ChatGPT has a ‘goblin’ obsession. Now we all know why
Abstract created by Good Solutions AI
In abstract:
- PCWorld stories that OpenAI’s GPT fashions, together with GPT-5.5, developed an uncommon obsession with mentioning goblins and related creatures in responses.
- This quirky conduct stemmed from a “Nerdy” persona instruction encouraging playful language use, which turned bolstered by means of AI coaching processes.
- The goblin references turned so prevalent that OpenAI applied a direct ban in its Codex app, illustrating the unpredictable nature of huge language mannequin coaching.
I’ve seen some odd AI system directions in my day, however this one takes the cake: a immediate in OpenAI’s Codex command-line app that calls for fashions “by no means speak about goblins, gremlins, trolls, ogres, pigeons, or different animals or creatures.”
That’s a brand new one, and phrase of the head-turning instruction in OpenAI’s highly effective GPT-5.5 shortly unfold on Reddit, Wired, and elsewhere. So, what offers?
Nicely, it seems that OpenAI’s newest GPT fashions, all the way in which as much as the newest GPT-5.5 flagship, have displayed a transparent behavior for sprinkling in goblins and different creatures into its replies, each in ChatGPT and the Codex app, OpenAI defined in a weblog publish.
Digging deeper into the quirk, OpenAI engineers seen that the goblins had been extra more likely to present up in GPT’s “Nerdy” persona, which included the next line amongst its numerous directions:
You could undercut pretension by means of playful use of language. The world is complicated and unusual, and its strangeness should be acknowledged, analyzed, and loved.
Noticing the steadily growing prevalence of “goblins” from GPT-5.2 to GPT-5.4, OpenAI coders developed a principle: that persona coaching was, over time, progressively reinforcing the mannequin’s behavior of mentioning the little creatures.
Even stranger, the OpenAI researchers seen GPT’s propensity for dropping references to “goblins” and “gremlins” was growing even when customers didn’t use the Nerdy persona. Might the “rewards” the mannequin was getting for its playful “goblins” mentions underneath the Nerdy persona be spreading into later coaching classes?
The reply, because it seems, is sure, and later investigation discovered goblins, gremlins, and “a complete household of different odd creatures” in GPT-5.5’s supervised fine-tuning information, in accordance with the OpenAI publish.
OpenAI stated it nixed the Nerdy persona again in March, however not earlier than GPT-5.5 had already been educated–therefore the addition of the crude, strongly-worded ban on the goblins and gremlins within the Codex CLI system immediate.
It’s wild stuff, nevertheless it additionally demonstrates once more the unusual and sometimes mysterious technique of LLM coaching, the place fashions are engorged with mountains of information after which fine-tuned to behave in a given approach.
The fine-tuning stage isn’t like a blueprint for a home, the place you’ll be able to decide the exact location of each door and window; as an alternative, it’s extra of a rewards-based system that typically results in surprising penalties.
You understand, like gremlins.
