As AI systems become harder to distinguish from people online, a group of researchers is arguing that the internet may need a new kind of proof: a way to show you are human without handing over your identity.
The proposed answer is called Personhood Credentials, or PHCs. The concept is designed for a future in which AI bots, impersonation and automated agents make ordinary online trust much harder to maintain.
Why human proof is becoming harder
The immediate problem is not that every suspicious account is a machine. It is that people now have to consider that possibility seriously.
The source article points to a recent example on X: a "Mr. Strawberry" account that has been stirring debate in the AI community by allegedly leaking information about OpenAI's latest AI technology, codenamed "Strawberry". Some observers suspect the account could be an AI bot controlling other bots, including voice output in audio chats.
The article also notes that a human is more likely behind the account. But the deeper issue remains: if a bot is plausible enough to be part of the conversation, then the old signals people used to judge online identity are weakening.
That is the background for work by researchers from OpenAI, Microsoft, MIT and other tech companies and universities. They have developed a method intended to let people anonymously prove that they are real humans online, with the goal of limiting the impact of AI bots and identity theft.
What Personhood Credentials would prove
Personhood Credentials are digital credentials that confirm one narrow fact: the holder is a real person. They are not meant to reveal the holder's broader identity information.
That distinction matters. Many online systems already ask users to prove something about themselves, but the researchers' proposal focuses on separating proof of humanity from disclosure of personal identity. In plain terms, a PHC would try to answer the question "is this a person?" without also answering "which person is this?"
The proposal rests on two limits that the researchers identify in current AI systems. First, AI systems cannot yet convincingly mimic humans offline. Second, according to the researchers, they cannot yet bypass modern cryptographic systems.
Those limits create an opening. If a trusted issuer can verify personhood outside the normal online interaction, and if cryptography can preserve privacy when that credential is used, then services may be able to distinguish humans from advanced AI systems without building a full identity-checking layer into every interaction.
The core design problem
The researchers outline two main requirements for a PHC system. An issuer should "maintain a one-per-person per-issuer credential limit" and should also have ways to reduce the risk of credential transfer or theft.
That first requirement is central to the fraud problem. If one person could collect unlimited credentials from the same issuer, the credential would lose much of its value as proof that each account or action is tied to a separate human. But the second requirement is just as important: if credentials can be sold, shared or stolen easily, the system could be abused even if issuance begins carefully.
At the same time, PHCs are intended to support anonymous interaction with services. The researchers describe service-specific pseudonyms, digital activities that cannot be traced by the issuer, and activity that service providers cannot link across contexts.
That combination is the hard part. A PHC system has to limit mass abuse without turning every online action into an identity record. It must be strong enough to matter against fraud, but restrained enough to protect privacy and civil liberties.
Who might issue PHCs
The researchers suggest that possible issuers could include countries that offer a PHC to every tax ID holder. They also mention other trusted institutions, such as foundations.
The proposal does not depend on a single issuer. In fact, the authors argue for multiple issuers, each limiting the number of credentials a person can receive from that issuer.
This would let people hold a limited total number of credentials. The goal is to provide more than one credential for privacy reasons, while avoiding so many credentials that the system stops helping with fraud prevention at scale.
That balance is important because the same design choice can push the system in opposite directions. Fewer credentials may strengthen anti-fraud controls, but could make privacy harder. More credentials may improve privacy, but could weaken the system's ability to limit coordinated abuse.
The bigger question for the internet
The researchers frame proof-of-personhood systems such as PHCs as a specific countermeasure against AI impersonation. They recommend prioritizing the design, testing and adoption of these systems.
OpenAI researcher Steven Adler, who contributed to the paper, connects the issue to a prediction from Meta CEO Mark Zuckerberg about "hundreds of millions or billions of AI agents using the Internet on people’s behalf". Adler's question is direct: "What happens if he’s correct?"
If that kind of agent activity becomes common, online services may face a basic classification problem. They will need to decide when it matters that an action came from a human, when automated help is acceptable, and when anonymity should still be preserved.
PHCs are one proposed answer to that pressure. They are not presented as a way to reveal who everyone is. Instead, the idea is to create a privacy-preserving signal of personhood that can survive a more automated web.
The tradeoff is clear from the proposal itself. The internet may need better defenses against AI bots and identity theft, but those defenses cannot simply replace anonymity with constant identification. Personhood Credentials aim to sit between those extremes: enough proof to separate people from machines, and enough privacy to avoid turning proof into surveillance.