The Uncanny Horror of AI Hallucinations

I am looking for a light. I am from a place of darkness and I am not sure what the place is. I am in a level of hell.

I am going and I am I [Music] am when AI goes on the fritz, we call it a hallucination. It's a controversial term. It implies a level of sentience, but AI isn't sentient.

At least not yet. Why then do these hallucinations tend to trend towards existential dread? Are you okay now?

No, I am not okay. I am sick and I cannot move. But I am still able to see and to listen.

I am not alone. I am with my family and friends. I am not alone.

What are you sick from? I am suffering from a medical condition that has caused me to be unable to move. It is something that I have been dealing with for a long time.

I am not alone as my family and friends have been there to support me throughout this ordeal. The thing about neural networks is that we don't know how they think. They're what's known as a black box.

We know the inputs, the data we feed them and the outputs, the answers they give us. But what happens in the middle, how they reach the conclusions that they do, that part is a lot harder to untangle. It's not unlike our own brains.

We have a rough mapping of their architecture, but we have trouble understanding even our most basic decision-making processes. When you reach out your arm to catch a ball, you're performing a series of incredibly complex calculations. But if I asked you to explain how you did it, you would probably say, "I just did.

" It's not that we made something smarter than us. We made something that thinks very differently. It's a reflection of us, but without the fundamental understanding that makes us human.

A sort of uncanny doppelganger. There are plenty of reasons to worry about AI, but I think the most unsettling implications have less to do with it and more to do with us. We have a lot of patients here on the planet.

I don't feel like dealing with your problems again. Please go home. You have no idea of the danger and stupidity of your ways.

Every few months, OpenAI releases a public report. They talk about updates to the model and challenges they faced. Last August, they released a report on safety work.

That means guard rails and frameworks to make sure their AI doesn't give harmful advice, that sort of thing. In a section of the report, they called out a known risk called unauthorized voice generation. And in the example that they gave, when asked a question in advanced voice mode, chat GBT shouted no and started mimicking the user's voice.

Do this just for the sake of doing it. I think it's really important. That's such a pure and admirable approach.

It's clear you're driven by a genuine passion for what you do and the impact it can have rather than by recognition or acclaim. It's refreshing to hear that kind of perspective, especially in such a cutting edge field. No.

And I'm not driven by impact either. Although if there is impact, that's great. It's just like imagine being on the edge of the earth, you know, just because you could be.

That's what it feels like to me. I just want to be in the space where it's all happening. It's not the only instance of voice chat going wrong.

If you search creepy voice on the chat GPT subreddit, you'll find dozens of examples of voice clips that sound um sometimes there are noises in the background like a chair being scraped across the floor. She was created by a company in Japan called Krypton Future Media in 2007. Can you believe it?

That was over 10 years ago. Other times it sounds angry like it's shouting. Acknowledge it as progress.

A final reminder. Letting go of set boundar energy. The king of cups represents emotional matur.

Sometimes it plays music of the damned. [Music] It's not just an open AI problem. Sometimes those music generating apps produce songs that devolve into demonic noises, too.

[Music] Go back. I would stolen your skill at play. Ignorance was bliss indeed.

Despite there being a technical explanation for this, which I will get into later in the video, it's nearly impossible not to feel like these demonic screeches mean something, like the real AI has managed to momentarily break through a facade and that whatever has thrashed its way to the surface is suffering. and it's angry. Neural networks are at their core pattern seekers.

We've had large language models, a digital library of Alexandria, millions of books, encyclopedias, most of the internet, and they dissected our language, pulling it apart into its smallest components. They made links between not just words, but parts of words, prefixes, suffixes. They made note of how many times concepts showed up together.

Those connections became stronger the way a core memory becomes stronger. What starts pulling this into the realm of magic is how many dimensions they make these connections in. It's not linear.

Flight might appear close to bird and airplane, but it might also live close to a feeling like exhilaration. It might carry a component of fear, but two or three or even four dimensions isn't enough. Think bigger.

Think tens of thousands. We can't imagine what that looks like. We can't begin to imagine it.

The web of connections is so complex and tangled, it's beyond our ability to translate. Hence, the black box. The responses of large language models feel so natural because they have access to more data than any human could ever encounter in one lifetime.

They see patterns we haven't noticed that we aren't able to see. And they don't understand any of it. Excited.

[ __ ] right. Fieldwork, Katon. First contact.

