Proving Humanity

What it is

A Galaxy Digital position paper arguing that agent traffic should be gated on "a distinct person stands behind this agent, counted once" rather than "is this a machine". It evaluates no proof-of-personhood projects. Its value to an engineer is a design brief and a set of specific data points, not a product.

Details the summary leaves out

• Why behavioral checks fail structurally. An ETH Zurich 2024 study found an off-the-shelf object-detection model solved every reCAPTCHAv2 image challenge. reCAPTCHAv2 also appeared to lean on cookies and browsing history, so the puzzle was partly theater. The page frames ChatGPT Agent passing Turnstile (July 2025) as a single anecdote, not proof.
• What signed-agent schemes attest to. Cloudflare's "signed agents" category covers agents directed by end users, with the hosting platform signing requests. The signature names the platform, not the user. It cannot show that one person isn't behind a thousand agents.
• Why AP2 mandates stop at a key. A mandate proves an account holder's key signed an instruction. It does not show the holder is a person or differs from other holders.
• Platform records don't close the gap. They are private and after the fact, they tie actions to an account rather than a person, and they vanish when an agent runs on the owner's own hardware.
• Real-world signals.
• Amazon blocked Meta's Muse agent, in part because it didn't disclose it was an AI.
• Imperva put automated traffic at over 53% of web traffic in 2025.
• Cloudflare Radar showed automated requests at 57.5% of HTML traffic, which the page says passed human traffic in June 2026.
• New York's AG found about 18M of 22M FCC net-neutrality comments fake. They were caught through bulk submission and recycled text, which agents can avoid.
• A design rule. Per-account limits can be multiplied away, while per-person limits cannot. This is the practical reason to anchor rate limits and quotas to a person.

Why KYC-everywhere is rejected

1. It multiplies exposure: every agent provider holds KYC data, including open-source agents with no company to protect it.
2. It excludes people without acceptable documents, for activity that never needed KYC.
3. It doesn't solve counting, because providers verify in isolation and can't tell whether two verified people are the same person. Pooling the records would concentrate the risk.

The proposed alternative is to verify once and reuse the credential across agents and services. Delegation should be limited and revocable, without handing over the credential itself.

Caveats

• Personhood must be established once, and the issuer has to be trusted to do it honestly and retain nothing extra.
• Credentials can be rented or sold, which weakens the one-credential-one-independent-person guarantee.
• The page notes that the ad-fraud figures (Juniper, $84B in 2023) come from vendors selling defenses and are only indicative.
• The METR-style doubling claim (task length roughly every seven months) is called narrow and possibly not generalizable.
• Proof of personhood bounds how many parties can act. It says nothing about truthfulness or alignment.
• The page does not say when a standard will exist. It argues only that the window to build one is while machines can still sometimes be caught.