ETH creator reveals privacy-focused AI experiment

TLDR

Ethereum creator Vitalik Buterin is testing a privacy-focused AI setup that uses his own health and travel data without fully exposing it to powerful models.

1. Buterin built a three-layer system using a local model, Ethereum-based zkAPI payments, and Tor to get personalized health advice while limiting data leakage.
2. The experiment showcases how Ethereum can underpin privacy-preserving AI agents, but current tools still leave some data and metadata visible.
3. Performance and security trade-offs remain significant, so the main thing to watch is how these ideas evolve into practical, audited infrastructure for private AI on-chain.

Deep Dive

1. How The Experiment Works

Vitalik Buterin describes a personal experiment where AI generates tailored diet and exercise recommendations from his health and travel data, while trying not to reveal that raw data to frontier models. In reports, he uses a local model, Qwen 3.8 Flash Next, to rewrite and orchestrate queries before they are sent to more capable remote models, masking personally identifiable information and his writing style in the process. The privacy stack then adds zkAPI for payment privacy and Tor for network-level anonymity, forming three layers that he argues are all needed for meaningful protection in this setup.

2. Ethereum, zkAPI And Private AI

zkAPI is a metered API payment system coauthored by Buterin that recently went live on Ethereum mainnet, enabling API providers, including AI services, to charge usage without directly linking specific requests to a funding wallet, using zero knowledge proofs for settlement. In this experiment, zkAPI separates his payment identity from individual AI calls, while Tor hides his normal IP address, showing how blockchain-based tools can complement traditional privacy techniques for AI workloads. The result is not perfect secrecy, but a demonstration that Ethereum infrastructure can support more private AI agents and services, where payments, identity, and data exposure are partially decoupled.

3. Limits, Risks And What To Watch

Buterin is clear that the system has limitations: Tor adds an estimated 10 to 100 times more latency than he finds acceptable, the local model runs at roughly 20 to 30 tokens per second instead of the 100-plus he wants, and limiting data to protect privacy also reduces the quality of help the remote model can provide. Independent security assessments are still missing, API providers can still see prompts, and timing or network metadata may remain linkable, so this is experimental rather than production-grade privacy. The next key signals will be code improvements in the zkAPI stack, external audits, and emerging applications where on-chain payments and privacy-conscious AI agents are used for real workloads such as health, finance, or trading.

Conclusion

Vitalik Buterins privacy-focused AI experiment is a small but important step in showing how Ethereum-based tools like zkAPI can support more private, agentic AI use without surrendering full control of personal data. The architecture highlights both the promise of combining local models, zero knowledge payments, and network anonymity, and the gaps in speed, usability, and verifiable security that still need to be closed before such systems can safely handle sensitive data at scale.