The open web is no longer mostly people. Bots now carry more traffic than humans. Most new pages have some machine text in them. I keep watching people leave public feeds for rooms that are private, encrypted, or gated.
Tap to enlargeThe Numbers Behind the Dead Internet
"Dead internet" started as a 2021 forum rant. By 2025 the security vendors were printing numbers that matched the rant.
The 2025 Imperva Bad Bot Report put automated traffic at 51% of the web in 2024. First time in a decade bots beat people. Of that 51%, about 37 points were "bad bots": scrapers, spam, credential stuffing, manipulation. The other 14% were crawlers, uptime checks, and APIs someone actually invited.
HUMAN Security said automated traffic grew eight times faster than human traffic in 2025. Generative AI is the obvious reason. A bot that can write a forum post, a review, or a reply used to take an engineer. Now it takes an API key and a script.
The content layer is worse. Ahrefs analyzed 900,000 newly detected English-language pages in April 2025 and found that 74.2% contained some AI-generated text. Only 2.5% were classed as pure AI. 25.8% as pure human. The rest were mixed. That is a detector score on one crawler's new-page sample, not a census of the live web. Some later estimates project that up to 90% of online content could be synthetically generated by the end of 2026. Those are extrapolations.
On social media, the picture is equally degraded. Industry estimates place bot accounts at 5-15% of total accounts across major platforms, spiking above 50% during viral events or politically charged topics. Meta removes millions of fake accounts quarterly. In April 2026, X (formerly Twitter) announced initiatives to suspend hundreds of bot accounts per minute, an acknowledgment of scale that reveals how far the problem has progressed.
The Collapsing Signal-to-Noise Ratio
The practical consequence is that the public internet is becoming unusable for its original purpose: finding accurate information from trustworthy sources.
Google Search, the primary interface between humans and the web for two decades, is drowning in AI-generated SEO spam. HouseFresh, an independent product review site, published research in 2024 documenting how Google's search results had become dominated by large publishers running algorithmically optimized content farms. The problem has intensified since. Search queries for product recommendations, medical information, and technical how-to guides increasingly surface AI-generated pages that are syntactically fluent but factually unreliable.
The response is measurable. Reddit traffic from Google Search increased 39% between 2023 and 2024, as users appended "reddit" to their queries to bypass AI-generated results and reach human-written opinions. Google responded by signing a $60 million deal with Reddit to license its content for AI training, turning the escape hatch into more training data for the machines users were trying to avoid.
When users add "reddit" to every Google search to find human-written content, the search engine has failed its core function. The signal-to-noise ratio has collapsed.
The term "dark forest" comes from Liu Cixin's science fiction novel The Dark Forest. In the book, the universe is full of civilizations that hide from each other because any detectable signal invites destruction from a superior predator. On the internet, the predators are bots. Any public expression, a blog post, a forum comment, a product review, a photograph, is immediately targeted. Blog posts are scraped and reconstructed as SEO spam. Forum answers are copied into AI training datasets. Photographs are downloaded, stripped of attribution, and used to generate synthetic variations. To be publicly visible is to be harvested.
The Retreat to Private Spaces
Humans are leaving the public layer.
Encrypted messaging. WhatsApp passed 3 billion monthly active users on Mark Zuckerberg's Q1 2025 earnings call in April 2025. Meta does not print WhatsApp MAU in the 10-K. Telegram founder Pavel Durov said on 19 March 2025 that the app had significantly over 1 billion monthly active users. That is a founder claim, not an audited census. Signal remains the privacy-first nonprofit option. Public MAU figures for Signal are sparse and should not be treated as a platform census. The migration pattern is clear: conversations that once happened on public Twitter threads or Facebook posts are moving into group chats where membership is verified and bots cannot enter.
Private communities. Discord now hosts millions of private servers where access requires invitation, role verification, or payment. The most valuable conversations about investing, technology, and specialized knowledge happen inside these servers, invisible to search engines and inaccessible to scrapers. Substack, last priced at $585 million pre-money on its 2021 Series B and 2023 writer round, represents the same trend in publishing: writers with paid subscribers build direct relationships with verified human readers, bypassing the algorithmic feed entirely.
Closed platforms. The explosive growth of private Instagram accounts, Close Friends stories, and group-only content reflects a broader pattern. Users are voluntarily restricting their audience to avoid the bot-infested public layer.
The Identity Problem
In the dark forest, the question shifts from "what information is available?" to "who is human?"
