The internet stopped being a mostly human place sometime in 2024, and the crossover was not close. Imperva's 2026 Bad Bot Report puts automated traffic above 53 percent of all web activity, up from 51 percent the year before, with human traffic down to 47 percent and still falling. Imperva is unusually direct about what this means: not an attack cycle that will pass, but a structural change in what the network is.
That number is the visible edge of a harder problem. Every model trained from here forward draws on a corpus that is already mixed. Every image, every paragraph, every voice sample scraped after late 2022 carries some probability of having been generated rather than made. There is no test that reliably separates the two, and the ratio is moving in one direction only.
Which means the body of verifiably human text, images and audio produced before that date has quietly become something nobody planned for: a finite, non-renewable resource.
This exact shape of problem has occurred once before, and the way it was solved is stranger than you would expect.
On 21 June 1919, the interned German High Seas Fleet scuttled itself in the cold water off Orkney. Seventy four ships opened their seacocks at anchor rather than be handed to the British. Fifty two went down. For the next twenty years they were treated as a navigational hazard and a salvage opportunity, nothing more.
Then came July 1945, and the world changed in a way nobody thought to connect to a Scottish anchorage.
Steelmaking pulls in enormous volumes of atmospheric air. After Trinity, and through two decades of above-ground nuclear testing, that air carried trace cesium-137 and cobalt-60. Which meant that every ton of steel smelted anywhere on earth after 1945 carried a faint radioactive signature. Harmless to touch. Ruinous, however, if you were trying to build the one class of object that must not be radioactive: the instruments that measure radiation. Geiger counters. Whole-body scanners. Sensors bound for deep space.
So engineers went diving at Scapa Flow. The hulls resting on that seabed had been smelted from air that had never known a bomb. Divers cut plate from them for decades, and the industry gave the material a name that has since escaped into wider use.
Low background steel.
The point of the story is not nostalgia for pre-atomic air. It is this: the contamination was invisible, it was global within a few years, and it was irreversible by any ordinary means. The only remedy anyone found was to go and physically recover the uncontaminated material that happened to have been sealed off before the event.
The internet had its Trinity in late 2022.
What the contamination costs
It is tempting to treat the decline of the web as a matter of taste. Search results feel worse. Recipes have preambles. Everything has a certain smoothness to it. Taste is easy to dismiss, so start with the two things that cannot be.
The first is fraud. The FBI's Internet Crime Complaint Center logged roughly 893 million dollars in losses from complaints referencing AI in 2025, the first year it tracked the category separately. Synthetic voice, synthetic video, synthetic colleagues on synthetic video calls. The verification habits that everyone relied on for thirty years, recognising a face and recognising a voice, stopped working inside about eighteen months.
The second ought to worry the people building the models most of all. Train a generative model on the output of generative models, iterate, and the distribution degrades. Rare events disappear first, then variance collapses, then the whole thing converges on a bland attractor. The 2024 Nature paper on model collapse built theoretical intuition behind the phenomenon. The tailings, in other words, do not just sit there. Fed back into the furnace, they make the next batch worse.
That is where the metaphor stops being a metaphor. Somebody, somewhere, has to go diving.
Four responses
The projects that have emerged fall into four families. They are rarely discussed together, which is a shame, because they are visibly four answers to one question.
Salvage
Cloudflare runs LowBackgroundSteel.ai, an index of datasets and corpora captured before the contamination event, on the explicit reasoning that a verifiably pre-AI reference corpus will be worth more over time, not less. The artist and environmental engineer Tega Brain built Slop Evader, a browser extension whose entire mechanism is a date filter: it restricts searches on Reddit, YouTube and a handful of other sites to results published before 2022. Brain is clear that it is not a solution. It is a way of showing people what has been lost by letting them briefly stand in it.
Beneath both sits the older infrastructure that made this possible at all. Archive Team has been rescuing at-risk services since 2009, and the Internet Archive has been quietly holding the only copy of enormous amounts of the pre-2022 web. Neither was built for this purpose. Both turn out to be the salvage fleet.
Gates
If the network is majority machine, the second response is to build rooms that machines cannot enter. This requires a way to prove a visitor is a unique human without learning who they are, which is a genuinely hard cryptographic problem that a dozen organisations are now attacking from different directions.
