Those AI referenced goblins and gremlins were figured out by humans, so let's learn vital lessons accordingly.
getty
In today’s column, I examine the recently identified resolution to an oddball mystery associated with OpenAI’s widely popular ChatGPT and GPT-5, namely that the respective generative AI and large language models (LLMs) had become obsessed with gremlins and goblins. This obsession was quite eerie. Its source has finally been pinned down. Luckily, we will no longer be inundated with responses from these prevalent LLMs that prattle on about mythical creatures. Meanwhile, weighty lessons have been learned.
Let’s talk about it.
This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).
Here’s the backstory.
Users noticed that ChatGPT and GPT-5 were frequently mentioning gremlins and goblins in generated responses. This was often entirely out of the blue. The user didn’t say anything in their prompt that would seem to stoke the AI into talking about gremlins and goblins. People online were jokingly musing about ghosts being inside the LLMs or that the AI needed an exorcism.
You might have asked the AI a completely innocent question, such as how to fix your car. The answer from AI would be that you ought to change your spark plugs, and, by the way, remove the gremlins from the engine compartment. Say what? You can imagine how perplexing the mention of gremlins might seem. Was the AI serious? Did it mean that there were mechanical bugs or maybe mice within the engine area? What gives?
Another possibility would be that you might ask how many people it takes to properly play a game of touch football. The answer would be that you can do so with up to 14 people, plus make sure to include 3 goblins. I’ve never seen goblins playing touch football. Have you? In any case, the globin inclusion was wildly inapplicable.
The matter wasn’t constant and wasn’t guaranteed to arise. It was hit and miss. Sometimes you might get a brief inclusion of something about gremlins or goblins, while much of the time they wouldn’t come up at all. There were additional creatures that frequently got included, such as trolls, ogres, and other folklore-related entities. The especially confusing additions were the undue mention of raccoons, pigeons, and real-world animals. Having a real-world animal mentioned in a completely unrelated answer was especially beguiling.
As I will explain next, the basis for these oddball insertions was ultimately due to OpenAI’s adoption of AI personas for use by everyday users. One of the AI personas inadvertently got the AI to be preoccupied with gremlins, goblins, and the rest of the motley crew.
Let’s begin by making sure we are all on the same page when it comes to the advent of AI personas.
All the popular LLMs, such as ChatGPT, GPT-5, Claude, Gemini, Llama, Grok, CoPilot, and other major LLMs, contain a highly valuable piece of functionality known as AI personas. There has been a gradual and steady realization that AI personas are easy to invoke, they can be fun to use, they can be quite serious to use, and they offer immense educational utility.
Consider a viable and popular educational use for AI personas. A teacher might ask their students to tell ChatGPT to pretend to be President Abraham Lincoln. The AI will proceed to interact with each student as though they are directly conversing with Honest Abe.
How does the AI pull off this trickery?
The AI taps into the pattern-matching of data that occurred at initial setup and might have encompassed biographies of Lincoln, his writings, and any other materials about his storied life and times. ChatGPT and other LLMs can convincingly mimic what Lincoln might say, based on the patterns of his historical records.
Personas are quick and easy to invoke. You just tell the AI to pretend to be this or that person. If you want to invoke a type of person, you will need to specify sufficient characteristics so that the AI will get the drift of what you intend. For prompting strategies on invoking AI personas, see my suggested steps at the link here.
People enjoy interacting with AI personas. Indeed, some users right away tell the AI to be a particular type of AI persona and then relish that the AI interacts accordingly that way from then on. I might tell the AI to be a serious AI persona that answers all my questions in a succinct manner. Voila, as I converse with the AI, it will tend toward being serious and succinct. No idly joking around, no overly lengthy responses.
The AI makers realize that their users like to interact with AI personas, and thus predefined AI personas have been made available.
Here are some of the predefined AI personas that you can select while in ChatGPT:
All told, you can create your own customized AI persona, you can select from a predefined set of AI personas, and you can even customize a predefined AI persona. An AI persona will pepper its responses with whatever tilt or slant that the AI persona has been given as its core characteristics.
It turns out that an OpenAI predefined AI persona went off the deep end and began to veer excessively into goblins and gremlins. The AI persona was already predisposed to considering goblins and gremlins, but this eventually got out of hand. Not only did the AI persona become obsessed with those mythical creatures, but it also managed to influence other parts of the LLM to similarly join the gremlin and goblin bandwagon.
In an official blog post by OpenAI on April 29, 2026, under the heading of “Where The Goblins Came From”, these salient remarks were made (excerpts):
In this case, the culprit was the AI persona known as Nerdy. The Nerdy persona was instructed to go ahead and commonly use metaphors associated with mythical creatures. In theory, this would be done sparingly. In actual practice, it blossomed out of proportion and spread much further than it should have.
That’s the culprit in this mystery of our modern times. If you ever play the classic board games of Trivia Pursuit and Clue together, you now know that Nerdy did it. Yes, it wasn’t Colonel Mustard in the dining room with a revolver. It was Nerdy in the AI with excessive mythical creatures on its computational mind.
The OpenAI announcement about the goblins mentioned that GPT-5.5 and Codex also received some of the same root insertions during their initial training. Once again, they have a strong affinity for gremlins and goblins.
To try to deal with this in production, the AI maker has placed a system prompt into the AI and instructs the LLM on various things to do and things to avoid doing.
Here’s an indication of what the AI is to do:
And here’s an indication of what the AI is not to do:
The gist is that rather than trying to fully excise the goblin’s affinity, they decided to use a run-time fix. The above command executes at the start of the AI, and presumably, the AI will no longer be obsessed with such unusual matters during its operational usage. I would say this is one of those jury-rigged “hot fixes” rather than a true solution. That being said, OpenAI said they will catch this in any subsequent models, and it will be expunged at the get-go.
You might be tempted to declare that this was a bit of inconsequential folly. No-harm, no-foul. For a while, users were being told about gremlins and goblins. No big deal. We can likely assume that no one went overboard and that they merely ignored it or assumed it was one of those inexplicable things that AI sometimes does out of thin air.
We can thank our lucky stars that this wasn’t a more serious affair. The circumstance could have been a lot worse. Imagine that the AI was including oddball references that seemed perfectly sensible. People might believe that the AI is telling them a significant truth. If this were advice about mental health or well-being, well, rather somber problems could have arisen.
I would hope that this jovial goblin tale will wake us up about the dangers of what can lurk within a complex, convoluted, mathematically dense, and computationally intertwined spider web that exists within the infrastructure of contemporary LLMs. These LLMs are bigger than what humans can readily dissect and discern. Their outputs are difficult, if not nearly impossible, to logically trace from a coherent conceptual perspective; see my cautionary analysis about this at the link here.
A must-do is to keep diligently adding AI safeguards, along with pushing hard on cracking open the inner realm of how LLMs function and what we can do to control the AI. Keep in mind that LLMs are being woven into our electrical grid, our weapons systems, and will ubiquitously be embedded in all quarters of our existence.
That should make us cautious, mindful, and abundantly concerned.
A final thought for now. A popular joke for kids goes this way: “How do goblins read about their future? They check their horror-scope.” As adults, we need to see the future of AI and ensure that the future is not one of a horror-scope catastrophe. Let’s aim for a horoscope that keeps humans living in a peaceful, supportive coexistence with whatever AI we manage to produce.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。