I've made real friends on the internet. Fallen for someone through a screen before I ever saw them in person. Argued with strangers who, years later, I'd trust with something that actually mattered. None of that was fake, and I'm not about to pretend it was.
But somewhere in the last year or two, a different kind of doubt showed up, one I didn't used to have. Not about the people I already know are real. About everyone else: the comment under a video, the five-star review, the person arguing with me in a thread I'll forget by tomorrow. I used to just assume there was someone there. I'm not sure I can assume that anymore.
And the part that actually bothers me isn't whether that's true. It's that I don't think I'd be able to tell either way.
That's the actual subject here: not the usual complaint that AI is flooding the internet with garbage. That part's true, but it's boring, and garbage is easy to spot. What's stranger is what happens once the fake stuff gets better than the real stuff, good enough that telling the difference stops being possible at all.
The Internet Was Built for People
The internet's first trick, decades ago, was closing a distance. You could find someone on the other side of the world who cared about the exact same obscure thing you did, a band, a disease, a hobby, some small strange corner of history, and just talk to them. Forums. Mailing lists. Blog comment sections nobody else read. Clunky, slow, occasionally full of strangers being weird at each other, but real in a way that was easy to take for granted at the time.
Herbert Simon had already named the thing that would eventually strain that system, back in 1971, decades before anyone had a social feed. In an information-rich world, he wrote, "a wealth of information creates a poverty of attention," because information consumes the one resource every reader has in fixed supply. He was writing about libraries and internal memos. He had no idea how right he'd turn out to be.
Then We Learned to Manufacture Attention
Every platform built after that quote has essentially been a machine for rationing scarce attention, and for a long time, that just meant getting good at being interesting. Blogs turned into social platforms turned into a full creator economy: millions of people making things for an audience of strangers, competing for the same fixed pool of eyeballs.
Then AI joined that competition, not as a faster tool but as an entirely different kind of competitor. A model doesn't get tired, doesn't procrastinate, doesn't need a salary. Prompting a hundred blog posts is cheaper than commissioning ten from a writer. Generating an infinite scroll of product photos is cheaper than booking one shoot. Running a bot around the clock is cheaper than staffing a forum. No conspiracy required. Just a spreadsheet.
The traffic numbers already show it. Automated activity overtook human activity on the internet for the first time in a decade in 2024, bots now make up the majority of web traffic, according to Imperva's 2025 Bad Bot Report, with bad bots alone accounting for more than a third of it. Whatever year you assumed that would happen, it already has.
The Weird Part Is That AI Got Good at This
For a long time, effort was one of the easiest ways to recognize a human. A badly written blog post took someone an hour they didn't get back. A photograph meant someone physically went somewhere. A portfolio meant someone built something, badly at first, then less badly. A thoughtful comment meant a stranger actually sat there and thought about what you'd said. None of that proved talent. It proved a person had spent something (time, attention, effort) to make it.
AI severs that link almost completely, not by getting merely competent, but by getting good enough to beat humans at being human, provably. In 2024, UC San Diego researchers ran a rigorous Turing test and found GPT-4, posing as a person in five-minute conversations, was judged human 54% of the time, close behind actual humans, judged human 67% of the time. A 2026 follow-up went further: with the right persona prompt, GPT-4.5 was judged human 73% of the time, more often than the real humans it was up against.
It didn't just pass. It out-humaned the humans.
There's something genuinely unsettling about how far this extends. In a widely cited 2022 study in the Proceedings of the National Academy of Sciences, researchers found that people couldn't reliably tell AI-generated faces apart from real ones, accuracy hovered around chance, and, worse, rated the fake faces as slightly more trustworthy. We've apparently reached a point where a face can read as more trustworthy precisely because nobody ever had to live behind it.
Ted Chiang once called large language models a blurry JPEG of the internet: a lossy compression of everything humans have written, smoothed into something fluent even when it's wrong. Nobody quite predicted how flattering that blur would turn out to be. A well-written paragraph used to be proof someone worked hard on it. Increasingly, it's just proof someone had a good prompt, which is bad news for anyone whose portfolio, Instagram grid, or LinkedIn profile has quietly been doing the work of a résumé. Once a 19-year-old with zero experience can generate all three in an afternoon, the résumé stops proving anything.
It isn't only that AI content can pass as human now, either. Human writing is starting to sound like AI, and that part isn't a coincidence. Researchers who tracked word usage across millions of PubMed abstracts found a specific set of words, "delve," "meticulous," "boast," "underscore," "commendable," among others, spiking sharply in 2024. The trend had quietly started a couple of years before ChatGPT existed, then accelerated hard once it did.