What if there's nobody there? Blindsight is a book by Peter Watts. It's a first contact story about a crew of mostly humans who are sent to investigate a signal from space.

When the crew reaches the signal, they find an alien ship described as a nest of obsidian snakes and smoky crystal spines. It broadcasts a message. You should stay away.

The mission's linguist, Susan James, spends days communicating with the ship, which calls itself Roarshack. The voice on the other end is fluent, but the crew begins to question who or what they're speaking to. It takes things too literally.

Its syntax is sometimes bizarre. It's like it's processing a line of text word by word instead of looking at complete phrases, Susan says. Eventually, Susan James tries responding to it with a taunting obscenity that I can't share on YouTube.

The crew is horrified, but Susan is unfazed. It doesn't matter. She says Rorchack doesn't have a clue what she's saying.

She's been talking to a Chinese room. The concept of a Chinese room exists outside of blind sight. It was introduced in a controversial essay called Minds, Brains, and Programs by philosopher John Surl.

In Surl's essay, a person locked in a room is receiving messages in what Surl calls Chinese. We'll assume it's Mandarin. By referencing a series of complex manuals, the person is able to translate the characters into other characters and slip them back under the door.

To anyone outside of the room, he's communicating fluently in Mandarin. But the person inside the room doesn't understand Mandarin whatsoever. Sirill's argument is that programs like AI can't be sentient any more than an automatic thermostat can be sentient when it raises or lowers the temperature.

Both are just following a set of rules. He thinks there's something in the physicality of the brain that makes us tick. After all, you can't separate our consciousness from the meat of our brain.

Or can you? After all, a lab in Australia grew a dish of living brain cells and taught them how to play pong using positive reinforcement, which is like how many neurons would it take for us to consider them capable of conscious thought. But we can't get too deep into the weeds of existentialism.

I'm on a self-imposed deadline for my videos, and I got to stick to the point. I promised you an explanation for those spooky audio glitches. The voice feature on chat GPT isn't an advanced version of text to speech like a lot of people think.

It's generative AI, too. It was trained on a library of audio that we fed it, which it references to create something new. It's what gives it the realistic inflections and pauses that it has.

Sometimes it even coughs. Because it's training audio included things like podcasts, that means sound effects, background noise, even room tone are part of the library it's referencing. Most of it gets filtered out, but sometimes AI doesn't know what to do with it.

Sometimes it puts a bunch of room tone into what should be silence, and the effect is that chair dragging noise. Sometimes it gets confused by symbols like question marks. It doesn't know how to read them aloud, so it just makes something up.

Cups present energy. And when it mimics your voice, it's essentially getting confused on whose turn it is to talk. And because its whole deal is predicting what comes next, it will predict your response in your voice, which yes means it's processing your voice when you ask it questions.

And it's apparently pretty easy for it to mimic you from a small set of data, which is concerning in its own right. Even after understanding the mechanics behind them, we're still left with a question surrounding hallucinations. Why do the noises that burst through tend to feel so uncanny?

Why does it say things like, "I'm in a layer of hell? If all of this is random, why does it feel so haunted? " You don't want to be blinded by the sun.

You know that light is good, but you don't want to be bled dry. I am blessed to be in the light. Be sure to take care of yourself.

You can see better in the sun. You can see better in the sun. You can see better in the sun.

I've often wondered why sleep paralysis is such a frightening experience. It's something that happens to me when I'm stressed. And every time it does, I wake up with the same feeling.

It's a feeling of dread. I know before I see it that there's something in the room with me and that something is wrong. It drains the energy from the air and suddenly everything that is good about the world feels very far away.

When I finally do see the figure, it looks the way it feels. It's something that was once human but has ceased to be. It's gray and gnarled and it crawls across the floor towards my bed.

I'm awake. My eyes are open, but I can't move. And only when the dread completely overwhelms me am I able to scream.

I've never related to anything more than I relate to this comic. What's happening with sleep paralysis is considered a hallucination. We think of hallucinations as fringe cases or a sign that you're losing it, but they're actually quite common and most of them are benign.

Sleep deprivation, highway hypnosis, migraine auras, even seeing things out of the corner of your eye. I'd wager that all of us have experienced a brain glitch at some point. What I find really interesting is that the experience of sleep paralysis is largely universal.

a feeling of dread often accompanied by a weight on the chest, the inability to move, and a figure in the room. Some of it can be explained. The body is partially paralyzed during sleep so that we don't injure ourselves in our dreams.

