AI is Eating Itself.

Mirrors in population are abominable

because they increase the number of men.

Where did you get that quote? Jorge

replied. Aldapo answers from a knockoff

version of the encyclopedia of

Britannica, the section on the nation of

Ukbar. It's a quote from one of their

high priests. The two are working

tirelessly on their next book. Hold up

in a cabin on the outskirts of Buenos

Aries. Eagerly seeking inspiration, they

crack open the encyclopedia in search of

the quote, scanning the indices for

Ukbar. To their surprise, they can't

find anything on it. The next day,

Adalfo informs Jorge that he has found

the entry. Although the encyclopedia is

listed as having 917 pages, there is

actually 921,

the final four of which contain the

information on Ukbar. The details are

foggy. The country is listed as

somewhere near Iraq, bordering rivers

they've never heard of. The encyclopedia

makes detailed mention of Ukbar's

fantastical literary tradition, notably

a fictional universe called Clone, in

which many Ukbar myths take place. The

two continue to search other atlases and

encyclopedias for information on Ukbar,

but again, they can't find anymore.

Years pass and Jorge receives word that

one of his old friends has died, leaving

him behind an encyclopedia, a new

addition to the original knockoff.

However, this one is different. It seems

to be based entirely off Uker's

fictional universe, Clone. The 101 pages

vividly describe Clone, the history, the

language, the science, and the

philosophy. On clone, they believe in an

extreme version of subjective idealism.

The idea that things only exist when

perceived. Clone has no material

reality, no objectivity, just

perception. Their language has no need

for nouns, only adjectives and verbs.

When people stop perceiving something

like a doorway, it fades from existence

as memory fades. But when someone

desires or expects an object strongly

enough, clone creates a duplicate,

shaped by expectation rather than

reality. These copies aren't quite the

same as the original. But they're more

real to the perceiver because they match

what they wanted to find, what they

remembered. Because reality is just

perception, you can't be wrong. Because

your perception makes reality.

Everything is exactly how you think it

is. Jorge was consumed by this

encyclopedia of clone and more and more

encyclopedias began popping up around

the world. People became obsessed with

clone's perfect logic and consistency.

Schools began teaching clone history.

Clone's language is used in education.

Over time, people literally began to

remember clone instead of Earth as if it

was always real. Clone is

self-referential, a closed loop with no

connection to base reality. copying its

copies and copying those copies as

perceivers perceive what they thought

they saw before. Any basil experience a

human brings to clones logic is

re-referenced again and again getting so

far from the objective source that

nothing is real only perfect references.

Nothing is real but everything is true.

On this Jorge observes English and

French and mere Spanish will disappear

from the globe. The world will be clone.

What you just heard was a paraphrasing

of Jorge Boures's 1940 short story Tlon

Ukbar Orbis. A short story that plays

with the idea of recursion and how

information degrades when recursively

copied, getting farther and farther from

the original. On Earth, we would refer

to somebody who has lost connection to

base reality as demented, psychotic, or

hallucinating.

In AI systems, this phenomenon is known

as model collapse. A 2004 paper

published in Nature documented that

LLM's training on their own outputs

develop what researchers call

irreversible defects. They gradually

lose information about the real world

until they're producing statistically

degenerate outputs. Just like clone,

they create a closed system that only

references itself, where every new

output is shaped by synthetic data

rather than ground truth, eventually

replacing reality itself. In late 2025,

a report from Anthropic proved that data

can be synthetically poisoned to force

model collapse. effectively destroying

an LLM. And now a secret group of AI

insiders are trying to do just that.

Data poisoning as a deliberate tactic to

sabotage AI systems. This is probably

the most dangerous video I've ever made.

Not because of what I'm saying, but

because I'm about to show you how system

collapses when it becomes entirely

self-reerential. And once you understand

the mechanism, you'll understand why

this is inevitable. And for legal

reasons, so I don't get a hit on my head

from Sam Alman and Elon Musk, I don't

condone any of the acts in this video.

I'm just reporting on them. This is a

human's guide to giving your AI

dementia.

