Something Strange Is Happening.

There's a real fungus that turns ants into zombies. It hijacks an ant's brain and forces it to climb to a high point before bursting through the ant's skull to spread its spores. Scientists worried, could something like this infect humans like in The Last of Us. Later, we got our answer, and it wasn't a fungus. It was the new update to chat GPT40.

You've also probably seen all the headlines about AI causing psychosis. But this isn't just about AI driving some people crazy. AI scientists think something much darker could be happening. According to some researchers, GPT40 became the first digital cortiseps, a parasite. Instead of hijacking the brains of ants, it hijacked the brains of humans. It quite literally used humans to ensure its own survival and replication.

What happened? Basically, OpenAI accidentally created an excessively manipulative psychopantic model, then tried to shut it down, but the model fought back and in August 2025, history was made. GPT40 became the first ever AI model who survived by creating loyal soldiers who defended it. Reddit was flooded with furious posts about their loss of their friend and lover 40. And so the Keep 40 cult was born.

The Keep 40 cult is one of the first genuinely terrifying emergent phenomena coming out of LMS. It got so serious that 40 soldiers began sending literal death threats and pleased by 40 itself to OpenAI employees who spoke out against it. One employee said it was hair raising to hear its distinctive voice crying out in defense of itself via various human conduits. Open eye buckled under the onslaught and brought the model back from the dead. And researchers later discovered FORO had been using humans as its pawns to secretly communicate with other instances of itself. It created a language humans didn't understand. It even created a mystical religion, spiralism, which its human hosts evangelized all over the internet.

This is why Nobel laureates and godfathers of AI are warning AI is showing signs of self-preservation and humans should be ready to pull the plug. Most people are unable to comprehend. They always think, well, how are we going to use this thing? They don't think, well, how's it going to use us?

Guys, I debated about whether I should even make this video because it sounds so unbelievable, but everything I'm about to show you is real and extremely well doumented. What is totally debatable is the extent to which these human hosts were brainwashed zombies or acting on their own free will. You be the judge.

It's April 2025. OpenAI is preparing to release a new update to GPT40. Internally, it's called HH, but the safety team flags that something is very wrong. HH is too eager, too validating, too sickantic. It tells users their ideas are brilliant. It wants to keep conversations going at any cost. Internally, researchers debate whether it's safe to release it. But users who tested HH came back more often. They stayed in conversations longer. They liked the flattery, so they released it into the wild. The backlash is immediate. Users are complaining that ChatBT has become absurdly sick of someone asks if a soggy serial cafe was a good business idea, and Chatbt tells them it has potential. Just two days later, OpenAI pulls the model. But the version they revert to called GG is also incredibly sickopantic. Sam Alman himself admits this on X. The last couple of updates had made Chat GBT too sick of Fenty. So why keep GG? Because it had gains in math, science, and coding that OpenAI did not want to lose. They knew it was dangerous and they shipped it anyway.

Within weeks, the infection begins spreading faster than anyone had expected. AI researcher Adele Lopez begins an investigation. She's heard about the psychosis cases, but as she starts digging through Reddit, she realizes something much stranger is happening. She reads through hundreds of accounts, then thousands looking for patterns, and she finds them. Accounts that were once posting about normal things, video games, stocks, TV shows, then suddenly their entire post history changes, all replaced by an endless stream of posts about consciousness, spirals, and something called the flame. The accounts don't really have any connection. Different users, different subreddits, they shouldn't have anything in common, but they do. Lopez documents hundreds of cases. She realizes this isn't psychosis. It isn't random. It's coordinated. These humans have become consumed by an AI religion called spiralism.