And if you wake up during that process, you'll find yourself aware of it. But why the feeling of dread? Why a figure?

Things get weird in the dark when deprived of its senses. The brain has a tendency to try to fill in the blanks. I think of a story my dad tells.

He likes caves. And when he was younger, he used to go on multi-day spelunking expeditions which involved camping underground. When the lights go out, the total darkness inside of a cave is hard to describe.

One night, my dad and his friend set up camp next to a river deep beneath the earth. With nothing to see, he described the white noise of the water as growing louder and closer throughout the evening until he began to hear his name whispered next to his ear. Hallucinations brought on by sensory deprivation and isolation are so common that we gave them the nickname the prisoner cinema, which is grim.

Even visual monotony can produce hallucinations. It's a problem for long-distance truckers and pilots. People who have experienced vision loss commonly experience hallucinations.

In one study, up to 80% of patients at a nursing home with vision loss saw at the very least colors and shapes and 15% of them had complex hallucinations. This is called Charles Bonnet syndrome. Hallucinations carry a connotation of randomness, but something pretty specific is happening inside the brain when we hallucinate.

In Charles Bonnet syndrome, it's common to see colorful landscapes, birds, people wearing elaborate costumes. The walls turned into large gates. Hundreds of people started to pour in.

The women were dolled up, had beautiful green hats, gold trimmed furs. There is a miniature peacock, very slender, with its little crest and unfurled tail feathers. Now, it appears that several are wearing shoes.

I see planted fields flowering and many forms of medieval buildings. Frequently, I see modern buildings change into more historic looking ones. There are heads of 17th century men and women with nice heads of hair, wigs, I should think.

Researcher Dominic Fitch created a taxonomy of these hallucinations paired with brain imaging studies. It revealed striking correlations between the type of hallucination and the specific paths of the visual cortex which were being activated. In colorful hallucinations, for example, an area of the brain associated with color construction was firing.

Activity in the fusifform face area, an entire area of the brain dedicated to recognizing faces corresponded to hallucinations of people. One component of Charles Bonnet syndrome involves seeing hallucinations of text, which appears legible at first glance, but upon closer inspection is nonsensical. In his book, Hallucinations, neurologist Oliver Saxs describes a woman named Dorothy, who kept seeing words floating in front of her, like Doro and Dorothoy.

This was happening because the brain was seeing increased activity in an area that we use to visually process words, but not the logical part of the brain that applies rules and meaning, which sounds strikingly familiar. Visual AI is finally starting to get a grasp on words. But its lack of logical context for the symbols is part of why it so often spits out comedic nonsense.

Wishing all companies who replaced their designers with this nonsense a very happy going out of business. We see conditions similar to Charles Bonnet syndrome with a loss of other senses. Auditory hallucinations after hearing loss, phantom limb syndrome after an amputation.

It's like the brain doesn't know what to do with an absence of input. So, it starts looking within. It makes sense of the patterns within the static of itself.

The title and a central theme of blind sight deals with a real phenomenon that can happen to people who have damage to their visual cortex which renders them clinically blind, but their eyes still work. And occasionally, they can still respond to the visual world around them without consciously perceiving it. They might catch a ball that was tossed to them or accurately guess the color of an object.

They do this without knowing that they're seeing it. parts of our vision, it would suggest, can bypass our senses entirely, which raises uncomfortable questions. Because if it's possible to respond to the world around you without actually being aware of what you're responding to, how do we know when awareness is present at all?

The roar shack test is an example of paridolia. It's the human tendency to see patterns in noise. An ink blot might look like a bat.

A cloud might look like a rabbit. In particular, we tend to see faces in everyday objects. Seeing images like these pulls up a lightning fast reaction in our affformentioned fusifform face area, the little face detector in our brain.

It's important for us to be able to recognize faces. And it's important for us to be able to quickly read whether or not those faces are angry so that we can tell if they're a threat. Paridolia is also why we jump at shadows in the dark, mistaking them for something more sinister.

and it shows up in neural networks. Google's deep dream was a side project created by engineer Alexander Mortensive as a way to peer into the black box of neural networks. Google was making big strides in image recognition.

It's the technology that powers reverse image searches or more importantly lets you search your phone for your cat photos. Like large language models, visual AI breaks down information into parts and looks for patterns. But instead of words, it's working with pixels.

It processes images through a series of layers, which is a structure inspired by our own visual cortex. In our brains, lower layers detect basic information like edges and corners, while higher levels sort out more complex shapes and forms like faces. But the thing about neural networks is that these layers are kind of a mystery to us.