World is the most aggressive. It verifies humans through in-person iris scans at physical devices called Orbs, and reports close to 18 million verified participants, up from around 12 million in mid-2025 and 6 million in September 2024. Its 2026 protocol upgrade moved to an account-based architecture with key rotation and recovery, and shipped integrations with Tinder, Zoom, Docusign, Okta and Vercel. When a dating app starts issuing verified-human badges, the idea has left the crypto conference circuit.
Others take different bets on the same problem. Humanity Protocol uses palm biometrics. BrightID and the original Proof of Humanity use social graphs and vouching. Human Passport, zkPass, Privado ID and Holonym work on privacy-preserving credentials that assert a property without revealing an identity. And there is a serious proposal to push the whole thing down the stack, using SIM and eSIM identity at the physical layer and propagating cryptographic attestations upward, so that telcos become infrastructure-level certifiers of human origin.
CAPTCHAs, meanwhile, are finished. Vision models solve them faster and more accurately than people do, and the solving services that once ran on human labour now run on models at a fraction of a cent per attempt. What remains of the CAPTCHA is a toll booth that stops only the humans.
Signatures
The third response gives up on detection entirely and moves the problem to origin. Instead of asking whether a piece of content is synthetic, sign it at the moment of capture and carry the signature with it.
This is C2PA and Content Credentials, now embedded in camera hardware and editing software, building a chain of custody rather than a verdict. Adjacent to it, Ben Goertzel's OpenWater project is building decentralised watermarking and provenance through SingularityNET and BGI Labs, with proof of humanity as a layer on top. Goertzel is refreshingly blunt about what it does not solve: no instrument proves absence of coercion, biometric systems exclude people and need appeals and recovery paths, certifiers collude, hardware gets compromised. The design goal is not perfection. It is ensuring that no single company or government becomes the sole arbiter of what counts as real.
Exodus
The fourth response is to leave.
In June 2019, a pseudonymous developer called Solderpunk published a specification for a document protocol sitting somewhere between Gopher and the web. He named it Gemini, after the NASA programme, and ran it on port 1965 in tribute. Its defining property is not simplicity but inextensibility. Gemini is deliberately built so that it cannot be extended, which means it cannot carry tracking, cannot carry targeted advertising, and cannot be commercially captured no matter who adopts it.
For years this was a curiosity with a few hundred sites. As of early 2026 there are at least six thousand capsules, almost all hand-written and hand-maintained. Gopher, the 1991 protocol Gemini descends from, had 411 active servers in January 2026 serving nearly six million selectors, sustained by nobody except the people who choose to sustain it. Around them sit Spartan, Nex, the finger protocol, the Marginalia search engine indexing the small and text-oriented web, and the broader IndieWeb and ActivityPub world. Further out again are the mesh projects, CJDNS and Hyperboria and Commotion, which answer a different question about physical independence but come from the same instinct.
None of this scales. None of it is meant to. What it demonstrates is that the machinery for a parallel, human-maintained internet already exists, is documented, and works. It is sitting there like tooling in a shed.
The part nobody has solved
Salvage is a hedge against a corpus that is already contaminated. Exodus is a subculture that has proved a point. The real fight is between gates and signatures, and it has an unresolved centre.
Every proof-of-humanity system needs a root of trust: someone who certifies the certifiers. Whoever holds that root holds the ability to decide, at internet scale, who is a person. That is a larger power than any platform has ever had, and handing it to a single company solves the bot problem by creating something considerably worse. The federated and zero-knowledge designs exist precisely because their authors understand this. Whether they can be made to work at the scale and speed the problem demands is, right now, an open question.
Which brings us back to Orkney.
The interesting epilogue to low background steel is that the shortage ended. Atmospheric testing was banned, fallout decayed, and by the 2010s ordinary new steel was clean enough for most instruments. The contamination turned out to be a phase, because its source was a specific human activity that specific humans agreed to stop.
There is no equivalent treaty here. The generative capacity is distributed, cheap, useful and permanent. Nobody is going to stop.
So the salvage matters more, not less. The archives being captured now, the provenance standards being fought over now, the small strange protocols being hand-maintained by a few thousand people right now, are the plate being cut from the hulls. Nobody at Scapa Flow in 1919 imagined what that steel would be for. The people doing this work do not entirely know either.
They only know that after a certain date, you cannot make it any more.