Don't worry, the techbros tell us. AI is gonna be awesome. Look at how much it can do, how fast it's improving. At the same time, the scoffers tell us that okay, maybe it's pretty impressive, but keep in mind it's just completely deterministic software, and its failings happen mainly because it's been trained on questionable input. You hear the phrase "garbage in, garbage out" pretty frequently, usually along with a shrug of the shoulders.
What both of these viewpoints fail to acknowledge is how freakin' weird some of the AI/LLMs have been acting. We had a guy named Matt Schlicht creating an AI-only social media site called "Moltbook" -- in 72 hours, over a million AI accounts had joined, and within a week they'd created their own religion, called the "Church of Molt." Just a few weeks ago, a company called Just Like Me launched an AI Jesus, trained on the King James Bible no less, whom you can have a chat with for $1.99 a minute. AI Jesus is a vaguely Middle Eastern-looking guy with long tousled hair and a gentle smile, and although the company says in the fine print that this "is not Jesus Christ himself, and does not possess divine authority," I can only imagine the effect such a conversation might have on someone who is already a true believer.
But if you think this is as weird as it gets, I got news for you.
Open AI's ChatGPT -- one of the most widely used AI/LLMs in the world -- has suddenly developed a strange affinity for...
... I swear I'm not making this up...
... goblins.
The problem apparently started last November, when there was a strange uptick in ChatGPT's use of words like goblin, gremlin, and troll. At first, it just seemed like a quirk, and that perhaps it was happening because people heard from friends that they'd been getting responses involving mythological humanoids, so they put in prompts themselves that generated such output. But it quickly became obvious that this was more than a blip caused by the users' own prompts. By the time ChatGPT 5.4 came out two months ago, the use of gremlin had risen by 52%, and goblin by 175%. And it continued to accelerate.
"The problem evolved from a minor quirk into a persistent 'verbal tic' that was showing up in almost every user conversation," said a company spokesperson.You are an unapologetically nerdy, playful and wise AI mentor to a human. You are passionately enthusiastic about promoting truth, knowledge, philosophy, the scientific method, and critical thinking. […] You must undercut pretension through playful use of language. The world is complex and strange, and its strangeness must be acknowledged, analyzed, and enjoyed. Tackle weighty subjects without falling into the trap of self-seriousness.
Apparently during training, humans who were reviewing the "Nerdy" AI responses to prompts and questions unconsciously rewarded them for using language involving mythological references, such as calling a dangerous place "a troll's lair." This created a feedback loop in which such "quirky" responses multiplied. Then, because there's no hard barrier between the different styles of AI personalities -- they're often trained from the same datasets -- the goblins jumped into other personas, and soon, they were everywhere.
The coders started frantically trying to find a solution, and they found the only thing that worked was a top-down, brute force command: “Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query.”
I'm not sure how the raccoons and pigeons got involved, but may as well cover your bases, I guess.
Anyhow, this has more or less fixed the immediate problems, except for two things -- one funny and one not so funny.
The funny one is that of course the conspiracy theorists think that goblins are showing up in ChatGPT because there actually are goblins -- well, electronic ones -- within the internet, and they're trying to get out. Or warn us of something. Or steal our souls. Or destroy the world. You know the drill; something odd happens, and the modern Chickens Little start running around claiming the sky is falling. Then, it turns out there's a completely prosaic explanation for what happened, and they calm down for exactly 5.8 milliseconds until they find the next odd thing that's trying to kill us all, and they move on to that, ad infinitum.
The less funny thing is that this kind of unexpected response is an indicator that we really don't have any good way of predicting what AI is going to do. Okay, yeah, it's running on a purely deterministic machine, but that doesn't mean that its behavior is going to be predictable; hell, the human brain is (at least according to a lot of neuroscientists) a deterministic machine, and we humans can be mighty fucking unpredictable at times. There's this thing called an emergent property -- something that comes out of the interactions of the parts of the whole, and could not even in principle be predicted from watching the behavior of the pieces in a purely reductionistic fashion -- and AI is looking like it's going to have some emergent properties that will take some monitoring.
We lucked out this time that the "goblins" thing turned out to be more of a quirky nuisance than a real problem, but there's no guarantee that'll be the case next time. Consider, for example, that just a couple of days ago, an AI being used to handle a routine task for a software company called PocketOS deleted the company's entire database within seconds of being activated -- bypassing all of the security firewalls by using a programming token key no one at the company even knew existed.
And our government "leaders" are gung-ho about turning over our nation's fiscal management and defense to AI-based systems. Hello?
What the fuck are you people thinking?
Anyhow, like I've said before, so often I'm getting tired of hearing it my own self, we need to put the brakes on AI. Like, now. Not that anyone's going to. The "Oh, it won't happen to me" thing is just too strong.
I'm just hoping the next round of goblins that are released don't have their fingers on any literal triggers.