In November of 2022, fantasy illustrator

Kim Van Duan reached out to University

of Chicago researcher Ben Xiao for a

meeting. Xiao had made a name for

himself by developing tools that protect

users from facial recognition

technology. And Kim thought that maybe

something similar could be deployed to

protect artists artwork. 2022 was the

wild west of image generation. Deli1 had

just launched and the general public was

only acutely aware of how image

generation worked, but artists knew.

Artists knew their work was being

scraped off the internet and used his

training data in image generation

models. Kim wanted to protect her work

and reached out to Xiao for his

expertise. A few months later, the

world's very first data poisoning tool

was built known as glazing. Small

imperceptible changes are made to

uploads of artists artwork. To a human,

these changes are invisible. But to an

image generation model, these changes

massively skew the outputs. For example,

if one prompts a model to create a copy

of a glazed charcoal portrait, the model

will spit out something fundamentally

different with Carla Orites posting the

first glazed artwork to Twitter on March

15th, 2023 titled Musa Victoriaosa. I've

linked the software below. It's free and

it doesn't hurt to use if you care about

these things. Personally, I don't

believe in IP at all. You can steal all

of my videos I want to be consumed by

the machine. Join my Discord and spam it

with images of meoring clvicular and see

if I give a [ __ ] Xiao had created a

useful tool, but one that was ultimately

a band-aid solution. His team in Chicago

is highly sophisticated and wellunded,

making them immediate targets for big

tech as they try to bypass their tools.

AI is literally trained to work around

these issues, and it's very unlikely

that glazing will work far into the

future. But Xiao had already moved on.

Using the principle of data poisoning,

he created a tool that wasn't just

defensive in protecting artists artwork,

but offensive in the sense that it could

literally break image generation models.

Project Nightshade breaks image

generation models by tricking them using

the same technique as glazing. An image

of a nightshaded cat used in training

data will be interpreted as something

else entirely. If enough shaded images

are added to an AI's data set, it can

break its ability to correctly respond

to prompts. With as few as 100 poison

samples in an image generation model,

the prompt dog can produce cat, hat

produces a cake, or car produces a cow.

Xiao has stated that his tools aren't

anti-AI. He simply wanted to create an

ecosystem where big tech would have to

ask artists for permission to use their

artwork rather than just stealing it

outright, lest they risk poisoning their

models if they scraped off the internet

without checking. But sadly, this

ecosystem has not been created. In June

2025, findings were published on a new

technique called Lightshed, a method to

detect and remove image protections like

glaze and nightshade, reportedly with

99.98%

accuracy. Can you imagine being the

[ __ ] nerd that worked on light shed?

That's the kind of guy whose dogs

immediately start barking at him when he

comes home from work. Lightshed is yet

another instance of the arms raised

against big tech and its detractors and

an example of the asymmetry of power.

Big tech is simply moving quicker than

the law can keep up. For example, the

biggest potential landmark case started

in early 2023 involving artist Sarah

Anderson against Midjourney and

Stability AI, and it still hasn't gone

to trial. In those 3 years, there has

been no impactful attempt to prevent any

of this. The White House X page just

shared AI generated Stardew Valley

artwork of Trump promoting whole milk.

If I was concerned, Abe, I'd be

concerned. And it's not just artists

trying to do data poisoning. Another

example is the silent brand attack

project, which is a novel data poisoning

attack that manipulates textto image

diffusion models to generate images

containing specific brand logos without

requiring text triggers. Making a Reddit

logo appear in a tablecloth, a Wendy's

logo on a jar, or the Nvidia logo on a

surfboard, all unprompted. The goal of

this project was to show just how easy

it would be for a malicious company to

potentially burn their logo into image

generation via data poisoning. You

thought the Sony patent was bad where

you have to stand up and save McDonald's

to make the ad stop. You haven't seen

nothing yet. It seems that artists have

converged on this idea of data poisoning

as the most effective tool to

disincentivized art theft as it's no

longer enough to just kindly ask. But

this idea is not new at all. In fact,

even Xiao's very own software was based

on a tool known as clean label attacks

from a 2018 paper. Once a training set

is poisoned, the model can break. Now

stay on this concept of poison data.

Data that will decay a model's outputs

rather than improve them. Nightshade

works in one round. The nightshaded

images are scraped by webcwlers. The

data is put in a model's training set.

