📰 Key Highlights

Ethereum co-founder Vitalik Buterin voluntarily launched a “de-anonymization” experiment on June 22, publicly admitting he had previously published a document of “medium importance” to Ethereum under an anonymous identity, and challenging the outside world to find that document. The backdrop comes from a February research paper by ETH Zurich and Anthropic, which pointed out that large language models can now perform online de-anonymization at scale by extracting identity-related clues from unstructured text, searching for potential matches, and reasoning about the most likely candidate — outperforming traditional techniques.

The AI tool Co-Invest later analyzed 27 documents and listed Buterin as the most likely author of an anonymous rewrite of EIP-7503 from December 2024, with roughly 20% confidence — about 10 times that of the second candidate. Buterin later confessed that he deliberately wrote the original in Chinese, then translated it to English using Qwen 2.5 with manual edits, trying to mask his personal writing style. He pointed out, however, that the stylistic clues AI picked up were his mathematical thinking habits and the way he explains algorithms — “intellectual fingerprints” that completely bypassed his disguise strategy aimed only at prose phrasing. Lighter CEO Vladimir Novakovski added that he had attempted a similar approach in 2023, using GPT-4 to compare writing styles in an attempt to track down Satoshi Nakamoto’s identity, but failed to reach any high-confidence conclusion. This experiment once again highlights the potential threat that AI text analysis poses to anonymous contributors in open-source communities.

💬 JudyAI Lab Perspective

Vitalik Buterin personally stepped into the ring to test it — deliberately drafting in Chinese and then translating through Qwen 2.5 — and still couldn’t dodge AI de-anonymization analysis. This makes one thing crystal clear: the boundary of identity lies not in language, but in thinking habits themselves.

What AI builders should pay the most attention to in this experiment is that Co-Invest analyzed 27 documents, and its judgment was based not on prose tone, but on the structure of Buterin’s mathematical thinking and the logic of his algorithm walkthroughs. He later admitted that these “intellectual fingerprints” completely bypassed his disguise strategy aimed at writing style. According to the ETH Zurich and Anthropic research, large language models can already extract identity clues from unstructured text, outperforming traditional techniques. For anonymous contributors in open-source communities, this is a security dimension that has barely been taken seriously. The granularity of AI analysis has evolved from “how you write” to “how you think.”

If you have anonymous writing that needs protection, it’s worth thinking ahead not just about changing language or word choice, but about this: has your reasoning framework itself already become a trackable feature?

📅 Original Source Info

🔗 Further Reading