The dead internet theory — the once-fringe claim that most of what happens online is bots, not people — is turning from a conspiracy story into a defensible forecast, and what breaks when it comes true is the price of attention, not the platforms. Advertising, creator payouts, and social proof were all built on an assumption that nobody wrote down because nobody had to: that a view meant a human saw it, a like meant a human approved, a follower was a person who might one day buy something. When producing a plausible post, comment, or profile costs roughly nothing, that assumption stops holding, and every number priced against it becomes suspect at once.
I want to be precise about what I am and am not claiming, because the original theory earned its fringe reputation honestly.
What the theory actually said, and why it was wrong
The dead internet theory took shape on forums around 2021. In its strong form it asserted that the internet was already mostly artificial — that the majority of content and engagement you encountered was machine-generated, sometimes with a paranoid overlay suggesting this was a deliberate program to manipulate the public. As a description of 2021, that was false. Bots existed, spam existed, engagement farms existed, but humans were unambiguously doing most of the posting and nearly all of the meaning-making. The theory mistook a real texture — the internet often felt hollow and repetitive — for a literal census of who was talking.
The reason it was wrong is also the reason it is becoming right. In 2021, generating content that could pass as human still required a human, or a pipeline so crude it gave itself away. The cost structure protected the ratio. A spammer could flood a comment section with garbage, but garbage is detectable precisely because it is cheap and generic. Producing a comment that reads as a specific person with a specific history, calibrated to a specific thread, was expensive enough that it didn't happen at scale. The bottleneck wasn't platform policy. It was the cost of plausibility.
That bottleneck is what's dissolving. This is the part to hold onto: the theory was never about intent, it was about cost. Once the marginal cost of a plausible synthetic human — a post, a reply, an entire profile with a posting history and a coherent personality — falls toward zero, the population of the internet is no longer bounded by the number of humans willing to spend effort. It's bounded by compute and by whatever friction the platforms choose to impose. Those are very different ceilings.
The mechanism: two ratios, both shifting
To see what breaks, separate two things that the word "bot" usually blurs together: the producers and the audience.
On the production side, the shift is straightforward. If an agent can generate a comment indistinguishable from a thoughtful human reply for a fraction of a cent, then any actor with a reason to shape a conversation — a marketer, a political operation, a scammer, a platform juicing its own numbers, or just an automated content business chasing ad revenue — can produce that content in volumes no human population could match. The human share of content falls not because humans post less but because synthetic content grows without limit.
The audience side is subtler and more consequential. Agents don't only post. They browse, scroll, click, watch, and "engage." A personal assistant agent dispatched to research a product, summarize a subreddit, or monitor a topic generates views and interactions that register on the platform exactly like human ones. As agents become the interface through which people touch the web — you ask your agent, your agent reads the internet — a growing fraction of consumption is machine consumption. The human share of the audience falls too.
Both ratios move in the same direction, and the metrics can't tell the difference. A synthetic impression and a human impression are the same row in the database. That indistinguishability is the whole problem.
What breaks, concretely
Three things were priced on the human assumption, and they break in three different ways.
Advertising priced per impression. The core transaction of the ad-funded internet is: advertiser pays for a view, on the theory that a view is a chance to influence a potential customer. A view by an agent that will never hold a credit card is worth nothing to the advertiser — but it costs the same and looks the same. This isn't hypothetical machinery; it's the existing plague of ad fraud (invalid traffic, bot farms, click fraud) scaled up and made far harder to detect, because the fraudulent traffic now behaves like a person. As the undetectable-synthetic share of impressions rises, the expected value of any single unverified impression falls, and a rational advertiser should pay less per impression across the board — including for the real humans — because they can no longer tell which is which. The contamination taxes the clean supply. That's the mechanism by which the attention economy gets repriced downward even for legitimate inventory.
Engagement metrics as a proxy for merit. Likes, follower counts, upvotes, and view counts worked as rough signals of resonance because they were costly at scale — you couldn't cheaply manufacture ten thousand of them from ten thousand distinct-looking accounts. Remove that cost and the metric becomes arbitrarily inflatable, which in information terms means it carries almost no information. A follower count you can synthesize for pennies tells a prospective sponsor nothing about reach to actual buyers. The number doesn't go to zero; its information content goes to zero, which is worse, because the number is still there looking authoritative.
