The scarce resource of the next internet is not attention, compute, or even trust in general. It is proof that there is a real, unique person behind an account. Once a bot can pass for a human in text, image, and voice — which for text it already does — the question "is this a real human?" stops being answerable by looking. It has to be established by protocol. My forecast, labeled plainly as a bet: verified humanity becomes a valuable, contested layer of the internet within a few years, and the design of that layer is one of the more consequential and least-discussed problems ahead. Because every system that can prove you are human is also a system that can watch you and lock you out.
Start with why inspection fails, precisely, because the precision is what makes this a protocol problem rather than a detection problem.
The moment inspection stops working
For most of the internet's life, "is this a person?" was answered by inspection. You read the comment, looked at the photo, heard the voice, and your own perception did the verification. It was cheap and it was usually right, because faking a human across those channels was expensive. A convincing fake profile took real human labor per account, which set a natural price floor on deception. Spam was crude because good fakes didn't scale.
That price floor is gone. A model now writes a comment indistinguishable from a thoughtful person's, generates a face that has never existed, and holds a voice conversation that passes for human on a phone call. The cost of a convincing fake human — per account, per hour — is collapsing toward the cost of inference. When the marginal cost of a plausible person approaches zero, inspection cannot separate the real from the synthetic, because inspection was only ever measuring the cost of the fake.
This is the mechanism behind what people loosely call "dead internet theory" — the idea, overstated as a description of today, that much of the web is already bots talking to bots. As a literal account of the present it's wrong; as a forecast of where the cost curves point, it names something real. If a synthetic participant is free and a real one is not, the synthetic ones dominate by sheer volume unless something external distinguishes them. Perception can't be that something anymore. Which leaves protocol: an out-of-band procedure that establishes personhood by evidence a bot can't cheaply produce.
Note what changed. We are not asking machines to be less capable so we can catch them — that's the CAPTCHA strategy, and it lost. We are asking for a positive credential of humanity, attached to the account, that survived some costly, hard-to-fake process. There are four families of such credentials, and each buys verification with a different currency.
Four ways to prove a human, and what each one costs
Verified identity (the KYC approach). Bind the account to a government-issued identity — passport, national ID, a bank that already did the check. This works. It is the most robust of the four because it inherits the state's investment in making identities hard to forge, and it's already deployed wherever "know your customer" law applies. The cost is paid in two currencies. First, pseudonymity dies: to prove you're human you reveal who you are, and the dissident, the whistleblower, and the person in an abusive situation lose the ability to speak without being named. Second, it centralizes power. Whoever operates the verification becomes a chokepoint that can deplatform, surveil, or be compelled by a government to do either. You've solved "is this a human" by making every human legible to an authority. For some contexts — banking, voting — that's acceptable. For the open internet, it's a different regime than the one we have.
Proof of personhood (uniqueness without identity). The clever ambition here is to prove two things while revealing a third: prove you are human and that you are unique — one person, one credential — without revealing who you are. Projects in this space, most visibly Worldcoin with its iris-scanning orbs, aim exactly at this: a token that says "one distinct person" and nothing more, ideally with zero-knowledge cryptography so the proof reveals no identity downstream. Done cleanly, it would defang most abuse, because uniqueness is what most abuse exploits — bot armies, sockpuppet rings, and vote manipulation are all one entity pretending to be many. Deny that and a huge class of attacks collapses while pseudonymity survives. The problem is that uniqueness is precisely the hard part. To guarantee one-person-one-credential you need something unique to the body, which pushes toward biometrics, which creates a permanent global registry of who exists and a biometric root that can never be revoked if it leaks. The cryptography can hide identity in daily use and still leave that enrollment root as a single, catastrophic point of failure and coercion. Promising, genuinely — and carrying a privacy risk shaped exactly like the problem it solves.
Reputation and web-of-trust. Skip credentials; accumulate a track record that is costly to fake. An account that has posted for years, is vouched for by other trusted accounts, and has staked something it would lose by misbehaving carries evidence no fresh bot can instantly manufacture. This is how trust worked before institutions — you trusted who your trusted contacts trusted — and it degrades gracefully, with no central authority and no biometric. Its weakness is symmetric to its strength: reputation takes time to build, so it can't verify a newcomer, and a patient adversary can farm aged, vouched accounts, which is already a black market. It raises the cost of fakery without setting a hard floor.
