IC3 study: Crypto provides limited fixes for AI problems

Researchers from the Initiative for CryptoCurrencies and Contracts (IC3) published a survey Monday examining how crypto tools do and do not address key challenges in artificial intelligence. The paper reviews applications such as wallets, blockchains and decentralized governance and identifies specific limits to their effectiveness.

The authors say giving AI agents access to crypto wallets allows programmatic on-chain transactions but does not make the agents more intelligent or immune to human control. “AI systems do not become more intelligent by possessing a wallet. Nor do they become more resistant to human manipulation or shutdown,” the paper writes. Wallets can automate tasks like trading or accessing on-chain services without human approval, the researchers note, but automation is not the same as autonomous agency.

The paper cites recent product moves as examples of growing industry interest in agent-enabled finance. One major wallet provider launched a non-custodial wallet intended for AI agents, and a trading platform has begun testing AI agents that will trade on users' behalf. The IC3 team points out that similar automation could be implemented with conventional payment systems and that agents remain dependent on human-designed infrastructure and governance.

On content provenance, the report says blockchains are effective for recording timestamps and registering digital artifacts, but they cannot determine whether a piece of content was created by a person or by a machine. “Blockchains are well-suited for timestamping and registering specific digital artifacts,” the paper states, adding that identifying AI-generated content would require an external classifier. If that classifier produces an incorrect label, the error would be permanently recorded on-chain.

The researchers also highlight a practical limit: most digital content is produced and shared on platforms that do not support cryptographic anchoring. As a result, only content that is deliberately recorded on-chain will have a verifiable on-chain record, leaving large amounts of material without reliable provenance in the system.

The survey questions claims that decentralization can resolve bias in AI models. The authors argue that algorithmic bias typically arises during model training and is addressed by changing training data or model design. “Algorithmic bias is unlikely to be solved by decentralized AI, because it arises inherently in the training process and is typically mitigated by revised training or inference techniques,” the paper states. The report finds little direct evidence that broader participation in governance or added transparency from decentralization will reduce bias in model outputs.

The paper was authored by researchers from Cornell, Carnegie Mellon, Princeton, Yale and ETH Zurich and was edited by Carnegie Mellon’s Giulia Fanti and Cornell Tech’s Ari Juels, who is also listed as chief scientist at Chainlink Labs. The authors present a detailed account of where crypto can add value-such as secure record-keeping and automating on-chain interactions-and where claims about crypto solving deeper AI problems are overstated.