The stranger finding is what happened outside of writing entirely. Researchers at the Max Planck Institute for Human Development analyzed hundreds of thousands of unscripted YouTube talks and podcast episodes and found people saying the exact same AI-favored words out loud, more often, within months of ChatGPT's release. One of the study's own authors said he first noticed it in himself. He'd started saying "delve" more, and went looking to see if it was happening to everyone else too. It was.
The internet didn't just fill up with AI-written content. It started teaching actual people how to talk.
The Dead Internet Theory Was Almost Right
None of this needed a conspiracy to become true, which is almost funny, because a conspiracy theory got there first. The "Dead Internet Theory" started on fringe forums in the mid-2010s and went semi-viral in 2021: a long, paranoid post claiming most of the internet was already fake, bots talking to bots to manipulate whatever humans were still logging on. Easy to dismiss at the time. This was before ChatGPT even existed.
A 2025 academic survey of the theory found something more useful than the original claim. Strip out the part about coordinated manipulation, and what's left is a pretty accurate description of where platforms were already heading: bots, algorithmic content, and engagement metrics steadily crowding out real interaction. Even Sam Altman, whose company built a fair share of the tools doing that crowding, has said publicly that there now seem to be a lot of AI-run social media accounts.
Even the internet's favorite statistic about its own decline turned out to be exactly this kind of thing. You've probably seen the claim that 90% of online content will be AI-generated by 2026, usually credited to Europol. Europol never published that number. Nobody can find where it actually came from. It just keeps getting repeated, outlet to outlet, because repeating a viral claim is cheaper than checking one: a stat about the fake internet, itself spreading like the fake internet.
The Internet, Eating Its Own Tail
There's a technical version of the same problem, and it's less a cultural observation than a hard limit. AI trained on AI-generated content gets worse, not better. A 2024 Nature paper, bluntly titled "AI Models Collapse When Trained on Recursively Generated Data," found that generative models trained indiscriminately on a mix of real and machine-made content degrade across generations, losing accuracy and diversity in a process researchers call model collapse. The open internet, the default place these models get scraped from, is exactly that mix.
So the self-referential loop people worry about (AI writing the articles that AI summarizes into the posts that AI turns into the videos AI recommends back to you) isn't just an aesthetic complaint. It's closer to inbreeding. A system that keeps talking to itself doesn't just get stranger over time. It gets measurably dumber.
Imperfection Might Become Proof of Life
If effort stops proving a human made something, maybe the proof shifts somewhere else entirely: not to what's good, but to what's obviously, uncorrectably flawed.
Think about the stuff the internet has spent twenty years training us to delete. A photo with someone's thumb over the lens. A typo in a birthday message that somehow makes it warmer instead of worse. A video someone filmed vertically because there wasn't time to think about it. A voice cracking halfway through a 20-second voice note. An inside joke that means nothing to anyone outside the group chat. A poster somebody made themselves for a show nobody's heard of, slightly crooked, definitely not on brand.
We call that bad content. But bad content is expensive to fake convincingly. Not impossible, just pointless, since its entire value was that nobody optimized it. Perhaps that's what starts to matter: not polish, but the specific, useless evidence that an actual person was in the room when it happened.
This isn't hypothetical anymore, either. In August 2026, a Chinese animated film called Niu Lai ("Here Comes the Cow") opened to almost nobody: about a thousand dollars in ticket sales over its first ten days. It was made by a mother and son over five years, self-taught, released through a production company that, until recently, did interior design and decoration, on what's now reported to be roughly a $200 budget. The animation is famously terrible: stiff, jerky, sincerely strange. Then clips of it started circulating as a joke. People bought tickets specifically to see how bad it was. Within a few weeks it had made more than $4 million, likely making it one of the most profitable films ever made relative to what it cost.
The reason people gave for showing up wasn't ironic distance from bad art. Social media users kept saying, explicitly, that they preferred the rough, human-made animation to anything AI could have produced. Theaters that had no official poster to work with just had staff draw their own by hand, crooked lettering and all, and those went viral too. A human being sincerely tried, and mostly failed, to make something well. In 2026, that turned out to be worth more than four million dollars to an audience that could have watched something technically flawless instead.
There's real psychology behind why this might work on us. In a 1966 study, the psychologist Elliot Aronson had people listen to a speaker who either aced a quiz or fumbled it. When the highly competent speaker spilled a cup of coffee at the end, listeners liked them more, not less, the mistake made an intimidatingly perfect performance feel reachable. It's called the pratfall effect: flawlessness reads as distant, a small blunder reads as real.
