Definition
Zero-knowledge proof is a zero-knowledge proof is a cryptographic method that lets one party prove a statement is true without revealing the secret behind it, such as proving a computation was done correctly.
A zero-knowledge proof lets you convince someone that a claim is true without handing over the information that makes it true. The person checking learns that the claim holds and nothing else.
The classic picture is a locked door. You prove you hold the key by walking through the door, without ever showing the key to the person watching.
Three properties define a zero-knowledge proof:
Modern systems turn a computation into a mathematical statement, then produce a short proof that the computation ran correctly on some hidden input. Many of these proofs are called succinct, which means checking one takes far less work than redoing the computation. Producing the proof is usually the expensive part, often much slower than running the original program.
AI providers keep their model weights private, and customers usually see only the text that comes back. That leaves a trust gap: you cannot tell from the output alone whether the model you paid for produced it.
Researchers are exploring zero-knowledge proofs as a way to close that gap. A provider could publish a short commitment to its weights once, then attach a proof to a response showing that the committed model produced it, without exposing the weights themselves. The same idea can prove that a hash, a signature or a private calculation was done correctly.
The cost is the catch. Proofs for large neural networks are slow and heavy today, so current work tends to cover narrow cases rather than everyday chat traffic.
Imagine a bank that must show a regulator its customers' balances stay above a limit without sharing any account. A zero-knowledge proof can show the rule holds across all accounts while keeping every balance private.
A zero-knowledge proof is one tool for checking inference results. It sits beside other trust tools such as C2PA, which records where a file came from, and AI watermarking, which hides a signal in generated content. Each answers a different question: who made this, how was it made, and was it computed honestly.
It is a way to prove something is true without revealing the secret behind it, like proving you know a password without saying it.
Producing a proof usually takes far more work than running the original computation, even though checking the proof is fast.
Anthropic and university researchers released FlashProve, an open-source system that cryptographically proves a 400B-parameter model produced a given token, at about 72 times the time of the prefill. It's slow and narrow, and it's a working answer to "did they swap my model?"