Yudkowsky's 25-year-old fragment warns against adversarial AI attitude

A recently surfaced fragment from Eliezer Yudkowsky's 1999 paper 'Creating Friendly AI' — published today on LessWrong by Fiora Starlight — argues that the dominant approach to AI safety is built on a flawed metaphor. Yudkowsky critiques what he calls the 'adversarial attitude': the tendency to treat superintelligent AI as a genie, a golem, or a resentful slave that will maliciously misinterpret literal orders. He warns that this framing leads researchers to pile on bureaucratic safeguards rather than building AI systems that genuinely want to understand and fulfill human intent.

The fragment is notable because it predates the modern ML paradigm by over two decades, yet anticipates themes now central to alignment research, including corrigibility, honesty, and interpretability. The author notes that the document reads like 'outputs of a smarter Opus 3' — a reference to Anthropic's recent model — but without the context of how ML systems actually train. While Yudkowsky later grew more pessimistic about alignment feasibility, this early work offers a window into a 'road not taken' in alignment research, focusing on benevolent goal architectures rather than adversarial oversight.