Costly-signal gating. Make participation expensive enough that running a million fake humans stops paying. A refundable deposit, a small stake, a fee, a scarce invite. This is the honest core of what proof-of-work and stake-based systems do: not prove humanity directly, but price out humanity's counterfeit at scale. It's simple and needs no identity, but it taxes the poor and the legitimate alongside the attacker, and a well-funded adversary just pays.
| Approach | Proves | Preserves pseudonymity? | Main cost |
|---|---|---|---|
| Verified identity (KYC) | Who you are | No | Centralization, surveillance |
| Proof of personhood | One unique human | Partially (with ZK) | Biometric root, exclusion |
| Reputation / web-of-trust | A costly track record | Yes | Slow; farmable over time |
| Costly-signal gating | Willingness to pay | Yes | Taxes the honest and the poor |
None dominates. Each trades verification strength against some combination of privacy, inclusion, and centralization. That tradeoff is not an engineering detail to be optimized away later; it is the actual product decision.
Why this is the trust primitive of the agent era
Proof of human is not a content-moderation footnote. It's the foundational primitive underneath the whole coming economy of software agents, which is why it connects directly to the coming agent trust crisis. When agents act on your behalf — booking, negotiating, transacting, posting — every counterparty faces a new question before it engages: is there a human principal behind this agent, and is this the only agent that human authorized, or one of ten thousand? Agent-to-agent commerce cannot assume the other side is honest or even human. It needs a verifiable answer, and proof of personhood on the human root is where that answer bottoms out. An agent is trustworthy in part because a real, accountable, non-duplicated person stands behind it.
There's a deeper, epistemic layer too. Even setting agents aside, the flood of synthetic testimony breaks how we know things from what others tell us — the problem I worked through in testimony at scale. We form most of our beliefs by trusting sources, and that machinery quietly assumed the sources were people, or institutions staffed by people. When a source's humanity is unknown, the reasonable prior on any unverified claim drops, and proof of human becomes an epistemic filter, not just a security one — a way to know whose testimony to weight. That is a large shift in how trust flows across a network, and it's why verified humanity is contested: whoever controls the credential controls whose voice counts.
The dystopian edge is not a side effect
Here is the counterargument, in its most uncomfortable form rather than defused. Every proof-of-human system is also a surveillance system and an exclusion system, and the cure can be worse than the disease.
Surveillance, because a credential that proves you're human is, by construction, a persistent handle on you. Even a privacy-preserving one has an enrollment moment — a scan, a document, a vouching — where a body meets a database, and that root can be leaked, sold, or compelled. Build the infrastructure to verify humanity at internet scale and you have built the infrastructure to track it at internet scale; the same rails carry both, and which one runs is a policy choice made by whoever holds the root, not a property of the math.
Exclusion, because verification always has a population that fails it: no document, no supported device, a biometric that doesn't match, or a principled refusal to be enrolled. Those people don't stop existing — they get shunted to an unverified underclass whose speech is discounted and whose access is throttled. Identity infrastructure has historically fallen hardest on exactly the people with the least power to contest it. A two-tier internet — verified-human and everyone-else — is a plausible and grim output of solving this problem badly. And "badly" is the default, because the parties most eager to build the verification layer, large platforms and states, are precisely those who benefit from centralization and legibility.
That outcome isn't avoidable by wishing the whole project away. The demand for proof of human is real and rising; something will be built to meet it. The question is whether it's built by people who treat the privacy and exclusion costs as first-order, or by people who treat them as someone else's problem.
What to build, and what to watch
For builders, the shift is small to state and large to implement: treat "is this a human?" as a first-class, protocol-level question in your system, not an assumption baked into your UI. Decide explicitly what you need to know — humanity, uniqueness, or identity, which are three different things — and buy only the weakest one that solves your actual problem. If you need to stop bot armies, you need uniqueness, not identity; don't collect a passport to solve a sockpuppet problem. Weigh the privacy cost of every answer, because each bit of verification you demand is a bit of surveillance you're now responsible for, and a set of people you're now excluding. Prefer schemes that reveal less: zero-knowledge uniqueness over identity dumps, refundable stakes over collected documents, multiple independent verification paths over a single mandatory one.
For everyone else, the thing to watch is which family wins in which context — because it won't be one system. Watch whether the open internet gets a pseudonymous proof-of-personhood layer or collapses into KYC-for-everything by default, which is the lazy equilibrium. Watch who ends up holding the enrollment roots. Watch for the two-tier internet appearing not by decree but by a thousand platforms quietly deciding unverified accounts don't count.
The internet was built on the cheap, usually-correct assumption that the thing you were talking to was a person. That assumption is now false, and rebuilding it as an explicit, verifiable, deliberately-designed layer is unavoidable. Whether that layer protects people or catalogs them is still open — which is exactly why it deserves attention now, before the defaults harden into something we didn't choose.