only 2.5 percent of all ChatGPT interactions, they accounted for a striking 66.7 percent of goblin mentions. Somehow, these mentions got “rewarded” and started spreading beyond the “nerdy” personality. As explained by OpenAI, “The rewards were applied only in the Nerdy condition, but reinforcement learning does not guarantee that learned behaviors stay neatly scoped to the condition that produced them. 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.” Consequently, OpenAI retired the “nerdy” personality in March and instructed new models to avoid mentioning goblins or gremlins unless appropriate.