Ethereum founder Vitalik Buterin announced on Tuesday via X that his 2024 AI anonymity challenge has a winner. The experiment proved that large language models can identify not just prose style but also deeper cognitive signatures in mathematical reasoning and algorithm design.
Setup: Write in Chinese, Translate Back to English, Double Camouflage
Vitalik described the process: in 2024, he wrote a revised version of EIP-7503 (Zero-Knowledge Wormhole) in Chinese, used the Qwen 2.5 model locally for translation, and then manually tweaked the output to erase all traces of his authorship. The strategy involved two layers of disguise — first, writing in Chinese and translating back to English to blur stylistic markers; second, manual editing to ensure natural fluency. Vitalik assumed that obscuring prose style would suffice to hide his identity.
AI Bypasses Prose Camouflage, Targets Math Thinking Habits
The result surprised him. The Qwen model ignored the prose-level disguise and went straight for unique patterns in Vitalik's mathematical and algorithmic explanations. Key features captured include: concrete numeric examples — a habit of using specific numbers to build intuition for abstract concepts; logical chains — distinctive connection patterns from premises to conclusions; algorithm explanation style — rhythm of language, choice of analogies, and depth of detail forming a recognizable signature. Vitalik noted that while the prose camouflage was effective, the model completely bypassed it and identified him through "thinking habit fingerprints."
From Stylometry to Cognitive Pattern Reading: A Turning Point in AI Text Analysis
This experiment goes beyond validating Qwen's capability. It marks a shift in AI text analysis: early identification models relied on prose style (sentence length, word preference, punctuation usage), but next-generation models can now capture deeper cognitive features — reasoning structure, conceptual organization, and problem-solving strategies. Practical implications range from academic authorship verification and technical document tracing to detecting multi-layer camouflage tactics (humanize then disguise) used to conceal AI-generated text. While modest in scale, Vitalik's experiment provides a concrete empirical case for the field of AI text fingerprinting.
Source: Vitalik Buterin X post, Jinse Finance; compiled by Flip from BlockTempo