On Working with Wizards (www.oneusefulthing.org)

The author reflects on the evolving relationship between humans and AI, shifting from the earlier concept of “co-intelligence” — where humans and AI collaborate interactively — toward a new dynamic likened to working with “wizards.” Unlike co-intelligence, these advanced AI systems such as GPT-5 Pro and Claude 4.1 Opus autonomously generate sophisticated outputs with minimal human intervention but operate opaquely, leaving users as passive recipients who must verify results without insight into the AI’s internal reasoning. This transition highlights a key challenge: while these AI “wizards” produce impressive, reliable work (e.g., critiquing academic papers with Monte Carlo analyses or transforming complex spreadsheets), their processes remain opaque, complicating trust and oversight. This shift is significant for the AI/ML community because it underscores the need for a new kind of literacy—one focused on evaluating and curating AI outputs rather than understanding their internal workings. Users must balance when to summon these powerful agents versus collaborate more actively, all while developing intuition for judging AI reliability across diverse tasks. The article emphasizes the paradox that as AI competence grows, so does opacity, making perfect verification increasingly impractical. Ultimately, embracing “provisional trust” in AI tools—that outputs are “good enough” despite limited transparency—will become essential, marking an era where human expertise adapts to working with enigmatic, highly capable AI wizards rather than fully controllable assistants.