When should humans remain in the loop, remain only as monitors, or be excluded because intermittent intervention degrades performance?
Nominal human-in-the-loop requirements can become safety theater: critics argue oversight mandates assume vigilance humans cannot sustain, while defenders reply that oversight with real authority, time, and information does add safety. Which conditions separate the two is the open question.
View on the map → · Open in Browse →
The literature is trying to move 'keep a human in the loop' from slogan to design. Proposals converge on exploiting the solve-verify asymmetry — engineer AI outputs so a human can check them without redoing the work — and a formal treatment proves conditions under which granting an agent more autonomy cannot hurt the overseer. A counterweight comes from experiments showing the human side of the loop degrades for social reasons: teams performed worse merely believing a competent teammate was an AI.
We also exploit the solve-verify asymmetry by designing AI outputs so that humans can efficiently check and contest them without having to resolve the task.
A concrete design answer to the oversight-as-theater critique: keep AI's operative agency but engineer outputs for human evaluative agency, exploiting the asymmetry that verifying is cheaper than solving.
a foundational framework, with a working definition, architecture and processes for effective human oversight of AI systems
A 16-author interdisciplinary attempt to give 'effective human oversight' a working definition, architecture, and research agenda — the synthesis that regulators invoking oversight requirements currently lack.
When this game forms a Markov Potential Game, we prove an alignment guarantee: any increase in the agent's utility from acting more autonomously cannot decrease the human's value.
Formalizes the ask-for-oversight vs act-autonomously decision as a game and proves conditions under which granting autonomy cannot hurt the human overseer — a rare theoretical handle on when humans can safely step back.
teams performed worse when a teammate was perceived as AI, especially when task difficulty increased.
Experimental evidence that merely believing a competent teammate is an AI degrades team performance and communication — the human side of the loop malfunctions for social-psychological reasons even when the AI is at human level.