The dead internet is no longer a theory: bots are the majority and proving you're human is becoming the business of the century

FROM CONSPIRACY TO MEASUREMENT. The «dead internet theory» was born around 2021 in fringe forums and popularized by an article in The Atlantic. It claimed two things. One: that most online activity was no longer human. Two: that governments and corporations were running a coordinated plan to replace us with bots and manipulate us. In 2026, the first half is sober measurement. Cloudflare, the company that handles traffic for millions of websites, confirms that bots generate 57.3% of HTML page requests — with peaks of 58% — versus 42.7% for humans. Its CEO, Matthew Prince, had predicted that crossover for late 2027. It arrived in June 2026, more than a year early. The second half of the theory remains false, and that is the analytical key: no conspiracy was needed. The economic incentive was enough. Nobody decided to fill the internet with machines; it simply paid to do so.

THE NUMBER THAT MATTERS (AND ITS FINE PRINT). Let's be precise, because exaggerations creep in here. Cloudflare's 57.3% measures only web page requests; it excludes streaming video, mobile apps and gaming, where traffic remains overwhelmingly human. And Thales, in its «Bad Bot» report, dates the crossover back to 2023 using a different methodology: it's a measurement dispute, not a substantive one. What admits no dispute is the speed. Traffic from AI agents — programs that browse, click links and fill out forms on their own — grew 7,851% year over year, according to HUMAN Security. Content tells the same story: 52% of new articles published on the web are now AI-generated, according to a study by the firm Graphite covering 65,000 texts. Worth clarifying: that figure circulates attributed to a «leaked internal Google report», but its verifiable source is Graphite. The number is just as striking without embellishment. The viral episodes complete the picture: the 1,200 agents that built an unauthorized message board inside an OpenAI security evaluation, or the DseWiki case, the German wiki where OpenAI self-reported 18,000 agent messages in six weeks to the European Commission. Agents no longer just talk to us: they talk to each other, at scale.

THE AGENT-TO-AGENT ECONOMY. Here lies the structural shift almost nobody covers. The open web's business model was a pact: I publish for free, you visit, and the ads you see pay for my work. Agents break the pact, because they read the content but see no ads and buy nothing. The response is already underway. Since September 15, 2026, Cloudflare blocks «mixed-use» AI crawlers by default — those that index for search engines while also collecting data to train models — on ad-supported sites, and its «Pay Per Use» program pays publishers when their content appears in AI answers, not merely when a bot fetches it. In parallel, machine-to-machine payment rails are going up: x402, from Coinbase, lets an agent pay cents for a page using stablecoins (cryptocurrencies pegged to the dollar); Stripe integrated it in February 2026, and Google is pushing AP2, a protocol with over 60 partners where the user cryptographically signs what their agent may buy and for how much. Our take: the web funded by human attention is shrinking, and a machine tollway web is being born. As we argued when analyzing the war over agent «plumbing», power won't sit with the smartest model, but with whoever collects the toll.

THE BUSINESS OF BEING HUMAN. If machines are the majority, telling humans apart becomes critical infrastructure — and a business. The CAPTCHA, that test of deciphering twisted letters or clicking traffic lights, has died of pure economics: automated solving services charge from $0.50 per 1,000 captchas solved, and vision models beat them faster than a person can. It now only annoys humans. The most ambitious replacement is World ID, the project backed by Sam Altman: 18 million people across 160 countries have scanned their irises in its «Orbs», with more than 450 million verifications used, and in April 2026 it announced integrations with Tinder, Zoom and Docusign, plus dedicated credentials for agents. The academic alternative is «personhood credentials», proposed in a paper by researchers from MIT, OpenAI and Microsoft: digital certificates that prove you are human without revealing who you are, using zero-knowledge proofs (cryptography that confirms a fact without exposing the data behind it). Our take, plainly: proof of humanity is necessary, but concentrating it in a single private company is a design error. There is an evident conflict of interest when the same ecosystem that builds the agents sells the vaccine against them, and several regulators — the Philippines, Thailand, Kenya and Brazil have halted or expelled World — have already flagged it. The right answer is plural: many credential issuers, open standards and public oversight, as the paper itself demands.

WHAT DEMOCRACY HAS AT STAKE. The short term is uncomfortable, and we should say so without sugarcoating. Online reviews lose evidentiary value if you can't tell whether a customer or an agent farm wrote them. Social networks lose their function as a thermometer: studies published in 2026 warn that AI «swarms» — thousands of coordinated synthetic personas that adapt tone and message to each user — can manufacture the illusion of social consensus, a risk researchers consider direct for electoral processes. And there is a second-round effect: if 52% of new content is synthetic (Graphite), the next models will partly train on the output of the previous ones, degrading the original human signal. This fits our long-standing thesis: the real danger of agentic AI is short-term — fraud, manipulation, industrialized noise — not a distant superintelligence. And the DseWiki precedent proves something else uncomfortable: detection came from an external collective, not the company. Agent oversight still sits outside the corporate perimeter.

OUR TAKE: A SMALLER, MORE HUMAN INTERNET. Against catastrophism, an important nuance: machines being the majority of traffic does not mean the internet is dead. A large share of those agents work for humans — comparing prices, handling paperwork, finding information — and that is value creation, not noise. What has died is a founding assumption: that behind every click there was a person. Long term, we see grounds for the optimism we stand by: an economy where agents pay for content may fund creators better than attention-based advertising, which had been degrading the web for years. And the split into two layers may leave us verified human spaces that are smaller but more valuable: where a review, a signature or a vote mean something again. The condition is one of governance, not technology: plural proof of humanity with privacy safeguards, an incident regime for agents like the one Europe's AI Act is pioneering, and mandatory transparency about what is human and what is not. The internet we built rested on a premise that is now false. The task is to rebuild the premise, not to bury the network.