This article was compiled and organized by BlockWeeks
Over the past year, the way I use the internet has changed noticeably. I open Google search less and less, and my entry point for browsing the web is gradually shifting to AI agents: agents buy dog food for me, remind me when I receive important emails, track package status, and even tell me "delivered" before the package system in my building notifies me. It also summarizes markets and news for me every morning, so I don't have to open X or media websites one by one.
As a result, the frequency with which I personally visit web pages, online stores, and my inbox has dropped sharply; the one actually "going online" is my agent. Something similar is happening to more and more friends, whether or not they are crypto-native. People are reducing their direct manipulation of the internet, and in scenarios that previously required three real people to click on web pages, there may now be hundreds of agents scattered about, each performing specialized tasks.
How will websites and the internet as a whole adapt to this migration? For most of the internet's history, bots were something websites tried to keep out: they usually appeared at the cost of consuming site resources, behaved rigidly, and were therefore easy to identify. Agents complicate both ends of this equation. More and more agent traffic comes from "people acting through software," so blocking agents means blocking these people; at the same time, agents behave more and more like humans, making detection harder, and making it easier for malicious actors to disguise themselves as large numbers of independent individuals.
Therefore, the main purpose of proof of humanity is no longer "keeping machines out," but ensuring that every machine can be traced back to a real person.
From "whether it is a machine" to "whether there is a unique real person behind it"
The internet has long assumed by default that there is a person on the other side of the screen, yet it has never formally verified this. Limitations in machine capabilities made this default acceptable: automated programs were either credible enough or cheap enough, but rarely both. A program that can act like a human is costly in both technical and financial terms, so such programs are rare; cheap automation can appear in large numbers, but it exposes itself through speed, repetition, and inability to handle unexpected situations.
What most people understand as a "bot" is still a script: repeatedly posting the same message, refreshing a page at fixed intervals. No matter how complex the branches, a script can only execute actions its author envisioned in advance; decisions are made beforehand, and rigidity makes it easier to detect. Generative AI first weakened the limits at the content and communication levels. In a 2025 Turing test, when GPT-4.5 was prompted to adopt a human-like persona, reviewers selected it as human 73% of the time, higher than the proportion who selected real human participants. A 2024 meta-analysis also found that people's ability to identify deepfakes was only slightly above random guessing.
Agents extend this breakthrough to the behavioral level. Scripts follow instructions; agents pursue goals: they are told what to accomplish, not how to accomplish it, and they decide the steps themselves. They can read pages they have never seen, decide where to click, and adjust when they fail until the task is complete. The real difference lies in when and by whom decisions are made—a script's decisions are fixed in advance by its author, while an agent makes decisions in real time based on the world it faces at that moment. This responsiveness was precisely what used to be regarded as evidence of "humanity." In on-chain scenarios such as DeFi, agents execute transactions and manage assets on behalf of users, which also makes the question of "who is acting" more prominent.
Blocking agents means blocking the people behind them
When agents become the primary interface, the difficult question websites face is no longer "is this a machine," but "is there a real person behind it, and is this person counted only once." Agents are beginning to identify themselves, and early attempts to associate agents with verified individuals are also appearing, but there is not yet a widely adopted standard that can establish "someone stands behind the agent" or ensure that "one person has not disguised themselves as thousands."
This directly challenges the security paradigm of the past, which centered on machine detection. CAPTCHA verifies that the operator is human by looking for gaps between machine and human capabilities; technical means such as rate limiting, timing patterns, and fingerprinting are used to identify and block unwanted or malicious machine traffic. These measures work because designing human-like machines at scale is expensive. The agent era is shaking this assumption: agents do not merely imitate human behavior; they can act legitimately in a person's name, and they increasingly resemble humans. Blocking one may mean blocking the very person it serves.
Proof of humanity is not equal to real-name identity: proving "uniqueness," not exposing "who"
Proof of humanity does not need to reveal a person's name to prove that they truly exist. At the moment an action occurs, the counterparty needs assurance that "a distinct person stands behind the agent," not identity documents. A name matters only when something goes wrong, and only to institutions with a legitimate claim to it.
This imposes a subtle requirement on identity systems: they must be able to confirm that someone is a unique and real human without exposing identity, and prevent the same person from infinitely replicating "personhood." If this cannot be done, many signals on the internet will lose their meaning. The key is not to recommend a specific project, but to understand why the problem has become urgent.
Two types of AI safety problems: alignment and institutional safety
"AI safety" covers two different types of problems, and proof of humanity is related to only one of them.
The first type is a technical problem: building AI systems that reliably carry out their designers' intentions, that is, "alignment." It involves models pursuing the wrong goals, deceiving supervisors, or providing substantive help to people building weapons. These are problems inside the system, and proof of humanity cannot solve them. An unaligned model does not need to pretend to be human to cause harm; dangerous capabilities do not become non-dangerous just because the requester is a verified real person. Alignment attracts most of the attention in the field and is at the core of public fear.
The second type of problem concerns the impact of AI capabilities on the surrounding world, especially on institutions. Much of social and economic life uses "people" as the unit of counting: a vote matters because it represents one person; a petition is effective because of the number of signatures; welfare or free trials are usually once per person; comments, trending topics, and public opinion are essentially counted by number of people. These signals are valuable because people are scarce.
Agents weaken this scarcity. When agents can write, speak, and act like humans, and can be replicated indefinitely, one person becomes something that can be produced cheaply. The effect is similar to counterfeiting, although the analogy is not perfect, because voices are not fully interchangeable like currency; what the two share is a value that depends on scarcity. Once counterfeiting becomes cheap, institutions that rely on counting people will become distorted.
The internet needs a new "personhood layer"
As agents become the internet's primary interface, the network needs a mechanism that can prove "a distinct person stands behind the agent" without revealing who that person is. This mechanism is not just an anti-bot tool, but infrastructure for future online identity and institutional trust. The question it must answer is not "is this a machine," but "is this a real, unique person who can be counted once."
This need is becoming urgent. The growth of agent traffic, the spread of human-like behavior, and the falling cost of identity replication—these three combined will gradually render existing verification, counting, and trust mechanisms ineffective. The internet used to assume by default that "someone is present"; in the future, it must prove that "there really is a person present," and that this person can be counted only once.
