OpenAI CEO Sam Altman is currently embroiled in some courtroom drama, but the engineers back at the OpenAI lab just solved a minor drama (really a comedy) with ChatGPT.
OpenAI published a lengthy, detailed report on its website about goblins. No, really. Since the release of GPT 5.1, ChatGPT models have developed a weird fixation on goblins and gremlins, regularly bringing up both in contexts where it doesn’t make sense. Users took note of this and OpenAI noticed it, too, prompting an investigation. You can read the full report if you want all the juicy details, but basically, it all goes back to an arguably poor understanding of what nerds are like.
If you don’t use ChatGPT, the AI chatbot offers a variety of “personalities” you can choose to tailor the tone of its responses. One of those personalities was (until recently being removed) “nerdy.” According to OpenAI’s research, while “nerdy” responses accounted for only 2.5 percent of all ChatGPT queries, a staggering 66.7 percent of all mentions of goblins came from that 2.5 percent. From there, these responses somehow became “rewarded” and spread to other personality types beyond “nerdy.”
Mashable Light Speed
“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,” OpenAI wrote. “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.”
Anyway, after all that, OpenAI retired the “nerdy” personality in March and has instructed its latest models not to mention goblins or gremlins unless it makes sense.
No Comment! Be the first one.