A personhood credential is a digital credential that lets someone prove to an online service that they're a real person, without revealing who they are. The term was popularized by a 2024 paper from researchers at OpenAI, Microsoft, Harvard, MIT, Oxford and other institutions, which argued that as AI gets better at passing as human, the internet needs a privacy-preserving way to tell real people from bots and fakes.
The problem they address
AI systems can now write, talk and act online in ways that are hard to tell apart from people, and they can do it at scale. The paper builds on two things AI still can't do:
- Convincingly pass as a person offline, for example in person or with real-world documents checked by a trusted issuer.
- Break modern cryptography, so a credential, once issued, can't be forged or copied.
Enrollment relies on the first; using the credential online relies on the second.
What makes a credential a personhood credential
Following the paper, a personhood credential should have two properties:
- Credential limits. Each issuer gives each eligible person at most one credential, so one person can't hold many.
- Unlinkable pseudonymity. A person uses a different pseudonym with each service, and services can't link or trace those pseudonyms back to the person or to each other.
The paper also suggests credentials could come from many kinds of issuer, including governments and other trusted institutions, and that they don't need to rely on biometrics.
How it differs from related ideas
| What it proves | What the service learns | |
|---|---|---|
| Personhood credential | A real, unique person | That fact, plus a pseudonym for this service |
| Identity verification (KYC) | Who the person is | Their identity |
| CAPTCHA | Probably not a simple bot, right now | Little, but it can't tell one person from many |
| Digital ID or wallet | Attributes the person chooses to share | Those attributes |
Personhood credentials are one form of proof of personhood. See also proof of personhood vs KYC.
Open questions
The idea raises real questions, which the paper itself discusses:
- Who issues them, and how much power that gives issuers.
- People without documents or access to an issuer, who could be shut out of services that require a credential.
- Selling or lending credentials: a person could let someone else, or an AI, use theirs. One credential per person limits the damage, since each person still only has one.
- Revocation and recovery when credentials are lost, stolen or misused.
How Human Pass relates
Human Pass works much like a personhood credential delivered as a sign-in. A person verifies a government ID once; Human Pass keeps only one-way fingerprints of the document, so one person can't hold two Human Passes; and each site gets a yes and an ID for that person that's different on every site, so sites can't link people. The difference from the paper's strongest designs: Human Pass is a trusted middle party that could, with its keys, link a person's IDs across sites, and our security overview explains how that capability is locked down.
Further reading
- Adler et al., Personhood credentials: Artificial intelligence and the value of privacy-preserving tools to distinguish who is real online, 2024.
- What is proof of personhood?