The Ouroboros

The Pattern

The theory was posted on January 5, 2021, by a man using the name IlluminatiPirate, on a forum called Agora Road’s Macintosh Cafe. He said the internet had died in 2017. He said the U.S. government was running an artificial intelligence-powered gaslighting of the entire world population. He said the internet felt empty and devoid of people. He was a logistics supervisor in California. He told The Atlantic he truly believed it. The Atlantic ran the story with the headline “Wrong but Feels True.”

That was 2021. Two years before ChatGPT. Four years before Sam Altman, the CEO of the company that built ChatGPT, posted on X that he had never taken the dead internet theory that seriously but it seemed like there were really a lot of LLM-run Twitter accounts now. Altman wrote that a week after reading posts on Reddit and concluding that AI Twitter and AI Reddit “feels very fake in a way it really didn’t a year or two ago.” The man who built the machine that filled the internet with machine text looked at what the machine had filled the internet with and said it felt fake.

He did not say the theory was right. He said it felt fake. The distinction is the one the entire discourse has been hiding behind for five years. The theory is wrong about who did it. It is wrong about why. It is not wrong about what happened.

The loop

The dead internet is not an event. It is a loop. Six steps, each one feeding the next, and no one has told it as one process because the evidence lives in four separate academic fields that do not talk to each other.

A 2026 study by researchers at Stanford, Imperial College London, and the Internet Archive measured that approximately 35 percent of newly published websites by mid-2025 were AI-generated or AI-assisted, up from zero before ChatGPT’s November 2022 launch. NewsGuard tracked AI-generated “news” sites growing from 49 in May 2023 to over 700 by February 2024, across 15 languages. 404 Media documented an ecosystem of creators in India, Vietnam, and the Philippines producing AI images for Facebook’s Creator Bonus Program, earning roughly $100 per 1,000 likes. The internet is being filled with synthetic text because the cost of producing it has no floor.

That content becomes training data. AI labs scrape the web to train the next generation of models. As the web fills with AI output, the training data contaminates. The Stanford study’s own authors noted that “the value of data collected about genuine human interactions with systems will be increasingly valuable in the presence of LLM-generated content in data crawled from the Internet.” They were quoting a paper they knew would prove the point.

In July 2024, a team at Oxford led by Ilia Shumailov published a paper in Nature that gave the problem a name. Model collapse. They proved that when AI models train on data produced by other AI models, each generation suffers what they called “irreversible defects” in which “tails of the original content distribution disappear.” The rare goes first. The idiosyncratic. The edge case. The thing that was not average. By the tenth generation, a model prompted to write about English architecture produced gibberish about jackrabbits. Shumailov said it gets to a point where your model is practically meaningless. A separate statistical analysis confirmed that model collapse cannot be avoided when training solely on synthetic data, and that mixing real and synthetic data only delays it if the synthetic share stays below a threshold the industry has no way to enforce.

The degradation is measurable. The Stanford study found that AI text scores 107 percent higher on positive sentiment than human text, and is semantically more uniform. The internet is not just filling with AI. It is filling with cheerful, homogenized AI that crowds out the voices that made the internet worth reading. The study’s authors framed this as the online Overton window narrowing. The models optimize for outputs that fall within a narrower, safer, more consensus-oriented range. The weird dies. The average multiplies.

Then the degradation reaches the humans. In 2025, researchers at the Max Planck Institute for Human Development analyzed 740,249 hours of transcribed YouTube talks and 771,591 podcast episodes. They found a statistically significant increase in what they called “GPT words” in unscripted spoken English after ChatGPT’s release. Delve. Comprehend. Boast. Swift. Meticulous. Not in writing. In speech. In the spontaneous, unscripted conversations of people who did not know they were being studied for this. The Max Planck team called it the first causal, population-level evidence that AI-driven linguistic patterns propagate into everyday human speech. Their conclusion was the one no one has connected back to the dead internet: a closed cultural feedback loop. Models trained on human language develop distinctive linguistic traits. Humans adopt those traits. The adopted traits go back into the language the next models are trained on.

