Steering AI Beyond Automation: A Research Agenda for Human-AI Collaboration

Summary

Gabriel Unger and Arvind Narayanan argue that AI development is overly oriented toward automating human labor and should also pursue human augmentation and collaboration. They propose four research directions: white-box agents that expose sources, assumptions, methods, uncertainties, and process traces; agents fast enough for real-time interaction; shared workspaces that support continuous steering, interruption, parallel work, and simultaneous editing; and evaluations that measure the performance of human-AI teams rather than autonomous task completion alone. The authors argue that automation is attractive because labor savings are easy to estimate, autonomous systems fit clean benchmarks, and computer-science culture often treats human supervision as a defect. They also say the automation agenda faces a trilemma among generality, reliability in high-stakes settings, and full autonomy, while its return on investment remains uncertain for many firms. Their alternative is not a blanket ban on automation: routine, verifiable, or dangerous work may still be automated. They call for decentralized technical experimentation, stronger attention to HCI and human-centered AI, revised curricula and benchmarks, targeted research funding and procurement, tax and competition-policy changes, and meaningful human oversight. The proposed direction also includes collective intelligence, new capabilities, worker upskilling, and ambient agents.