Vitalik praises Qwen model's ability to recognize anonymity: AI sees through his mathematical thinking habits, prose disguise completely fails.

Vitalik Buterin stated that an AI anonymous identification challenge he initiated in 2024 has a winner. He first wrote the revised version of EIP-7503 (Zero-Knowledge Wormhole) in Chinese, then used Qwen 2.5 for local translation and manually corrected it, attempting to conceal his authorship, but was successfully identified by AI through his unique thinking habits in mathematics and algorithm explanations.
(Previous context: Ethereum's third major iteration》Vitalik analyzes the Lean Ethereum roadmap: almost all important components will be replaced)
(Background supplement: Alibaba will integrate Qwen AI into Taobao's 4 billion products: automatic price comparison, ordering, and logistics management)

Table of Contents

Toggle

• Experimental Setup: Write in Chinese first, then translate back to English
• AI's Approach: Mathematical thinking is harder to disguise than prose style
• Extended Implications: AI identification advances from "style comparison" to "thinking pattern reading"

Ethereum founder Vitalik Buterin posted on the X platform on Tuesday, announcing that an AI anonymous identification challenge experiment he initiated two years ago has a winner. This experiment verified that large models can not only recognize prose styles but also capture human thinking imprints in mathematical reasoning and algorithm design.

Experimental Setup: Write in Chinese first, then translate back to English

According to Vitalik's description, the entire experimental process was quite straightforward but ingeniously designed: In 2024, he wrote the revised version of EIP-7503 (Zero-Knowledge Wormhole) in Chinese, then used the Qwen 2.5 model for local translation, and finally manually corrected the translation results, attempting to completely erase his own authorial traces.

The core of this experiment lies in two layers of disguise: the first layer is writing in Chinese and then translating back to English, using the translation process to blur the prose style; the second layer is manual correction to ensure the translation results are natural and fluent. Vitalik's strategy focused on obfuscating the prose style—he believed that as long as the style was blurred, the author's identity could be hidden.

AI's Approach: Mathematical thinking is harder to disguise than prose style

The results, however, showed that the Qwen model bypassed the prose-level disguise and directly targeted Vitalik's unique thinking habits in mathematical and algorithmic explanations. Specifically, the key features captured by the model include:

• Specific numerical examples — When explaining abstract concepts, Vitalik habitually uses concrete numbers to build intuition
• Logical chains — His reasoning steps exhibit a unique connection pattern, with a personal style in deriving conclusions from premises
• Algorithm explanation style — The rhythm of language, choice of metaphors, and depth of detail when describing algorithms all have identifiable patterns

Vitalik remarked that although the prose style disguise was quite successful, the AI completely bypassed his obfuscation strategy and identified him directly through "thinking habit fingerprints."

Extended Implications: AI identification advances from "style comparison" to "thinking pattern reading"

The significance of this experiment goes beyond verifying the capabilities of the Qwen model. It reveals a turning point in AI text analysis: early identification models mostly relied on prose style (sentence length, word preference, punctuation usage), while newer models can capture deeper cognitive features—reasoning structure, concept organization, and problem-solving strategies.

This capability has multiple implications in practical applications: identification of academic paper authors, tracing of technical documents, and even detecting multi-layer strategies that first "humanize" and then "disguise" in AI ghostwriting will face greater identification pressure. Although Vitalik's experiment was small in scale, it provides a concrete empirical case for the field of AI text fingerprinting.

This article is based on Vitalik Buterin's X post and Jinse Finance flash news, translated by Flip, editor of Dynamic Zone