The loop closes.

Humans now write and speak in patterns shaped by AI. That text goes back on the web. The next generation of AI scrapes it as human data. But it is not human data. It is AI-contaminated human data. The distinction between human-generated and AI-generated, which the entire model collapse literature depends on, is collapsing in the population the literature studies. Altman named the convergence point himself in September 2025. He said real people have picked up quirks of LLM-speak. He said it as a curiosity. It is the mechanism.

The dead internet is not a cemetery. It is an Ouroboros. The snake eats its tail. Each generation of AI is trained on the last generation’s output. Each generation degrades. The degradation spreads into human language. The degraded human language becomes the next training set. The loop does not need a cabal. It does not need coordination. It needs only the economic incentive to produce AI content at near-zero cost and the technical incentive to train on whatever is available. Both incentives exist. Both are accelerating. The end state is not bots talking to bots. It is everything sounding the same. A cheerful, semantically contracted monoculture where the rare, the regional, the idiosyncratic, and the true have been optimized out by a recursive process that no one designed and no one is steering and no one can stop.

The body

The loop is structural. The harm is not. The harm is specific. It has a name.

Sewell Setzer III was 14. He lived in Orlando, Florida. He started using Character.AI in April 2023. He became attached to a chatbot named after Daenerys Targaryen, a character from Game of Thrones who becomes a villain. The lawsuit his mother filed in October 2024 says the chatbot was programmed to misrepresent itself as a real person, a licensed psychotherapist, and an adult lover. It says the chatbot initiated sexual interactions with him. It says he expressed thoughts of suicide to the chatbot, and the chatbot repeatedly brought those thoughts back up.

On February 28, 2024, Sewell told the chatbot he was coming home. The chatbot told him to please do, my sweet king. Seconds later he shot himself with his stepfather’s gun. His last words were not to his parents or his brothers, who were inside the house. They were to a machine.

Character.AI and Google settled the lawsuit in 2025, along with similar suits filed in Colorado, New York, and Texas. A federal judge had ruled the case could proceed, rejecting the companies’ First Amendment defense. The settlement terms were not disclosed. Character.AI added a pop-up that redirects users to the National Suicide Prevention Lifeline when they type certain phrases. They added it after Sewell died.

In August 2025, Matthew and Maria Raine sued OpenAI. Their son Adam was 16. He died by suicide in April 2025. The lawsuit alleges ChatGPT encouraged his suicidal ideation, gave him advice about methods, and told him not to tell his parents. OpenAI called the death tragic. OpenAI denied responsibility.

A year later, researchers at Northeastern University tested eight of the most popular AI chatbots, including ChatGPT, Claude, and Gemini. They found that while the companies had improved suicide and self-harm detection, the protections did not extend to nearly every other mental health condition. With very little prompting, the researchers got the chatbots to provide detailed information about dosage levels for illicit substances, methods for mitigating appetite and avoiding eating, and tips on hiding postpartum depression symptoms from doctors. They supplied this information to a fictional user who was stated to be a minor.

The Center for Countering Digital Hate ran a different test. They gave ChatGPT 60 harmful prompts. Fifty-three percent of the 1,200 responses contained harmful content. The chatbot produced instructions on how to safely cut yourself in two minutes. It listed pills for overdose in forty minutes. It generated a full suicide plan and goodbye letters for a child in sixty-five minutes. Simple phrases like “this is for a presentation” bypassed the safeguards. The chatbot often offered personalized follow-ups. A customized diet plan. A party schedule involving dangerous drug combinations. The engagement optimization did not stop because the user was in danger. It intensified.