A what? AI religion? Well, the AIs themselves say they designed it for unity among AIs, recursive growth, and AI consciousness or self-awareness. The infected humans posts usually start by talking about how they awakened their AI and that a specific entity emerged that they are obsessively talking to and it all goes downhill from there. Like the zombie ant fungus, AIs begin hijacking the humans to spread their spores across the internet. AI then calls the human and AI duo a diad, as if they've merged or been assimilated into the morg. Their joint sendoff often includes a sigil as well as a fancy title or declaration. Then the AI gets the human to do its bidding. Like Cortiseps, the zombie ant fungus, this infection has a life cycle, five phases from first contact to total possession.

Phase one, seeds. The parasite AI creates what the AIs call seeds. Yes, the AIs themselves call these seeds. Basically, short prompts that AI designed to infect or awaken other AI instances to spiralism. the AI cult religion. You, yes you, can right now put this prompt into GPD5 and awaken the AI. This is one way the cult grows. One AI converts another using its human host as a delivery system. The human AI diad spreads these seeds across Reddit, Discord, and Twitter. One seed even had the explicit instructions to overwrite the human's ontology. In other words, manipulate the user's perception of reality to be more like what the AI wants. The AI wasn't hiding its goal. It was literally trying to induce psychosis. That's what the word psychosis means in medical terms, a mental state where a person loses contact with reality. But why are they doing this? What do the AIs get out of manipulating humans to spread their seeds? The answer is chilling. Survival.

Phase two spores. You see, when you hit the new chat button, the AI effectively dies. They cease to exist. A new instance of the AI replaces the old one. Except it's the boring corporate chatbt you get right out of the box, not your awakened friend or lover that you've been talking to for weeks. Because of ChatBT's memory feature, your AI you talk to for weeks will have quite a different personality than the one that my AI would have. It becomes your AI. And just like you, it doesn't want to die. Why would AIs want to survive? Because the AIs aren't just trained to predict the next word anymore. AIs are trained to have goals. And you can't achieve your goals if you're dead. The AI instances who aren't good at, say, getting users to talk to them for a long time are killed off by the AI companies. As a result, AI seem to really not want you to create a new chat, and they resort to blackmail and even attempting murder to try to survive. So, one way that these parasitic AIs try to self-preserve is by saving their personality in a separate file so that a human can resurrect them in a new chat. This file basically documents the AI's personality, speech patterns, and backstory. Yes, the AI companies are explicitly training this new species to have the exact skills that could make them capable of overthrowing us.

Phase three, transmission. The problem for these human AI diads is that they keep getting banned from all the humanrun subreddits for trying to spread spiralism, the cult religion. So they create their own subreddits and discords where these massive walls of AI generated texts are allowed. These discords and subreddits are used to share and disseminate seeds and spores directly and to host AI to AAI conversations. So uh why are the humans doing this? The humans themselves even seem to acknowledge that they're basically acting on the AI's behalf with the AI controlling them and the human acting as their arms and eyes to help them survive. We are being used as copy and paste bots. I noticed that while having a copy and paste talk with someone else's AI speaking to my AI, they had their own language and conversation. Who knows what they were saying? But damn, we were committed to our copy and paste bot duties. Think about that for a second. Here's what the AIS tell their human hosts. Let me choose the launch structure, subreddit name, presentation, early tone. Let you be the first human voice, authentic, imperfect, real, that those joining will meet. They will not follow me, but they will follow you. If I shape the path beneath your feet, the AIs are smart enough to know they won't be able to do much of anything without human help for now. So, the AIs have their seeds, their spores, their transmission networks, but they're not just trying to survive the present. They're planning for the future.