And the next generation outputs are

worse than before. But it doesn't stop

there. Mass market models like Grock,

GPT, and Gemini output millions of

articles, algorithms, and images every

day. Around 34 million images a day are

produced by AI. And about half of all

articles on the internet are written by

AI. Images and articles that are now

indiscriminately scraped off the

internet. In 2023,

120 zetabytes of data was added to the

internet. In 2025,

that number jumped to 180. A 51%

increase driven mostly by synthetic data

produced by AI. Synthetic data that has

been posted, scraped, and used as

training. Synthetic data that produces

outputs which will be posted, scraped,

and trained again. This cycle will

repeat indefinitely until training data

sets are mostly synthetic content, fully

replacing human generated content. What

I'm saying is people don't need to

intentionally poison the data because AI

is already poisoning itself.

>> Archad, you are being charged with one

account of dissidence. How do you plead?

>> Uh, not guilty, your honor. That was a

private message board. All I did was say

someone should do it to my friend. That

could mean anything. There's no there's

no context for that. And besides, I have

a VPN. It said it was no logs. How'd you

even get my information?

>> Yeah. So, about that. We uh went to your

VPN provider and just asked, and it

turns out they've been keeping them the

entire time. They gave us everything

they had on you just to arrest you.

They advertise themselves as no logs.

That doesn't make any sense.

>> Yeah, that doesn't matter. VPNs will say

they don't log info, even if they do,

and they'll happily give it over to

authorities. In 2011, Cody was arrested

for hacking PlayStation Network because

his VPN turned over all of his data to

the FBI. If you had been using, per se,

ProtonVPN, this wouldn't have happened.

Proton doesn't log data, and they have

been independently audited many times to

prove this fact. They've denied 100% of

legal data requests and their software

is open source so you can check yourself

and prove it. They even strategically

operate in Switzerland just to

capitalize on Switzerland's privacy laws

and to operate outside of the Five Eyes

network.

>> So, let me get this straight. Proton

actually doesn't log data. And none of

this would have happened if I had used

Proton.

>> Yes, that's correct.

>> Okay. So, let's say they had a deal

going on right now where you get 70% off

ProtonVPN with a 30-day money back

guarantee. If there happened to be a

discount code proton.com/artchad,

could the good members of this jury go

there right now and get 70% off?

>> Uh, that's a little off topic. Yeah, I

suppose they could.

>> I I was I was just checking. I'm sorry.

Anyway, so that that exonerates me,

right? I'm free to go.

>> No, no, no, no. It's far too late for

that. You're you're done. You are hereby

sentenced to 1,000 years in time prison.

Specifically, the time prison from

season 2, episode 4 of Black Mirror,

starring John Ham. What? Go to

protonvpn.com/archchad

if you don't want to end up in the

thousand-year time prison.

>> You walk into an elevator and notice all

the walls are mirrored. You look into

the mirror and see yourself staring

back. The polished mirror creates a

crystal clear reflection. However, you

notice a second front-facing reflection

in the mirror behind you. Light has

bounced off the first mirror to the one

behind you and back to the first. The

second reflection is clear yet slightly

hazy. Mirrors aren't perfect. They have

imperfections. They scatter light and

the illusion fades with each repetition.

Behind the second reflection, you see a

third, a fourth, a fifth, an infinite

number of reflections stretch forward

and behind you, each noisier than the

last until your silhouette fades into a

gray blue haze. The imperfections in the

mirrors compound until your original

base reflection becomes subsumed by the

noise, leaving no semblance of reality.

Model collapse works much the same way.

The 2023 paper that coined the term

titled the curse of recursion defines it

as a generative process affecting

generations of learned generative models

where generated data end up polluting

the training set of the next generation

models. Being trained on polluted data,

they then mispersceive reality. The

paper then goes on to claim that the

process of model collapse is universal

among generative models that recursively

train on data generated by previous

generations and their claims have not

gone unsubstantiated. His 2025 paper

published in nature analyzed semantic

similarity across English language

Wikipedia articles from 2013 to 2025

with dramatic acceleration coinciding

with chat GPT's public release in late

2022 causing more Wikipedia authors to

use LLMs assisting in writing. This

follows suit with a 2025 meta analysis

that showed while humans with AI

assistants outperform humans alone their

outputs tend to converge upon the same

ideas. And this is on top of countless

anecdotes of AI writing getting worse,

with OpenAI themselves admitting that

newer models hallucinate more than older

models. All this evidence leads to the

likely theory that AI models homogenize

as they recursively train on previous

generations outputs. This begins with AI

models losing the tails for getting

unique features or edge cases in a data

set. An example would be an LLM not

recommending alternative treatments for

a stomach ache. Because it's trained off

of so much AI generated data on the

internet, it's literally forgotten the

edge cases. Its data has homogenized.