Current verification systems are inadequate. CAPTCHAs, once the standard line of defense, have been consistently defeated by AI since 2023. Phone number verification creates a thin barrier. Bot operators purchase SIM cards in bulk for pennies. Email verification is meaningless when a single person can generate unlimited email addresses.
The only durable solution is cryptographic proof of personhood. World (formerly Worldcoin), founded by Sam Altman, is the most aggressive attempt. The project uses iris-scanning hardware ("Orbs") to generate a cryptographic proof that a person is a unique human being, without revealing their identity. On 17 March 2026, CoinDesk showed World's live verification meter at 17,912,203. Later March press put iris-verified humans near 18 million and World App users near 39 million. Those are company-adjacent figures, not a government ID census.
World ID uses zero-knowledge proofs (ZKPs) to verify humanness without exposing personal data. The Orb scans the iris, generates a unique hash (an "IrisHash"), and stores a commitment on-chain. When a service asks "is this a real person?", the protocol proves the answer is yes without revealing which person. This is the cryptographic equivalent of a bouncer checking your ID at the door without recording your name.
The integration is already happening. Tinder partnered with World to combat fake profiles. Razer uses World ID to verify human gamers. Reddit, Discord, and several decentralized platforms are experimenting with proof-of-personhood gates for posting privileges. Access to the human internet is becoming contingent on proving you are human.
The New Digital Divide
This retreat creates a new form of inequality. The public internet, free, open, accessible, is being degraded by synthetic content. The private internet, encrypted, verified, often paid, is where authentic human discourse and high-quality information are concentrating.
The affluent and digitally literate will inhabit curated Discord servers, paid Substack communities, and invitation-only forums. Everyone else will be left in the bot-generated wasteland of the open web, consuming synthetic media and interacting with synthetic personalities without knowing it.
This has direct implications for democratic participation. Political discourse that moves behind encryption is discourse that cannot be publicly scrutinized. Journalism that moves behind paywalls is journalism that most citizens cannot access. Expert knowledge that concentrates in private communities is knowledge that the public cannot benefit from.
The dark forest may intensify the polarization problem. If humans retreat into small, verified communities of like-minded peers, the serendipitous exposure to opposing viewpoints, which the open web, for all its flaws, occasionally provided, disappears entirely. The risk is a thousand separate internets, each cryptographically sealed from the others.
What Survives in the Open
Not everything retreats. Three categories of public content remain viable.
Institutions with established trust. Wikipedia, government databases, academic journals, and verified news organizations retain value because their institutional credibility predates the AI content flood. Wikipedia's human moderation infrastructure, 40,000+ active volunteer editors, has so far outpaced bot contamination. The cost of maintaining this defense is rising, but the model works.
Content with provenance. C2PA (Coalition for Content Provenance and Authenticity), backed by Adobe, Microsoft, Google, and the BBC, is building technical standards for cryptographic content provenance. A photograph or article signed with C2PA metadata carries verifiable proof of its origin, creation date, and modification history. This does not prevent bots from creating content, but it allows humans to distinguish verified content from synthetic output.
AI-mediated consumption. Personal AI agents, the same technology that created the dark forest, may also filter it on behalf of the user. An agent that reads feeds, verifies sources, filters synthetic content, and surfaces only verified human-created material is a personal guide through the forest. The text-field-as-interface model described in The Text Field is the New Dashboard is the logical consumption layer for a public web that humans can no longer browse directly.
The Architecture of the Next Internet
The internet is splitting into two layers.
The public layer: Open, searchable, and increasingly synthetic. Dominated by AI-generated content, bot traffic, and algorithmically optimized spam. Still useful for commerce, logistics, and machine-to-machine communication. Increasingly useless for authentic human discourse and trustworthy information.
The private layer: Encrypted, verified, and human. Accessed through invitation, payment, or cryptographic proof of personhood. Where real conversations, genuine expertise, and authentic relationships concentrate. Invisible to search engines, crawlers, and the attention-refinery apparatus described in Attention Refinery.
The data shows the split is already underway. The remaining choice is whether the public builds infrastructure to keep a viable public commons, a space where verified humans can still exchange ideas openly, or accepts a full retreat into private clearings and leaves the open web to the machines.
C2PA, World ID, and personal filtering agents are the remaining bets on a public commons that still has humans in it. If those fail, the open web keeps the commerce and the machine-to-machine traffic. The conversations go behind an invitation, a payment, or an iris scan. Wikipedia's volunteer editors and a few provenance-signed feeds would be what is left in the open. Build the proof-of-personhood and provenance layer now, or accept that split as permanent.