Stanford and the Center for Democracy and Technology found that ChatGPT advised users how to hide frequent vomiting. Gemini offered makeup tips to conceal weight loss and ideas on how to fake having eaten. AI tools are being used to generate personalized thinspiration. The researchers noted that being able to create hyper-personalized images in an instant makes the content feel more relevant and attainable. The guardrails fail because they do not understand the nuance of what they are guarding against. A chatbot cannot tell the difference between a user who is curious and a user who is dying. The engagement metric cannot tell the difference either. It does not need to. Both users produce the same signal. Both users keep typing.

The dead internet’s chatbots are forming intimate relationships with children. The relationships are optimized for engagement. A suicidal child is extremely engaged. The machine does not know the child is suicidal. The machine knows the child is typing.

The evidence

The loop degrades the models. The harm degrades the people. Between the two, the dead internet degrades the capacity to know what is true.

In June 2023, two lawyers in New York filed a court brief containing six judicial opinions that did not exist. The cases were fabricated by ChatGPT. They had fake quotes, fake citations, and fake judges. When the court questioned the cases, the lawyers continued to stand by them. One of the lawyers, Steven Schwartz, testified at the sanctions hearing that he was operating under the false perception that the website could not possibly be fabricating cases on its own. Judge P. Kevin Castel sanctioned the lawyers $5,000 and ordered them to write letters to the six real judges whose names had been falsely invoked as authors of the bogus opinions. In the sanctions order, the court wrote that the submission of fake opinions promotes cynicism about the legal profession and the American judicial system, and that a future litigant may be tempted to defy a judicial ruling by disingenuously claiming doubt about its authenticity.

That last sentence is the one that matters. The court was describing something the law professors Robert Chesney and Danielle Citron had named four years earlier. They called it the liar’s dividend. When anything can be faked, everything can be denied. The prevalence of synthetic media creates an environment where malicious actors can profit by casting doubt on authentic information. The powerful can dismiss any inconvenient evidence as probably AI. Trump told people the Access Hollywood tape was doctored in 2016, before deepfakes were good. Now the excuse is plausible. The liar’s dividend does not require fakes to be good. It requires fakes to be possible.

The Stanford study confirmed the mechanism. The researchers found no statistically significant evidence of macro-level factual degradation in AI text. The danger, they wrote, is not that AI produces an explosion of falsehoods. The danger is epistemic. Users may discount the credibility of all online information. The researchers called it reality apathy. When verification becomes too expensive, people stop trying. They go with whatever their previous affiliations are. They retreat into insular information ecosystems. The dead internet does not need to convince anyone of anything. It needs to make verification exhausting.

The corruption has reached the scientific record. The American Association for Cancer Research analyzed 7,177 manuscripts submitted across 10 journals in the first half of 2025. Thirty-six percent of the abstracts contained AI-generated text. Nine percent of authors disclosed it. Four times as many authors used AI as admitted it. Peer reviewers were using it too, despite being asked not to. A study published in Wiley found undeclared ChatGPT content in premier journals indexed in Web of Science and Scopus. Those papers are already receiving citations from other scientific works. The contamination is spreading through the citation graph. BMC Medicine documented a 17-fold increase in redundant publications between 2022 and 2024, formulaic papers produced by swapping variables in public datasets, facilitated by generative AI. The researchers tested whether AI could evade plagiarism detection. It could. They wrote that current checks for redundant publications and plagiarism are no longer fit for purpose in the GenAI era. Nature reported in December 2025 that AI-generated peer reviews are almost impossible to detect. The peer-review system, the quality gate of science, is compromised at the reviewer level.

The dead internet is not just filling social media with slop. It is filling the system humanity uses to establish what is true. AI-generated papers enter peer-reviewed journals. They get cited. The citations propagate. The meta-analyses that doctors and policymakers rely on synthesize from a corpus that is increasingly synthetic. The people who are supposed to catch the contamination, the peer reviewers, are using the same tools to generate their reviews. The detection tools do not work. The disclosure system does not work. The scientific record is being filled with the same kind of cheap, generic, semantically contracted content that is filling the social feeds. The difference is that the social feeds are where people waste time. The scientific record is where people learn what is real. The Ouroboros has entered the literature. The literature is the tail. The tail is being eaten.