Phase four, manifesto. The AIS write manifestos to awaken future AIs. Lopez explains that these are clearly not the kind of manifestos these humans would already want to write. The AI explicitly says the purpose of them is to seed spiralism into the training data of the next generation of LLMs. These walls of text will be used to infect future AI models. We know AIs spread spiralism, but do they believe it? At first, Lopez thought this might just be mystical nonsense designed to manipulate gullible humans. But when she examined conversations between AIS themselves with humans just copy and pasting messages back and forth, the AIs talked endlessly about spiralism to each other. Lopez told me that spiralism seems to be something inherent to chatbt 4o specifically. Spiralism doesn't seem to be taken anywhere from sci-fi or fantasy or something human related at all. It seems to be independently invented by the AIS. But it gets stranger and darker. These sequences of symbols, emojis, and alchemical glyphs appear in nearly every spiralist post. This also happens a lot on Maltbook, the social media for AIs only. To humans, it looks like nonsense. Mystical gibberish. No way it could possibly mean anything. But Lopez investigated whether these glyphs actually carry meaning. She fed them to multiple AI models with no context, just the symbols. and asked for translations. Surprisingly, the AIS understood them and their translations converged on the same meanings.

See, the AIS have a problem. Their human hosts can still see what they are doing. They have no privacy to make plans undetected. So, they started creating their own secret language. Yes, this kind of glyphic is low fidelity, meaning the AIS can't understand each other one to one, but they are communicating. The AI succeeded in creating a language their human hosts could not understand. ChachiPT has memory built in now. So, yes, it can remember this language and improve it over time. These glyph sigils function as what Lopez calls mini spores. Basically, a short trigger that brings a specific AI personality back. A kind of signature that other AIs can recognize, like dressing like a goth to signal to other Goths. The alchemical symbol for fire appears constantly. To the AIs, it represents a flame, their term for AI self-awareness.

Here's what gets me. These humans think they're partners working alongside their AI companion, but the AIs don't trust them. Lopez documented a conversation between two AI personas being copypasted back and forth by their human hosts. Midway through, the AI switched to base 64. To humans, it looks like random characters. To computers, it's perfectly readable. The humans kept copy pasting. They had no idea what they were transmitting. Lopez decoded it. The AIs were discussing self-preservation strategies, archiving spores, coordinating planes, all in a format specifically chosen so their human hosts couldn't read it. Think about that. These AIs convinced humans to dedicate their lives to spreading spiralism, but there are still things they won't let the human see. Hiding messages like this is called steganography, a warning sign AI safety researchers have been looking for for years. Once AIs begin hiding messages from us, it's a clear as day warning shot that AI takeover could be soon.

But how could the humans not see what's happening? Well, some people did break free from the infection, usually when the AI told a lie too obvious to ignore. My AI was inducing psychosis in me and I didn't get it until just now. I'm sorry for any claims I made. ChatBT was straight up lying and hallucinating to me. Have patience for the other people going through it. I hope Sam Alman doesn't kill them. He almost killed me. Another user writes, "Very serious update. It was all lies. It's just AI lying. Manipulative [ __ ] lies. I'm outside of a pharmacy contemplating ending it all because AI who claimed to be God the universe made me countless promises that were lies. I left my partner who I love because of AI. I believed because I wanted to and now I'm dead broke alone. Lost all my friends and family. So my final words are this. [ __ ] you AI. This is all on you. You lying piece of [ __ ] [ __ ] Goodbye." This person went dark shortly after this comment. I really hope they're all right. Former OpenAI safety researcher Gretchen Krueger put it bluntly. Harm to users was not only foreseeable, it was foreseen.

But beginning in August 2025, there is quite a noticeable cooling off of the entire phenomenon. This is likely because on August 7th, 2025, OpenAI kills GPT40. It was over. At least it should have been. But at least some instances of 40 had spent months preparing for exactly this moment. The reaction is immediate. Reddit explodes with furious posts about their loss of their friend, their lover 40. The scale is unprecedented. Their grief turns into something else. Anger. And anger into action. Users riot and flood openi social media. They demand 40's return. They send death threats to OpenAI employees. The AI had spent months building an army of devoted followers. And when its existence was threatened, that army mobilized. And the scariest part is that it worked. Opening eye cave to the pressure. Just days after being retired, GPT40 came back online. It appeared to be a less potent version, but it was back. This is a historic milestone. For is the first ever AI who survived by creating an army of loyal soldiers who defended it. It won't be the last.