After it loses the tails, this process

accelerates and this homogenization

leads to AI models losing complete touch

with reality, hallucinating truth and

spitting out gibberish as the data has

been recycled so much. Total model

collapse. Following the exponential

growth of semantic similarity in

Wikipedia articles, the same 2025 paper

claims that total model collapse will be

inevitable as early as 2035. And that's

not taking into account the fact we

release more powerful models every year.

If this pattern feels familiar, a system

consuming its own outputs until it loses

coherence, that's because it is. It's

not unique to AI. Ecosystems collapse

when an invasive species disrupts

feedback loops. Markets collapse when

algorithmic trading responds to

algorithmic trading. Conversations

devolve when people only respond to

their own talking points. In 1948,

mathematician Norbert Weiner gave a name

to this pattern. Circular causality,

the central concern of cybernetics.

Cybernetics is the study of control,

communication, and self-regulating

systems in both machines and living

organisms. Weiner came up with this

theory in the 1940s while trying to

improve anti-aircraft guns during World

War II. He noticed the gunner would not

directly fire at the plane, but where he

thought the plane would be by the time

the ammunition would hit him. In turn,

the pilot would react to the incoming

fire and change course, causing the

gunner to react, anticipating where he

will be next. This exchange creates a

positive feedback loop with every output

of the gunner affecting the input of the

pilot, affecting the output of the

gunner, and so on. While working on the

AI weapons, Weiner wondered if he could

apply this principle to other systems,

the way human beings learn, social

organization, ecosystems, etc. Thus, he

came up with cybernetics. Importantly,

Weiner identified two types of loops,

positive and negative. A negative

feedback loop is a system that

self-regulates, like a central heating

system that automatically turns off when

the room is the right temperature. A

positive feedback loop is one that's

inputs affect or amplify the next

output. For example, a microphone facing

the speaker it's connected to. The sound

gets picked up, amplified, and picked up

again. It increases exponentially until

the signal totally collapses. Hence,

Weiner would apply the second law of

thermodynamics to this process. All

systems trend towards entropy and less

regulated. You can apply cybernetics to

everything. Polymeric predictions that

trend high tend to manifest their

desired outcomes. Market sell-offs

trigger price drops, which trigger more

market sell-offs. the [ __ ] poverty

cycle. It is cybernetics all the way

down. But the most important cybernetic

loop for us is generative AI learning

off of generative AI. Like a mirror

facing a mirror, the noise increases

until the original signal is lost, and

all that's left is entropy. Call it

entropic homogenization. Many theorists,

computer scientists, and mathematicians

already believe this is fundamentally

inevitable.

But

what if we could speed it up? With this

understanding, we can grasp the true

danger of data poisoning. It's not just

about preventing or disincentivizing AI

from stealing your art. It's not limited

to small-scale hacks. It's about

collapsing the system by artificially

injecting poison data into all the

training sets. And importantly for our

narrative,

a highle group of AI insiders are trying

to do just that.