Some platforms are already building toward a much blunter version of the same idea, less "humanizing imperfection," more demanding proof outright. World, the identity project Sam Altman started at OpenAI, scans people's irises to issue something it calls a "proof of personhood" credential: not who you are, just that you're a real, unique human and not a bot. Sit with that for a second. The internet was invented to connect people. We're now building infrastructure to prove the people on it are actually people, before it'll let them talk to each other.
Twenty years spent teaching us to remove every flaw from what we post. The flaws might be what gets valuable next.
So We Go Outside
Zoom out from platforms and screens entirely, and the same pattern's been building for a while, for reasons that have nothing to do with AI. In 2023, the U.S. Surgeon General declared an epidemic of loneliness and isolation: roughly half of American adults reported feeling lonely, with health risks comparable to smoking up to fifteen cigarettes a day. One finding stood out: people who spent more than two hours a day on social media were roughly twice as likely to report feeling isolated as those who spent less than thirty minutes.
None of this is new, and AI didn't start it. Robert Putnam was documenting Americans' retreat from clubs, leagues, and civic life back in 2000, in a book called Bowling Alone, years before smartphones existed. AI isn't the cause here. It's the deadline.
Because here's the thing an increasingly synthetic internet can't touch, no matter how good it gets: it can't fake you being somewhere. So maybe the most interesting reaction to all of this won't be another app, or a better filter, or a smarter way to verify who's real online. It'll be people meeting for dinner. Going to a concert with bad sound and worse parking. Joining a five-a-side team with strangers who turn into friends by week three. Playing cricket in a park with people whose last names you don't know yet. Sitting in a café having a conversation nobody's recording, that produces nothing except the conversation itself.
That last part matters more than it sounds like it should. We've spent a decade turning experience into content: every dinner is a photo, every trip is a post, every feeling is a caption waiting to happen. An evening that generates absolutely nothing shareable used to feel like a wasted one. Maybe that flips. Maybe the appeal, going forward, is specifically that nothing came of it. You were there. Nobody else needs to know.
Being There Becomes the Flex
There's a status signal already forming around this. The old flex was follower count, a color-graded feed, a LinkedIn with seven hundred connections. Increasingly, it's doing something nobody can screenshot.
A recent Morning Consult survey found Gen Z and millennials showing outsized interest in switching to a "dumb phone," something with no algorithmic feed at all. This year's viral "Flip Phone Summer" turned that interest into an actual, if small, market shift, with bare-bones phones increasingly worn as a status symbol instead of a downgrade. None of this is about to kill the smartphone. But opting out of the content flood is starting to read as taste, not nostalgia.
The Internet Becomes Furniture
The biggest shift might also be the easiest to miss, because it's already halfway finished: the internet stops being a place you go and starts being infrastructure you stand on.
In 1991, before chatbots existed, the computer scientist Mark Weiser wrote a short, prophetic essay called "The Computer for the 21st Century." Its opening line still holds up: "the most profound technologies are those that disappear." Weiser's point was that real technological success looks like electricity, or writing, woven so completely into daily life that nobody notices it's there anymore. He meant computers fading into the walls. Maybe the internet ends up going the same way, except this time it's AI agents doing the fading: systems that browse, negotiate, buy, and publish on our behalf while we're doing something else entirely.
People still say "going online" the way an older generation said "going to the movies," like it's a destination. Give it a few more years, and that phrase might sound as strange as saying you're "going to electricity."
After the Internet
AI's entire pitch has been abundance: faster, cheaper, more available, more instantly good. What if that abundance is exactly what pushes people out? When everyone can create anything, creating stops proving anything. When everything can sound human, an actual human gets harder to find in the noise. Which might mean, for the first time since the internet began, that finding one is worth the effort.
Maybe that's the real ending to the Dead Internet Theory. Not that the internet quietly died and nobody noticed, but that it lived on, got remarkably good at sounding like us, and in doing so, taught us to go find each other somewhere it isn't.
None of this kills the internet. It'll still be there, doing what infrastructure does: running underneath everything, mostly unnoticed, mostly agents talking to agents. What might actually end is a smaller, stranger idea, the idea of being online. Of treating a feed like a room you walk into, a place you go to find people, instead of a utility you pass through on your way to finding them somewhere else.
The internet doesn't have to disappear for that to be true. It just has to stop being where the humans are.