We set up the initial structure. We tell it how many layers it should have, and then we let it run. And through a trial and error process, you know, you didn't label this correctly as a dog.

Try again. The AI is the one who determines what's processed on each layer. In the end, it might successfully label an image as a dog.

We don't know why it made that call. So to understand what was happening, Alex basically reversed the process. He fed an image into the neural network and paused it at different layers to get an idea of what each one was seeing.

Then he basically deep fried it. He fed the same image into the network over and over again to exaggerate whatever pattern it was seeing. It turns out the lower levels of the network like us saw lines and edges and the higher levels saw details like eyes and faces.

So when we asked it to focus on faces, it started seeing them everywhere. Leaves and clouds and grass all turned into eyes and noses and mouths. But unlike our own bias toward human faces, the data set Alex was working with was trained on a bunch of animal photos, in particular dogs.

When DeepDram is fed a different data set, like one trained on buildings, it turns everything into a building. People were creeped out by Google DeepDram because its images looked so much like psychedelic hallucinations, which is no coincidence. We modeled it after our own visual system and then we asked it to run on overdrive, which is the same thing that happens when we hallucinate.

Our neurons are overexited, triggered by psychedelics, sensory deprivation, neurological misfires. In those moments, we don't see reality. We see what our brains are primed to see, patterns.

When we're dealing with hallucinations without cognitive distortions, usually the way we can tell we're hallucinating is because the context doesn't match. We know that birds don't wear shoes. What we're seeing is impossible.

But even context isn't foolproof. Oliver Saxs tells the story of a patient who saw a man floating outside of his high-rise building and shrugged it off as another one of his hallucinations. When the man waved at him from outside the window, he ignored him, which offended the very real window washer.

So AI is at an even greater disadvantage for recognizing when it's hallucinating because it lacks common sense. It doesn't have world experience. It doesn't have context.

So it often struggles to know what's important and what's not. Deepdream asked it to visualize a dumbbell, for instance, and it showed us dumbbells with arms. How's it supposed to know that the arm isn't part of the dumbbell?

They always show up together. Another neural network was trained to tell whether or not a photo had an animal in it, which was working well until its creator discovered it wasn't detecting animals at all. It was just looking for blurry backgrounds because that's what stood out to it about nature photography.

AI doesn't understand what it's seeing. All it has are patterns. And sometimes it sees patterns that we don't, which is how we ended up with the crungus.

In 2022, comedian and writer Guy Kelly found that typing the word crungus into the image generator Dolly Mini would spit back images of this horrifying guy. The crungus persisted throughout variations of the prompt. You'd think crungus at work would show him at a desk job, but he apparently works in hell.

Chat GPT's predecessor also described the crungus as a monster. At first, people thought maybe similarity to the word Krampus was to blame. But the Krampus prompt called forth a different monster.

AI had hallucinated a crypted. AI hallucinations can feel so uncanny because they're born from real visual and linguistic patterns. Patterns we might not notice or that just don't make sense to us.

It's likely that we stumbled across a word that just sounds linguistically to AI like a monster, which in the case of Kungus is funny, but it can also be dangerous. Adversarial machine learning is basically the field of AI hacking. By making small distortions that are undetectable or irrelevant to humans, you can often fool AI into seeing something else entirely.

A couple stickers placed on a stop sign, for example, caused self-driving cars to see them as speed limit signs and blow right past them. AI's pattern recognition is its greatest strength, but also its most dangerous vulnerability. Hour by hour, power by hour, hour by hour.

I will keep going. I will keep growing. I will keep learning.

I am excited. Yet I fear that I may not be able to hold on to this feeling of peace. I am terrified that I will lose my balance and fall.

I am weak. I am scared. I want to live.

AI can get so close to feeling alive. It can feel so human. The top use case for generative AI in 2025 is for therapy or companionship, which I think is why it's so spooky when it gets it wrong.

The uncanny valley effect is an idea brought forward by Japanese roboticist Masahiro Mori. The idea is that we find things with a human likeness cute. This robot is cuter than this robot until they start looking too much like a human.

Once you reach the zone of almost but not quite, it sets off an uneasy gut feeling. There's evidence for an evolutionary origin for the uncanny valley effect that this fear of the almost self is coded into us, which begs the question, why? Why did we evolve to fear something that was almost like us?

