Vitalik Buterin Tests Private AI Health Advice Using Local Models, zkAPI and Tor
Ethereum co-founder Vitalik Buterin has tested a privacy-focused system for obtaining personalized health advice from advanced AI models without directly exposing sensitive personal information to remote services.
The experiment used a local AI model to process health and travel information and generate recommendations related to diet and exercise. The local system could also consult more capable remote models when additional knowledge or reasoning was needed, while controls were designed to limit the information disclosed during those interactions.
How Buterin's Private AI System Worked
According to Buterin's X post, a local model managed the workflow and determined when external assistance was necessary. A skill file provided instructions on when to request help from remote models and how to construct queries that disclosed as little personal information as possible.
The local model also generated the external queries rather than having Buterin write them himself. This was intended to reduce another potential source of identification by preventing his writing patterns from being directly included in requests sent to outside AI systems.
The approach was designed around multiple potential privacy risks, including the information contained in prompts, payment activity and network connections.
zkAPI and Tor Add Payment and Network Privacy
Buterin used zkAPI to address the possibility that payment activity could reveal his identity when accessing remote AI services. He also used Tor to provide an additional layer of network privacy during communications with external models.
The setup ran zkAPI through Tor as a command-line tool, bringing the payment and network protections into the same workflow. Buterin said the three privacy measures were important because they addressed different ways his identity could potentially be exposed.
Privacy Comes With a Trade-Off
Buterin reported that the experiment worked and that input from more advanced remote models improved the health recommendations produced by the system. The external assistance was used to support personalized diet and exercise planning while keeping sensitive details away from those models as much as possible.
However, the experiment also highlighted a limitation of privacy-preserving AI systems. The less personal context made available to remote models, the less information those models had to produce highly tailored recommendations.
That created a trade-off between privacy and the usefulness of external AI assistance. Buterin's test attempted to balance the two by preparing requests locally, protecting payments through zkAPI and routing network traffic through Tor.
Implications for Private AI Assistance
The experiment illustrates how local AI models can be used as an intermediary between users and more powerful remote systems when sensitive information is involved. Rather than sending complete personal profiles to external models, the local system can determine what information is necessary for each request.
Buterin's results also show that stronger privacy controls can reduce the amount of context available to remote models. His test therefore focused not only on keeping personal information private, but also on finding a practical balance between limited disclosure and the quality of AI-generated assistance.
Writer: Marcus RenfieldCrypto Market Analyst & Onchain WriterMarcus Renfield covers cryptocurrency markets with a focus on onchain data, Bitcoin price action, and emerging market narratives. His writing examines how capital flows, network activity, and broader market structure influence short- and medium-term trends.He aims to provide clear, data-informed analysis for readers seeking a deeper understanding of crypto market dynamics.