AI and Blockchain: A Record of Origins or an Overblown Promise?

Where did the data this model was trained on come from? And who owns the right to use it? These are two simple questions that many teams building systems cannot answer. Researchers audited more than 1,800 text datasets (2023) and found that over 70% had no licence stated on popular hosting sites, and that the error rate in licence categorisation exceeded 50%. Into this murk, blockchain is proposed as a remedy.

What does blockchain actually do?

Satoshi Nakamoto described it in 2008, in a paper on electronic cash: a peer-to-peer network records transactions in blocks, each linked to the one before, and producing a block takes computational effort, so what has been recorded is hard to change. In 2018 America's National Institute of Standards and Technology (NIST) described it as a distributed ledger with no central repository, designed to reveal tampering and resist it, and qualified its immutability as holding "under normal operation" of the network.

So it tells you: this record has not been changed since it was entered. It does not tell you: what is in it is correct.

What is expected of it alongside AI?

The proposal is to record on the ledger the fingerprints of datasets, their licences, model versions and decisions, so that one can return to them when asked: what data was this model trained on? Were its weights altered? The idea is reasonable, but we found no independent measurement of its success at scale, so we present it as a hypothesis, not a result.

Two limits apply. First, an honest ledger preserves error as faithfully as truth; if false or unlicensed data goes in, it is recorded as it is, and the problem of the bridge between the physical world and the digital ledger remains: who vouches for the first moment of entry? Second, many provenance tools in use need no chain at all. The C2PA standard attaches to an image or file a statement of its origin and edits and signs it digitally, so that any later change is detected; trust in it rests on signing certificates and trusted timestamps, not on a blockchain. It does not establish that what the statement says is true; it checks that the statement is tied to the file and has not been tampered with.

Proving a computation without revealing it

One promising idea is the zero-knowledge proof, a way for a party to show that an output came from a particular model without revealing its weights. Kang and colleagues (2022) presented what they described as the first proof of this kind for neural-network inference at ImageNet scale, but producing the proof for the largest model they tried (79.2% top-5 accuracy on images) took about 2,458 seconds, nearly 41 minutes. The technique remains costly and slow.

Energy also bears on the judgement: not all chains use the same energy. Ethereum cut its consumption by about 99.95% after moving in 2022 to proof of stake.

How to judge

1. Ask first: what problem would the ledger solve? If the answer is "verifying origin", a digital signature and a well-audited conventional record may suffice.
2. Distinguish "unchanged" from "correct"; the first is the ledger's business, the second is the business of verification and of whoever entered the data.
3. Require human governance to remain the final authority; technology executes but does not bear responsibility.

Trustworthiness in safeguarding rights and giving honest testimony begins with the human being; the ledger is only a tool to help.