Ethereum co-founder Vitalik Buterin said on October 29 that artificial intelligence models remain “extremely knowledgeable” but “not very wise” when they encounter newly published writing that was not part of their training data. In a post on X, he argued that AI can detect references with precision, but often makes only surface-level judgments about what a new piece is actually arguing.
Buterin’s comments came amid an online debate over a Wall Street Journal article about Harvard’s Humanities 10 course. He noted that when models evaluate famous texts, they can draw on auxiliary literature and years of commentary. With brand-new essays, whitepapers or governance proposals, however, that supporting material is absent, leaving the model to rely mostly on wording rather than substantive reasoning.
He connected the issue to crypto security, cautioning that teams relying on AI summaries for smart-contract audits, research and token-supply schedules could receive fluent but shallow overviews that miss subtle flaws. Buterin has previously warned that AI could weaken key parts of crypto defenses within two years and has backed human-in-the-loop tools such as a “paranoid mode” for AI chats.
The comments follow a July experiment in which Franklyn Wang’s team at Co-Invest used AI to identify Buterin as the likely author of an anonymous rewrite of EIP-7503. The model assigned roughly a 20% probability to Buterin’s authorship after evaluating reasoning style rather than vocabulary.
