Conditions under which people delegate decisions to generative AI

Identify the circumstances under which people are willing to delegate decisions to generative artificial intelligence in social interactions that have consequences for others.

Background

Delegation to AI is increasingly feasible in everyday interactions, yet the behavioral drivers of when individuals choose to cede decision authority to AI remain insufficiently characterized. The authors flag this as an unresolved issue central to understanding the societal impact of AI-mediated decision-making.

References

In particular, several fundamental questions regarding human behavior in AI-mediated social interaction remain unaddressed. Under what circumstances are people willing to delegate decisions to generative AI?

Generative AI Triggers Welfare-Reducing Decisions in Humans  (2401.12773 - Dvorak et al., 2024) in Introduction (Section 1)

Our data locate this variability in the person but cannot say what property of the person produces it; the account below is an interpretation, not a result. Prior studies have established that automation use is driven substantially by stable, person-level dispositions . Merritt and Ilgen showed that a person's general tendency to trust machines predicts how much they use automation, even after accounting for the system's actual qualities. Since then, this personal tendency, often called dispositional trust, has been treated as a key factor shaping how people interact with automation . Our result is compatible with this: reliance on a conversational agent looks less like a response evoked by a particular task and more like the expression of a relatively stable individual tendency that users bring with them into the system. On this reading, the chat assistant is not a tool whose use is driven by task demands, but one whose adoption reflects a disposition that varies from person to person. We therefore treat dispositional trust as a hypothesis rather than a conclusion. Testing it would require pairing this behavioral logging with a propensity-to-trust instrument administered before exposure to the system.

Delegating or Doing? Understanding User Behavior in Hybrid Human-Agent Interfaces  (2608.19551 - Dizon et al., 20 Aug 2026) in Discussion, Section “Reconsidering Risk-Calibrated Delegation” (Section 5.1)