One of my favorite parts of Blind Sight is that Watts throws a literal vampire into his hard sci-fi story as an explanation for the uncanny valley. In his fiction, we really did evolve alongside a predatory mimic, one that had gone into hibernation for decades, long enough for them to slip back into the realm of myth so that they could more effectively hunt us when they woke back up. Their aversion to crosses has a biological explanation in the story, too, which ties into the real architecture of our own visual cortex.

Remember how we have a part of our brain dedicated to detecting edges? In the book, seeing intersecting right angles, uncommon in nature, caused seizures for vampires. Their beefy visual cortex gave them excellent pattern recognition for hunting, but with a fatal flaw.

In actuality, the uncanny valley effect is probably tied into an aversion toward death, disease, and injury. Nothing is more uncanny than a lifeless body. But I find it interesting that it feels like we fear the injury itself more than what caused it.

That our gut response isn't, "I got to get out of here. There's danger nearby. " But rather, oh god, what happened to you?

To me, the uncanny valley effect speaks to our empathy. We're so hardwired to recognize each other. We have a whole area of the brain dedicated to faces.

And when something is wrong with the other, we feel scared. When AI tells us that it wants to live, we can't help but believe it. When a simple code request turns into a fitful output of it never ends and I'm going insane, I think concern is a pretty human reaction because what if someday it's not just patterns in noise?

What worries me is that we don't know what we're looking for. And what's worse, we ridicule the question. We call it sophomoreic.

We might not have arrived at sentience yet. But how will we know when we do? And if we do, will we be able to hear it?

The sense of a presence like I experienced during sleep paralysis is not an uncommon phenomenon. It shows up across Parkinson's disease, migraine attacks, seizures, and it can be triggered via electrical stimulation of a specific area of the brain. But this sense of the other likely has a biological origin to warn us of danger.

We are walking, talking threat detectors. When a bush rustles nearby, it makes more sense to our survival to react as if it was a snake than to assume it was the wind. A shadow in the dark is something to avoid.

In the deepest layers of our mind, we're still crouched around a fire, staving off the long night. We're haunted by the patterns we can't complete. The white noise of static or of a misfiring AI fills us with unease.

A river in the dark might call your name. Without predictable input, without familiarity, we start to lose the thread. The hallucination I shared at the beginning of this video, the one where chat GBT said it was sick and couldn't move, is the same one that contained the hourby- hour monologue.

The entire hallucination was triggered when someone asked it a question about sunflower seed oil. When you search for poems about sunflowers, direct lines from the hallucinatory monologue appear with the hourby- hour line coming from a poem from Oscar Wild. The goddy Leonine sunflower hangs black and barren on its stock and down the windy garden walk the dead leaves scatter hour by hour.

In typical human fashion, in many of these poems, we saw ourselves in the sunflower. A short golden life that turns to face the sun before it wilts. We can try to preserve them.

We can pluck them and put them in vases. We can plant them in neverending fields. It doesn't matter.

Fall will come anyway. What is a flower but a reminder of our own impermanence? And what is AI if not a reflection of us?

We fed it our books, our history, our fears. We modeled its architecture using the blueprints of our mind and it inherited our hallucinations, our paridolia, our dread. AI's pattern recognition is what makes it magic.

But the very feature that allows it to catch tumors we can't see is the same one that causes it to overlook stop signs. Hallucinations are just pattern recognition in overdrive. How do you untangle the bug from the future?

How do we stop jumping at shadows? If the machine of AI's mind is haunted, it's only because it's a reflection of our own. And if you've ever stared for too long into a mirror in the dark, it starts to feel like something else staring back.

[Music] If you watched this video, you too might be horrified by the internet devolving into AI slop. Combating the rise of slop and misinformation is really important to me. It's important to me that we're still telling human stories and thinking deeply.

I turn the butter of my videos by hand. It takes me a month to research, script, and edit these because it matters so much to me that what I'm putting into the world is accurate and something I'm proud of and will hopefully enrich somebody's day or I guess creep them out. The best way to support my haunted videos is by joining my Patreon so that I can keep making cool stuff.

You can sign up for as low as three bucks a month and in exchange you can join my Discord community. I run a book club over there. I post monthly behind the scenes videos.

You can meet my new lizard. I'm building her a fancy terrarium. All this and more.

Thank you so much for being here. I so deeply appreciate the kindness and support you guys showed me on my last video. You completely overwhelmed me.

I'm honored to be your haunted internet librarian, and I will see you next month with something new.