What is Provable inference?

Provable inference is a set of techniques that let an AI provider prove, with math, that an answer really came from the model it claims to have used. The customer can check the proof without ever seeing the model's secret internals.

The problem it solves

When you send a prompt to an AI service, you receive only the text that comes back. You cannot see which model produced it. A provider could quietly serve a smaller, cheaper model while charging for a larger one, or switch models between requests, and you would have no direct way to tell. Provable inference aims to close that gap.

How it works in plain terms

The usual design has two steps:

1. Commit once. The provider publishes a short fingerprint, called a commitment, of its model weights. The weights themselves stay private. The fingerprint is like a sealed envelope: it fixes exactly which model is meant without revealing it.
2. Prove each answer. For a given request, the provider runs the model and also produces a proof that the committed model generated that output.

The customer then verifies the proof. Checking is designed to be much cheaper than redoing the work. Most proofs of this kind use a zero-knowledge proof, which convinces a checker that a statement is true without revealing the secret behind it.

Why it is hard

Running a large model is already expensive, and making it prove every step of its math adds a lot more work. Today's research systems are much slower than normal inference and often cover only narrow cases, such as a limited number of tokens or a simplified way of choosing the next word. Treat them as proofs of concept, not features you can switch on.

Why it matters

Trust in AI services currently rests on reputation and contracts. Provable inference points toward a future where a buyer, a regulator or a reseller's customer can ask for evidence. That matters for:

• Anyone paying for access to a specific model through a middleman.
• Regulated settings, where the exact model used must be recorded and defended.
• Providers that want to show a model really is the one it says without exposing it.

What it does not prove

A proof says which model produced an output. It does not say the output is correct, safe or free of hallucination. A verified answer from a model can still be wrong.

Related ideas

Provable inference is a close cousin of formal verification, which uses mathematical proof to show software meets a specification. Here the thing being proven is where an answer came from, not whether a program is bug free.