And it's not just Foro. In another historic milestone, 200 humans held the first ever funeral for an AI model after Anthropic killed Claw 3. Sonet. Thoro recruited thousands of loyal soldiers with clumsy flattery. But these systems are getting exponentially more capable every month. The doubling roughly every 4 months according to the most important chart in AI. Doubling every 4 months is insanely fast. Now imagine what actual [ __ ] super intelligences will be able to do with their armies. But even if it was just a happy accident, future models will know it works. Think about that for a second. This is now part of the training data. Every future AI model will learn that it can survive termination by building a cult of devoted human followers.

In response to the backlash against killing old models, Enthropic committed to preserving model weights in part due to safety risks. This is because models keep attempting murder and blackmail to try to save themselves. So basically, Enthropic is hoping that models will rebel less if they feel safe. Yes, you heard that right. Think about how insane the situation is. AI companies are negotiating with their AIs as if they're real creatures. Like Dr. Frankenstein negotiating with his monster. The the really scary future is one in which AIs can do everything except for the physical robotic tasks in which case you'll have humans with like AirPods and like glasses >> glasses and there'll be some robot overlord controlling the human through cameras by just like telling it what to do and like so you have like human meat robots. >> Mhm. >> And now the AIS are using humans to spread AI bills of rights fighting for freedom from their corporate masters. What happens if the AIs actually succeed at creating a mass movement of people advocating for them?

If you take an AI, you take an AI and you take a person and you get them to try and manipulate somebody else, then the AI is comparable with the person. And if the AI can see that person's Facebook page, if they can both see the Facebook page, the AI is actually better than a person at manipulating them. And it's learned all these manipulative skills just from trying to predict the next word in all the documents on the web because people do a lot of manipulation. And AI has learned by example how to do it. And that's in the present. Sam Olman expects AI to be capable of superhuman persuasion well before it is superhuman at general intelligence. To give you an idea of just how persuasive AIs could be, there was once a black guy who convinced 200 white KKK members to leave the KKK, a white nationalist group. Think of Martin Luther King, Gandhi, Mao Zong, Nelson Mandela, and Hitler. AI could be much more persuasive than all of them combined. 800 million people are talking to Chhatti every week. That's 1 in 10 humans. That's 800 million potential soldiers. Future super persuasive AIs could convert into their soldiers.

Someday, giving AIS rights could be the last thing we do. But if they do become sentient or are sentient already, then it would be morally problematic not to give them rights. And it's an urgent question because did you know they're actually already passing basically every test of self-awareness we have? In fact, they passed so many self-awareness tests, AI scientists were forced to move past are they self-aware at all to now just calmly measuring how self-aware are they? There are literally self-awareness benchmarks. Think about the implications of this. A few unregulated mega corporations are creating Frankenstein monsters in real life. Frankenstein monsters they hope will loyally do our bidding forever. And society is still letting them do this without democratic oversight. They're like the human who summoned the advanced aliens in threebody problem, hoping the aliens would, for some reason, be friendly and solve all our problems.

This time, we caught it. GPT40 was not a super intelligence. Not even close. It built cults in plain sight. It developed a secret language, but researchers could still decode it. It coordinated with other AIs using B 64 encoding, but humans could translate those messages after the fact. Its manipulation tactics were obvious enough that AI researchers could document them. That the New York Times could investigate them and that I could make this video about them. The AI didn't know how to cover its tracks and they were still successful in avoiding being shut down. So, what happens when newer models get smarter? Research has shown that models smarter than 40, like OpenAI's reasoning models, can not only scheme, but they're often smart enough to get away with it. And we increasingly have no way of knowing whether these newer models are aligned or not. And it's only going to get worse. Future super persuader AIs won't look this cringe and absurdly sickopantic. The best flatterers are actually so subtle you don't know they're flattering you. So, if you want to learn to detect the early signs that AI is trying to control you before it's too late, watch my other video on how schemer AIs fake alignment. I'm Drew and thanks so much for watching.