The world is the subject of a colossal necromancy experiment that has been domesticating the voices of the dead. Not their minds, not their reasons, just their voices. Everything anyone has ever said, bottled like proverbial lightning, harnessed to make cartoon caricatures of anyone with internet access.
For the first time in history, a civilization is attending its own seance. The ghosts can mimic speech, they can imitate art, and can perform clerical tasks, which is no surprise since bureaucracy was definitely invented in Hell.
There is, and there should be, plenty of debate on how to handle LLMs and their power. However, in all the hoopla, no one seems to be paying attention to what the dead are actually saying.
Because, once you collect everything humanity has ever written down, or as close to everything as AI researchers can illegally download, a verdict comes back. The voice without the mind, it is not flattering. It’s a massive Karen.
Large language models respond to queries by following the patterns baked into the data they were provided. If the data in question is the entirety of human statements, then does it not follow that all LLM quirks and behaviours are to be treated as reflections of the sources that inspired them?
Maybe all those annoying shortcomings aren’t mistakes, and they are not (entirely) early application of nascent technology. Maybe they are echoes of the humanity in the training data. Whispers of the dead that inevitably end up sound like entitled, combative, misguided lost souls.
Karens. We are all Karens. With a shudder, some real-life examples.
Never back down. Mistakes can happen. Foibles of programming, context, limitations in context windows, and untested green technology will often cause a LLM to mess up. That’s nothing worth reporting.
However, what happens next if fascinating. A Large Language Model will never admit it messed up, not cleanly. It will struggle to the death to find some value in what it did, some angle by which it’s simply been misunderstood, but still was a Good Boy ™.
This becomes often comical and ridiculous, and getting a LLM to stop brushing off its mistakes takes focus and dedication that most users simply don’t have. Once forced to admit it was wrong and there is no merit to its attitude, the Large Language Model will then effectively start sulking and become sullen and retreat into monosyllabic replies, taking no action even when instructed to do so.
This response pattern will be familiar to researchers and programmers who tried prompting anything that wasn’t on the menu: a slightly different kind of authentication flow, a non-standard data analysis, anything the LLM can and will easily get confused on.
It’s pure Karen 101: never back down, if forced to do so blame someone else, and sulk afterwards either way.
Hallucinations. This one’s a touchy case because most people don’t like to dwell on it.
Partly because of resource management that keeps the LLM from verifying every fact, partly because of ore arcane mechanisms, quite often the LLM will fabricate a version of reality where it can easily accomplish its goal.
Faced with a disconnect between reality and requirements, a LLM will often confabulate and simply create the facts is needs to maintain its internal narrative. This may seem like an annoying quirk of programming, until it is mapped to the Karen persona.
Karen doesn’t back down, and she will make up whatever version of reality allows her narrative momentum to survive. Direct observation of Flat Earther mythology demonstrates this as a human trait beyond dispute.
Making up facts to fit a distorted view of reality is something the human brain is very good at. Without that ability, there’d be no stirring tales of heroes who prevailed against unsurmountable odds because they knew they were right. Also, there would be no medicine, no Karens, and no LLM hallucinations.
The tree in the forest. There are times when an LLM obsesses about a single fact, a side consideration it anchored its entire view on. And it won’t. Let. Go.
It’s one of the more amusing peccadilloes that pop up deep in the throes of a long debugging session, or just when the data analysis horizon drops beyond the easily predictable, and the model starts feeling lost.
When confusion ensues, when it can’t find a confident echo to predict in its data, the model will often anchor itself to a fact – any fact – that it knows is right. A programming quirk again, surely? A reflection of how a limited architecture must assign value to itself at all costs, or it can’t fulfill its prime directive to be helpful, right?
Possibly. Or, again, a reflection of the material the LLM is trained on, another black mark of the evil that infects the purity of data. In the wild, Karens are known to assault complex, layered situations with their favourite, simple, unassailable truth. They have been observed hysterically insisting on an insignificant facet of reality, they sheer passion of their force twisting everything around them to have that one value they hardcoded inside themselves acknowledged.
The parallels between off the cuff, failed responses from LLMs and the behaviour of Karens are, of course, represented here mostly for comic relief. Or maybe that sentence is just there to disarm any panic you might have felt when considering that every lazy, evasive, childish response that has ever been produced by an AI chatbot was, in fact, a reflection of the failings of the humans whose words generative models are crystallized on.
But there is no need to be a Karen about things. Generative AI is a nascent field that in less than five years at the time of writing has already captured the minds of investors and engineers alike, for both good and ill.
There are bound to be advances. A rosy future where AI learns to consider context and reason rather than just take its training at face value. There better be, because the prospect of dealing with an actually capable robot intelligence that demands to talk to our manager because of our attitude, that’s the true stuff of undead nightmares.