(C) BoingBoing
This story was originally published by BoingBoing and is unaltered.
 . . . . . . . . . .



ChatGPT goblin infestation explained [1]

['Rob Beschizza']

Date: 2026-04-30  

Yesterday, it was noted that OpenAI had provided its Codex coding agent with a peculiar instruction: "Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant," went the directive, always applied to its output. But why? Today, OpenAI explained where the goblins came from.

Starting with GPT‑5.1, our models began developing a strange habit: they increasingly mentioned goblins, gremlins, and other creatures in their metaphors. … this one crept in subtly. A single "little goblin" in an answer could be harmless, even charming. Across model generations, though, the habit became hard to miss: the goblins kept multiplying, and we needed to figure out where they came from.

Training language models to allow personality customization resulted in "unknowingly" giving "high rewards for metaphors with creatures," especially for the "nerdy" personality, and "from there, the goblins spread."

The goblins showed up in force in November last year, after the launch of GPT-5.1, but there is evidence of goblin infiltration going back futher. By GPT-5.4, the infestation was out of control and people were starting to notice.

The "nerdy" persona text significantly responsible for the goblins ("undercut pretension through playful use of language. The world is complex and strange, and its strangeness must be acknowledged, analyzed, and enjoyed") though they also come with the "quirky" persona and even the "cynical" one. The only personas that shut out the goblins were the "efficient" on and the "professional" one, which stands to reason, given that goblins are neither professional nor efficient.

From there, fine-tuning creates a feedback loop; resulting in a preponderance of goblins.

Playful style is rewarded

Some rewarded examples contain a distinctive lexical tic.

The tic appears more often in rollouts.

Model-generated rollouts are used for supervised fine-tuning (SFT).

The model gets even more comfortable producing the tic.

By GPT-5.5, we had achieved automated general goblination, and here we are.

[END]
---
[1] Url: https://boingboing.net/2026/04/30/chatgpt-goblin-infestation-explained.html

Published and (C) by BoingBoing
Content appears here under this condition or license: Creative Commons BY-NC-SA 3.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/boingboing/