A recent arXiv preprint reports on an online study in which 56 AI-literate participants completed 392 Video Question Answering trials with a multimodal AI. After seeing each AI response, participants rated their trust in the system for that task and chose one of three policies for a similar future task: reject AI help, use it with verification, or delegate fully.
The authors found that post-interaction trust strongly predicted delegation choices (odds ratio 14.97), yet trust and delegation were not the same thing. Verification was the most common policy, and it frequently appeared even when trust was high, suggesting that users often prefer to keep some oversight rather than hand over full control.
Delegation also increased with task order after controlling for both momentary and participant-level trust (odds ratio 1.22). This points to an experience-related effect that self-reported trust did not capture. The authors argue that researchers should move from asking about intended delegation to observing enacted delegation, and they propose "enacted delegation traces" as a methodological framework for studying human oversight in multimodal and agentic AI systems.
As the study is a single preprint, its findings are not yet independently replicated, and the sample is limited to AI-literate participants. Still, the results offer a concrete starting point for understanding when and why people hand control to AI. The authors do not claim to resolve this question, but their proposed trace-based method may help future work study delegation as it actually happens. (arXiv:2610.21948)