Vitalik Buterin's Anonymity Pierced by AI Analyzing Reasoning Patterns
In brief
- Buterin challenged AI to identify an anonymous document written in Chinese and machine-translated on June 22.
- Franklyn Wang's AI ranked Buterin as most likely author with ~20% confidence, 10x higher than next candidate.
- Wang identified the document via intellectual habits and mathematical reasoning style, not prose patterns.
- Research from ETH Zurich and Anthropic confirms large language models enable practical deanonymization at scale.
The Challenge
Buterin issued his anonymity challenge on June 22, stating he had published a document of "medium importance" to Ethereum under a different name at some point in the past decade. He wanted to test whether current AI tools could discover his hidden contribution.
To obscure his identity, Buterin employed a deliberate obfuscation strategy. He wrote the anonymous rewrite of EIP-7503 in Chinese, translated it using Qwen 2.5, and manually corrected the translation to disguise his prose. The document was an alternative specification for an Ethereum protocol proposal.
How Wang's AI Cracked It
Wang's analysis identified the document by examining something Buterin hadn't anticipated: his reasoning patterns. Wang's winning submission identified the anonymous rewrite by analyzing the way it explained mathematical and technical concepts.
The distinction matters. "The tell wasn't his words, it was his reasoning," Wang said. Co-Invest ranked Buterin as the most likely author with roughly 20% confidence, about 10 times higher than the next candidate.
"Notice that the stylistic hints that his AI picked up on were intellectual habits and style of math and algorithm explanation, which bypassed my obfuscation strategy (which only covered prose) completely" — Vitalik Buterin
Buterin acknowledged the insight. His intellectual habits and mathematical reasoning style were the vulnerability his obfuscation strategy had overlooked.
Broader Implications
The result aligns with recent academic work on AI-driven deanonymization. Researchers from ETH Zurich and Anthropic claimed in a February paper that large language models have made online deanonymization practical at scale.
Wang himself has explored this terrain before. Vladimir Novakovski said he worked with Wang in a 2023 project using GPT-4 to try to identify Bitcoin creator Nakamoto by matching writing style, but the effort failed to produce a high-confidence result. Buterin's case differs: Wang had a bounded search space (Ethereum contributors) and access to Buterin's known writing.
The episode raises questions about anonymity in an era of large language models. Prose obfuscation alone isn't enough. Reasoning patterns, mathematical intuition, and conceptual frameworks are harder to hide — and harder to strip away without fundamentally changing the work itself.
Frequently asked questions
Why didn't Buterin's obfuscation strategy work?
Buterin wrote in Chinese and machine-translated to English to hide his prose style, but AI identified him through his intellectual habits and mathematical reasoning patterns instead. His obfuscation only covered word choice, not the deeper conceptual framework he used to explain ideas.
How confident was Wang's AI in identifying Buterin?
Co-Invest ranked Buterin as the most likely author with roughly 20% confidence, about 10 times higher than any other candidate. The bounded search space (known Ethereum contributors) and access to Buterin's prior writing strengthened the analysis.
Is this the first time AI has deanonymized someone online?
Researchers from ETH Zurich and Anthropic published a February paper claiming large language models have made online deanonymization practical at scale. Prior efforts, like Wang's 2023 GPT-4 project attempting to identify Bitcoin's creator Nakamoto, failed to produce high-confidence results.