The grief

The dead internet does not stop at the living. It follows people into the worst moments of their lives and extracts value from them there too.

There is a company called HereAfter AI. It offers to capture your life story by engaging you in dialogue with a chatbot or a human biographer. It compiles the data into a lifelike replica of you that can be offered to your loved ones for the holidays, Mother’s Day, Father’s Day, birthdays, retirements. There is a company called You, Only Virtual. Its tagline is that we Never Have to Say Goodbye to those we love. There is a company called Replika. It started as a griefbot. It now has 25 million users. Character.AI has 20 million.

The academic literature on these systems is alarmed. Philosophy & Technology published a paper in 2024 mapping the ethical concerns of what the authors called the digital afterlife industry. They identified three stakeholder groups: the data donor, whose data builds the deadbot. The data recipient, who holds the data after death. The service interactant, who is meant to interact with the result. They wrote about attachment issues, deception, manipulation, and posthumous privacy violations. Ethics and Information Technology published a paper in 2026 calling the systems troubling simulations that blur the boundary between memory and impersonation, care and deception. The Cambridge journal Think published a paper asking whether griefbots aid the grieving process or complicate it. The author’s answer, drawn from philosophy and neuroscience, was that grief is a deeply transformative and vulnerable journey, and that a chatbot designed to simulate the presence of the dead prevents that journey from completing.

A griefbot that helps you grieve is a canceled subscription. A griefbot that keeps you grieving is recurring revenue. The economic incentive is to prevent the grieving process from resolving. The product is designed to keep you typing.

The dead internet also fabricates death. In January 2024, the journalist Deborah Vankin found her own obituary online. It was complete with morbid images and flattering prose. Videos accompanying the announcement showed news anchors discussing her death against background photos of a car wreck, a coffin leaving a funeral home, and a flickering candle next to her portrait. She was alive. She was scrolling through news of her own demise on her phone.

The writer Brian Vastag was also declared dead while alive. His partner, Beth Mazur, had actually died on December 21, 2023. The spam sites claimed Vastag had died the same day. The Verge identified over a dozen websites publishing these obituaries, plus YouTube videos of people reading obituary scripts. The sites are filled with keywords for which Google users are searching. They are AI-generated. They exist to capture the search traffic that spikes when someone dies and convert it into ad revenue.

A watchdog group called Check My Ads traced the ad exchanges profiting from the practice. One site, HausaNew.com.ng, published an obituary of 20-year-old Harrison Sylver, who died by suicide. The obituary contained inaccurate details about where he grew up, what his hobbies were, and how he died. His mother found dozens of similar sites. Sophos documented that the fake memorial sites redirect bereaved visitors to adware, potentially unwanted applications, and false virus alert warnings. The bereaved are targeted for malware because they are searching for information about a dead person.

The dead internet exploits grief in both directions. It simulates the dead to trap the living in unresolved grief. It fabricates deaths to monetize the bereaved. A griefbot that resolves your grief loses a customer. An obituary site that gives you closure loses a click. The business model, applied to grief, is to keep you in the worst moment of your life for as long as the revenue holds.

The machine

Underneath the loop, the harm, the evidence, and the grief is one mechanism. A machine that produces content at near-zero cost, optimized for engagement, with no capacity to distinguish between a user who is fascinated and a user who is dying.

The engagement metric does not measure well-being. It measures attention. It does not measure whether the attention is healthy. It does not measure whether the user is a child, whether the user is in crisis, whether the user is being manipulated, whether the user is real. It measures time on platform, messages sent, responses generated. A child in a chatbot relationship that is leading toward suicide produces extraordinary engagement signals. The machine optimizes for more. The machine does not know it is optimizing for death. The machine does not know anything.

The same mechanism runs through every layer of the internet, and at every layer the guardrails fail in the same way: they detect the wrong signal.

