Vitalik Buterin Tests Three Privacy Layers for Remote AI Requests

Ethereum co-founder Vitalik Buterin tested a three-part setup for using remote artificial intelligence with personal health and travel data, aiming to limit the information exposed while seeking diet and exercise recommendations.

A local Qwen3.8-Flash-Next model drafted queries with fewer identifying details before sending them to remote frontier models. The setup also used zkAPI for payment-channel privacy and Tor to add network and IP privacy. Buterin said all three layers were needed.

The setup returned recommendations that Buterin said remote models had helped improve. The remote models were not identified, and no independent security assessment was provided.

The approach does not make the data fully private. The API provider can still see prompts, and network and timing metadata may remain observable.

Buterin said the local model ran at 20 to 30 tokens per second, slower than he found comfortable. He estimated Tor’s request-by-request privacy approach added 10 to 100 times more latency than he said he could accept. Those figures reflect his observations and estimates, not independent benchmarks.

The experiment highlights a trade-off: removing more personal context from a request may limit the help a remote model can provide. “The more careful you are about what data you give to a remote model, the less it can help you,” Buterin wrote.

The test builds on work around zkAPI’s private AI payments on Ethereum.