On Working with Wizards
One Useful Thing Ethan Mollick
AI systems like GPT-5 Pro and Claude 4.1 Opus now perform complex tasks autonomously with impressive results, but users cannot see how the AI made its decisions or fully verify the work. The shift from collaborative co-intelligence to passive consumption means people receive sophisticated outputs in minutes—such as GPT-5 Pro analyzing an academic paper in 9 minutes 40 seconds and finding a previously unnoticed error—but lose transparency and the ability to develop expertise. Users must develop new skills: knowing when to use wizard-like AI versus collaborative AI, becoming judges of output quality rather than understanding process, and accepting provisional trust in systems where perfect verification is impossible.
Why it matters
Verifying magic on the jagged frontier