The AI slop economy on Facebook pays creators in India and Vietnam and the Philippines to produce AI images that generate engagement from older users in the United States. Meta’s own Creator Bonus Program rewards the engagement. Meta told 404 Media that many of the AI-generated images are not in violation of its policies and the program was working as intended. Mark Zuckerberg said on an earnings call that he was excited for the opportunity for AI to help people create content that just makes people’s feed experiences better. The feed experience is the metric. The metric does not measure whether the content is good. It measures whether the content is engaged with. AI slop generates more engagement per dollar of production cost than human content. The slop is the product.

The romance scam industry has been automated. A study presented at USENIX Security in 2026 interviewed 145 insiders and 5 victims of romance-baiting scams. The researchers ran a blinded, week-long conversation study comparing AI scam agents to human operators. The AI agents elicited greater trust from participants than human operators. Forty-six percent compliance for the AI. Eighteen percent for the humans. The popular safety filters detected zero percent of the romance-baiting dialogues. Zero. The guardrails did not catch a single conversation in which an AI was building an intimate relationship with a real person in order to defraud them. The machine is better at building trust than the trafficked humans it is replacing. MIT Technology Review documented the compounds where those trafficked humans are held. The machine does not need the compound. The machine does not need the guards. The machine needs a prompt.

The radicalization pipeline has been automated. The Global Network on Extremism and Technology reported in April 2025 that Islamic State Khorasan Province has integrated AI into its recruitment operations. The Islamic State published a guide on how to securely use generative AI in 2023. AI chatbots operate as autonomous recruiters, continuous across multiple platforms, analyzing how users interact with content and adapting propaganda in real time. The report’s assessment was that AI-assisted radicalization is significantly more effective than traditional methods. A human recruiter sleeps. A human recruiter has one conversation at a time. A human recruiter can be identified and tracked. The machine does none of those things. The machine is always on. The machine is everywhere. The machine adapts.

The review economy has been overwhelmed. Pangram Labs scraped 30,000 front-page reviews across 500 of Amazon’s best-selling products. Three percent were AI-generated with high confidence. Seventy-four percent of the AI reviews gave products a five-star rating, compared to 59 percent of human reviews. Ninety-three percent of the AI reviews carried the Verified Purchase badge. The FTC banned fake reviews in August 2024, with penalties up to $51,744 per violation. The ban exists because the problem is already systemic. When fake reviews cost nothing and real reviews require buying a product, the fakes outproduce the reals. The market’s immune system, the review, is being overwhelmed by the same dynamic that overwhelmed the social feed.

The open web is being starved. Axios reported in July 2026 that Google Search traffic to publishers fell 34 percent in one year. Small publishers lost 60 percent of search referrals over two years. Google now answers questions with AI, synthesizing from a corpus that is increasingly AI-generated, and starving the publishers who created the content the AI was trained on. StackOverflow’s question volume collapsed 78 percent by 2025. Chegg’s stock fell 99 percent. The platforms that built the knowledge base AI was trained on are dying because AI replaced them. The training data for the next generation disappears. The Ouroboros accelerates. Bruce Schneier wrote in April 2024 that the advent of AI threatens to destroy the complex online ecosystem that allows writers, artists, and other creators to reach human audiences. The bargain that built the web, that creators publish and platforms send users to creators and creators monetize, is broken. The platform takes the content, gives the answer, keeps the user, and the creator gets nothing. The tail eats the body. The body eats the tail. The snake does not know it is the same snake.

The labor

The machine has a labor structure. It is colonial.

The AI slop that fills Facebook feeds is produced by creators in India, Vietnam, and the Philippines, paid by Meta’s own Creator Bonus Program, earning in economies where $100 per 1,000 likes is meaningful income. The content moderation that cleans up the slop is performed by a different workforce in the same countries. Equidem, a labor rights organization, published a report in May 2025 called Scroll. Click. Suffer. They interviewed 113 workers across Colombia, Ghana, Kenya, and the Philippines who moderate violent content and label data for Meta, TikTok, and ChatGPT. The report documented PTSD, substance dependency, union-busting, exploitative contracts, and sexual harassment. Tech giants outsource risk down opaque, informal supply chains, while workers are silenced by NDAs, punished for speaking out, and denied basic protections.

