AI Hallucination & Blockchain

Try out my Fraud Proofing LLM for yourself here:

verifiable-claude.vercel.app

Verifiable Claude - Fraud Proofs for LLMs

Cryptographically verifiable AI responses. Every claim is committed to a Merkle tree and verified against web sources — making hallucinations provably detectable.

Recently I asked Gemini for a synopsis on a book and it gave me completely wrong information.

I told it it was wrong, it apologized. Then proceeded to give me different but still totally wrong information. I told it "only tell me information you can verify is fact, do not tell me something if you don't have access to the contents of the book itself, just tell me you don't know". And it STILL confidently gave me totally wrong information.

And so I reached the jagged frontier of AI, where the magic is less sparkle and more dog poop.

In blockchain, particularly in my world of Ethereum L2s, we have this thing called fraud proofs. If someone lies about the state of the chain, there's a system of checks to verify and punish the liar.

I talk about this in a video with visuals here.

So what about fraud proofs for AI?

LLMs hallucinate with confidence. Blockchain systems fail explicitly.

In Arbitrum (a blockchain):

• If a sequencer makes an invalid state transition → fraud proof catches it

• If data is unavailable → the system halts, doesn't guess

• If verification fails → transaction reverts, doesn't pretend it worked

In AI agents:

• If the model doesn't know → it makes nonsense up confidently

• If retrieval fails → it fills in plausible-sounding baloney

• If verification would help → there's no mechanism to verify

I decided to borrow from blockchain architecture and build fraud proofs for AI.

I'm most familiar with a style of blockchain called an Optimistic Rollup. So I used that logic.

Optimistic Rollup is a pattern where systems assume data is valid but enable verification. (Hence, "optimistic")

Verifiable Claude

When Claude makes a factual claim, it should provide a fraud proof - evidence that can be independently verified.

Making AI more accurate is hard to do. People are paid zillions of money for it. But what if I give AI some optimistic rollup energy.

1. Optimistic execution: Claude responds quickly with claims

2. Challenge period: Users can flag suspicious claims

3. Fraud proofs: System verifies claims against sources

Just like Arbitrum assumes validity but enables verification, Verifiable Claude assumes truthfulness but enables fraud proofs.

verifiable-claude.vercel.app

Verifiable Claude - Fraud Proofs for LLMs

Cryptographically verifiable AI responses. Every claim is committed to a Merkle tree and verified against web sources — making hallucinations provably detectable.

Tech used:

• Brave API for web searches

• Anthropic API for Sonnet/Opus queries

• Claude for vibe-coding