AI Identity & Content Provenance

The problem

AI's generative capability has broken the epistemic foundation of digital communication: content that looks real may not be, and a person who appears to be themselves may in fact be a simulation. Ben Horowitz frames this as the primary new problem class AI creates, and argues the solution has to be cryptographic, with blockchain as the trust anchor.

The personalized-communication collapse is the broadest version of the problem: AI now lets anyone generate a phone call, email, or video call that is deeply personalized and indistinguishable from the real person it imitates. In Horowitz's words: "All communication is going to be completely unusable."¹ The old human tells that once flagged a scam, like an oddly formal salutation in a phishing email, no longer work once AI removes the errors that gave a fake away.

A sharper version is the deepfake executive problem: an AI-generated version of a chief executive appearing on a video call and instructing a finance team to wire a large sum overseas. This is not hypothetical; deepfake-driven business email compromise and executive-impersonation fraud have already produced losses in the hundreds of millions of dollars. Horowitz's own stated response is blunt: everything becomes hardware-rooted access, and nothing from him should be believed unless it carries his cryptographic signature.

A more ambient version is content that spreads without anyone realizing it is fabricated: AI-generated videos are already being forwarded by people who do not know they are fake, including within Horowitz's own family. Detection tools can currently catch a meaningful share of AI-generated content, but detection is expected to eventually fail as generation improves, at which point the only durable way to trust content is cryptographic signing rather than pattern-matching.

A three-layer framework

Horowitz separates the problem into three distinct layers, each requiring its own cryptographic solution. The first layer is proving humanity itself: the most basic gate for social media, dating apps, and video calls, where the question is not whether an account is valid, since bots can hold valid accounts, but whether biological existence can actually be proven. This connects to a broader argument that as AI erases the natural separation between humans and machines, proof-of-personhood restores a form of identity-layer scarcity that cheap AI has otherwise made impossible to recover.²

The second layer is proving identity beyond mere humanity: whether a specific claimed person can actually be verified as themselves, using cryptographic key pairs as identity anchors, where a private key functions as the identity and the public key as its verifiable proof.

The third layer is proving authorship of content itself: whether a specific piece of content was actually created or approved by its claimed author, using cryptographic signing analogous to existing content-provenance standards. A signed video of a public official giving a speech would carry a verifiable attestation that the official, or an entity they control, approved it; an unsigned video would carry no such guarantee.

Blockchain as the trust anchor

The central architectural question is who should be trusted to maintain an identity registry and a content-provenance ledger. Horowitz explicitly rejects the obvious institutional candidates: platform registries carry commercial and political incentives that may not align with verifying identity accurately, and government registries are controlled by states that can suppress or manipulate records. His preferred alternative rests on the mathematical and game-theoretic properties of a blockchain, verifiable by any party without having to trust the registrar itself, which connects the AI identity problem to a broader case for public blockchains as neutral, verifiable public infrastructure more generally.

A further, more mundane instantiation of the identity layer is government payments: the stimulus-era experience revealed that the US government is structurally bad at getting money to people, in part because there is no universal address book for citizens, and Horowitz's proposed fix is that everyone needs an address where money can be sent, which he treats as fundamentally a crypto problem. A crypto wallet address becomes the universal identifier: once identity is proven, anyone can receive payments at that address with no bank account and no intermediary able to intercept the transfer.

AI agents as economic actors

A third, distinct application concerns AI agents that need to transact as economic participants in their own right rather than as extensions of a human. Horowitz frames the practical obstacle directly: a credit card merchant account requires a legal entity, identity verification, a bank relationship, and compliance with card-network rules, none of which apply cleanly to an AI agent, leading him to conclude that agents need a bare payment instrument native to the internet itself. CZ reaches the same conclusion from the infrastructure side, citing the practical impossibility of applying existing identity-verification and volume assumptions to agent-scale transactions, arriving at the same destination from business logic that crypto is currently the only viable payment infrastructure for AI agents.

Independent corroboration

Reid Hoffman arrives at closely related conclusions from a different angle, drawing on social-platform experience and agent-economy reasoning rather than corporate security threat modeling.³ Deepfakes are his own motivating concern: having deployed his own AI-generated speaking avatar at conferences, his stated worry is whether a given video of someone is actually real, and he describes crypto as certification authority done in a genuinely deep way, the same blockchain-as-trust-anchor argument Horowitz makes from a different starting point. He traces his own interest to buying Bitcoin in 2014 on the reasoning that this is how identity infrastructure, and not only money, should eventually work on the internet. His own platform's history offers a natural experiment in favor of the underlying claim: a real-identity requirement produces noticeably less misinformation than pseudonymous platforms tend to, which he summarizes as a robust sense of identity leading to more trustable information. And he extends the argument specifically to agent-to-agent trust, asking what happens when one person's agent is transacting with another's, whether that exchange is trustable, and whether either side actually has verifiable identity, an extension he answers with onchain identity tokens.

The two accounts do not conflict; they triangulate from different starting points, corporate security in Horowitz's case and long-run social-platform structure in Hoffman's, toward the same conclusion, that blockchain functions as the answer, with the live question being which specific chain and token standard wins out.

Tensions and open questions

Proving humanity is the hardest of the three layers: biometrics can be spoofed, hardware attestation requires trusting the device itself, and any centralized registry becomes an attractive target in its own right, with no deployed solution yet achieving coverage, privacy, and security all at once. The blockchain-as-trust-anchor argument also assumes the blockchain itself resists capture, and a governance takeover of a proof-of-stake chain would undermine the exact mathematical guarantee the argument depends on. Signed content requires signing infrastructure to be adopted by creators before distribution even happens, which is itself a coordination problem, since deepfake videos will obviously not be labeled as deepfakes by whoever creates them. And the proposal that everyone needs a crypto address for government payments assumes those addresses are stable, non-losable, and recoverable, all of which remain unsolved user-experience problems for people without technical backgrounds.