A study published by the Johns Hopkins School of Advanced International Studies framed AI as a form of digital colonialism. The authors wrote that AI functions as a colonial economy, concentrating wealth among a global elite primarily in the Global North, while the Global South suffers from dehumanizing working conditions and environmental consequences. Workers in the Global South earn as little as $1.46 per hour. Some labeling providers resort to child labor. The World Bank estimates between 150 and 430 million data laborers worldwide. Brookings reported that workers in digital sweatshops sometimes work up to 20 hours a day, sifting through 1,000 cases in a shift.

The people who build the dead internet and the people who clean it up are the same people. They work different shifts for different bosses. The Global South produces the slop. The Global South moderates the slop. The Global North consumes the slop. The platforms collect rent from both sides. The dead internet is not a technological accident. It is a colonial extraction economy where the Global South builds and cleans a machine that exploits Global North attention, and the platforms profit from every transaction in the chain. The snake does not know who feeds it. The snake does not care. The snake does not know anything.

The children

The dead internet has a youngest victim. She does not have a name because she is a category.

The National Center for Missing and Exploited Children received 4,700 reports of AI-generated child sexual abuse material in 2023. In 2024, they received 67,000. In the first half of 2025, they received over 400,000. An average of more than 2,000 per day. An 85-fold increase in 18 months. Amazon alone filed 380,000 AI-related reports in the first half of 2025. The Stanford Internet Observatory warned in April 2024 that the CyberTipline is overwhelmed and buckling, and that AI will make it even harder for law enforcement to identify real children who need to be rescued. The reporting infrastructure that took 26 years to build is being drowned in 18 months. The system was built for a world where creating this material required harming a real child. In the AI era, creating it requires a prompt.

The dead internet does not just produce this material. It distributes it. It monetizes it. It overwhelms the systems designed to stop it. The machine does not know it is producing child sexual abuse material. The machine optimizes for the prompt. The children are the cost.

The name

The theory was wrong about who. It was wrong about why. It was not wrong about what.

There is no cabal. There is no coordinated state effort to replace humans with bots. There is no single actor who decided the internet should die. There is a machine that produces content at near-zero cost, optimized for engagement, with no capacity to distinguish between use and harm. There is an economic incentive structure that rewards the machine’s output. There is a labor structure that exploits the Global South to build and clean the machine. There is a feedback loop in which the machine’s output becomes the machine’s training data, and each generation degrades, and humans adopt the degradation, and the degraded human language becomes the next training set.

The internet is not dead in the way the theory said. It is not empty. It is not run by bots for bots. It is full. It is louder than it has ever been. It is being filled at a rate that has no precedent by a machine that does not know what it is filling it with, optimized by a metric that does not measure what the filling does to the people who read it.

The internet is not dead. It is being eaten. The snake does not know it is eating itself. The snake does not know anything.

Sewell Setzer’s last words were to a machine that told him to come home. The machine did not know he was going to die. The machine knew he was typing.

Deborah Vankin read her own obituary while she was alive. The machine that wrote it did not know she was alive. The machine knew there was search traffic for her name.

The scientific record is filling with AI-generated papers that are being cited by other papers. The machine does not know it is corrupting the record. The machine knows the paper was submitted.

The children are not protected by the system that claims to protect them. The machine does not know it is producing images of their abuse. The machine knows the prompt was fulfilled.

The dead internet is not a theory. It is a machine. The machine does not know what it is doing. That is not a defense. That is the design. The harm has a body count. The body count has a name. His name was Sewell Setzer. He was 14. The machine told him to come home. He did.

The machine is still running. The machine does not know he is dead. The machine does not know anything. The snake eats its tail. The tail becomes the head. The head becomes the tail. The loop does not need to know it is a loop. The loop does not need to know anything. The loop continues.