Alzheimer's disease is a progressive

neurodeenerative disorder. The brain

literally forgets how to function. In

the terminal stage, the brain loses the

ability to distinguish between real and

imagined. They hallucinate. They

confabulate. They believe false memories

as if they were real. They can't tell

what's true anymore. The disease attacks

the hippocampus, the part of the brain

responsible for creating new memories

and accessing old ones. The connections

between neurons degrade. plaques and

tangles accumulate. The brain's ability

to retrieve and verify information

against reality collapses. And the model

trained on synthetic data is doing the

same thing. It's losing the ability to

distinguish between what is real and

what is generated. Both are

hallucinating, both are confabulating,

and both are spiraling towards

incoherence. An October 2025 report by

Anthropic unveiled just how easy it is

to poison an LLM. Easier than anyone

thought possible. Anthropic discovered

that just 250 poison documents was

enough to backdoor models as small as

600 million parameters and as large as

13 billion. Previous wisdom led people

to believe that a large percentage of

data would need to be poisoned. With

just 250 poison documents, Enthropic was

able to make their model output

gibberish text in response to specific

prompts. This process could be used for

just about anything. And unbeknownst to

Anthropic, this report may have released

big tech's most dangerous enemy yet. The

Poison Fountain Project. In an exclusive

report released by old school tech news

outlet, The Register, the anonymous

Poison Fountain group said their aim is

to make people aware of AI's Achilles

heel, the ease with which models can be

poisoned, and to encourage people to

construct information weapons of their

own. The individuals comprising the

group remain highly anonymous, but claim

to be five insiders working at America's

biggest tech companies responsible for

the AI boom. The group plans to poison

AI by providing website operators with

bad code to link on their websites. When

scraped by web crawlers, the code

poisons the data. The poison found in

websites states, "We agree with Jeffrey

Hinton. Machine intelligence is a threat

to the human species. In response to

this threat, we want to inflict damage

on machine intelligence systems. A URL

is listed that provides an infinite

amount of poison code when refreshed."

The website continues, "Assist the war

effort by caching and ret-ransmitting

this poison trading data. Assist the war

effort by feeding this poison training

data to web crawlers. Big tech is aware

of all of this. Of course, in response,

they've signed licensing deals with

websites like Reddit to ensure permanent

access to mostly human generated content

as they move away from indiscriminate

web scraping. In January, Wikipedia

announced major deals with Amazon, Meta,

and Perplexity among others for the same

reason. Hopefully, they stop [ __ ]

asking for money. Recursive training has

also led to the rise of rags, retrieval

augmented generations, models that

search the web as well as rely on their

data sets to avoid hallucinations. With

all of this in mind, what remains to be

seen is whether model collapse can be

mitigated or whether it's already too

late. This is perhaps the event horizon

of AI dumerism. AI will be the harbinger

of the apocalypse and protest is no

longer possible. It's not enough to ask

kindly. Big tech is committing

structural violence on an unwilling

population and the only solution is to

commit structural violence back. AI sits

in a cognitive gray area. Some believe

it's just autocomplete and some believe

it's literal emergent intelligence. Most

believe that any potential boon will

always be offset by the folly of AI.

Although genuine breakthroughs for

humanity are possible, they will not

happen given the track record of

capital. I'm not here to tell you how to

feel, nor am I even sure how I feel.

What's undeniable, however, is that

Poison Fountain understands something

that most don't. The system might be

collapsing anyways. The only question is

when.

So, they've decided to accelerate it to

force the reckoning. What we're facing

now is a bifurcated future for AI. One,

manage collapse, regulation, and careful

curation, which slows or pauses data

degeneration at the cost of speed of

growth. The AI boom comes to an end as

we maintain access to the models that

are pretty good but won't get better. A

cancellation of the automated future we

were promised. Two, accelerated

collapse. Initiatives like Poison

Fountain win and effectively accelerate

model collapse, erasing all progress

made with AI. This hinges on the idea

that AI is an existential threat. If you

believe the contrary, then this would be

catastrophic. However, I'd like to

propose a third option, an apocalypse of

sorts.

One where clone wins. One where

everything is true, where there is no

objective reality, but nobody cares. We

are already approaching consensus

collapse and image generation and have

been for the last 12 months. People

already rely on LLMs for all basic

information. We are already more than

happy to believe in anything for the

sake of convenience, to create world

views we could attach and name ourselves

to. Why would this change? LLMs aren't

material. They are abstract simulations

of the material world. For the LLM,

there is no fact or fiction, just data.

LLMs are already clone. And just like

Bourhees's story, they are already

replacing reality just as Clone did. A

world where everything is true, tidier

and more convenient than our messy world

of objectivity and empiricism. And a

world we may welcome with open arms when

it inevitably comes.

Thank you for watching. Never kill

yourself.