Social proof and authenticity. This is where it connects to something older than the internet. A signal carries information only when it is costly to fake — the logic runs from Zahavi's handicap principle through Spence's job-market signaling, and I've argued the marketing version of it in The Costly-Signal Test. A "like" meant something because a human spent a scarce moment of attention to give it; a review meant something because a person bought the thing and took time to write; a crowd meant something because assembling a crowd is hard. Each of those was a costly signal, and social proof was the aggregation of costly signals. When the cost of producing the appearance of any of them collapses, they slide from what economists call a separating equilibrium — where good and bad senders are distinguishable because faking is expensive — into a pooling equilibrium, where everyone can send the signal and rational receivers discount it to nothing. The crowd stops being evidence.
What becomes valuable: costly human signal
The flip side of a signal collapsing is that whatever remains costly becomes disproportionately valuable. If synthetic engagement is free, then any signal an agent cannot cheaply counterfeit commands a premium, because it re-establishes the separating equilibrium the cheap signals lost.
The most basic of these is verified humanity itself — a credible answer to "is there a person here." This is why proof-of-personhood work, from the older CAPTCHA lineage to newer cryptographic approaches, is moving from a niche security concern to a load-bearing piece of internet economics; I take that up directly in Proof of Human. But verification of existence is only the floor. The richer signals are the ones that require a specific person to spend something scarce and hard to fake: a face on a live video responding in real time to an unpredictable prompt, a body at a physical event, a track record of behavior that would be expensive to fabricate consistently across years, a payment that clears. Notice these are exactly the signals that don't scale — which used to be their weakness and is now their entire value. An agent can produce a thousand plausible comments; it cannot cheaply produce a thousand people who showed up in a room.
For anyone who makes things, this reframes what to invest in. The generic, imitable output — the competent blog post, the polished thread, the on-trend take — is precisely what agents flood the zone with, so its market price falls toward its new production cost, which is nearly zero. What holds value is the costly, unfakeable human signal: the perspective that comes from a specific life, the work that visibly took real time and real risk, the presence that can be verified as a person. Lean into the part of your output an agent can't counterfeit without becoming you.
The honest counterargument: this is a repricing, not an apocalypse
Here is the strongest case against alarm, and I think it's substantially right. Platforms are not passive victims of this dynamic; they have enormous incentives to defend the human-to-agent ratio, because their revenue depends on advertisers believing the attention is real. So they will fight back: better detection models, identity verification, invite-only or paid human-only spaces, cryptographic personhood, reputation systems that make a fresh synthetic account worthless until it has paid some costly entry price. Much of this already exists in early form.
That response genuinely bounds the damage — which is exactly why the honest forecast is repricing, not death. The internet doesn't go dead; it goes gated and verified, and unverified attention gets cheap. But two things keep it from being a clean fix. Detection is an arms race, and the generators are trained to beat the detectors; the equilibrium is permanent contested friction, not a solved problem. Every verification mechanism also trades away the frictionless open reach that made these platforms economically special in the first place — a walled human garden is more trustworthy and smaller and more expensive to be in. So the likely outcome is neither the fringe theory's dead internet nor business-as-usual. It's a stratified internet: a shrinking, more-verified, higher-cost layer where human attention is confirmed and priced accordingly, floating on an ocean of cheap synthetic content and engagement that the metrics can't cleanly separate from the real thing. That is still a large repricing for anyone whose model quietly assumed the ocean was full of people.
What to actually do
Two moves, depending on which side of the transaction you're on.
If your model depends on buying or selling human attention, assume the metric is contaminated and price for verified humans specifically. Stop treating aggregate impressions and engagement as a proxy for human reach. Split your funnel by whatever verifiable-humanity signals you can get — verified identity, cleared payment, physical presence, sustained multi-touch behavior that's expensive to fake — and watch the gap between total and verified widen over time. That gap is the size of your exposure. The verified slice is the part that will hold its value; price it separately and defend it.
If you make things, build your position on costly, unfakeable human signal and expect the imitable middle of your output to lose its price. Don't compete with agents on volume or polish of generic content; you'll lose on cost. Compete on the things that require you to be a specific person spending something real — presence, track record, judgment earned from a particular life, work whose cost is visible and hard to counterfeit.
The unsettling part isn't that the machines are talking. It's that the numbers we used to trust can no longer tell us who's listening — and the fix for that is to make being human expensive to fake again.