[Paper Review] Belief Updating and Delegation in Multi-Task Human-AI Interaction: Evidence from Controlled Simulations

Shreyan Biswas, Alexander Erlei+1 more|arXiv (Cornell University)|Feb 2, 2026

Explainable Artificial Intelligence (XAI)0 citations

TL;DR

The paper shows that users carry over beliefs about AI accuracy across multiple tasks, update them conservatively, and delegate mainly based on subjective AI accuracy beliefs rather than self-confidence, using a preregistered multi-task experiment with controlled AI accuracy levels.

ABSTRACT

Large language models (LLMs) increasingly support heterogeneous tasks within a single interface, requiring users to form, update, and act upon beliefs about one system across domains with different reliability profiles. Understanding how such beliefs transfer across tasks and shape delegation is therefore critical for the design of multipurpose AI systems. We report a preregistered experiment (N=240; 7,200 trials) in which participants interacted with a controlled AI simulation across grammar checking, travel planning, and visual question answering, each with fixed, domain-typical accuracy levels. Delegation was operationalized as a binary reliance decision: accepting the AI's output versus acting independently, and belief dynamics were evaluated against Bayesian benchmarks. We find three main results. First, participants do not reset beliefs between tasks: priors in a new task depend on posteriors from the previous task, with a 10-point increase predicting a 3-4 point higher subsequent prior. Second, within tasks, belief updating follows the Bayesian direction but is substantially conservative, proceeding at roughly half the normative Bayesian rate. Third, delegation is driven primarily by subjective beliefs about AI accuracy rather than self-confidence, though confidence independently reduces reliance when beliefs are held constant. Together, these findings show that users form global, path-dependent expectations about multipurpose AI systems, update them conservatively, and rely on AI primarily based on subjective beliefs rather than objective performance. We discuss implications for expectation calibration, reliance design, and the risks of belief spillovers in deployed LLM-based interfaces.

Motivation & Objective

• Investigate how users form, update, and transfer beliefs about an AI system’s accuracy when interacting with the same AI across multiple tasks with different accuracies.
• Examine how users’ beliefs about AI accuracy and their self-confidence jointly influence delegation decisions in human–AI interaction.
• Assess how dispositional trust and related individual differences shape users’ initial beliefs about AI accuracy before observing performance.

Proposed method

• Conduct a preregistered, within-subject experiment (N=240; 7,200 trials) across three tasks with fixed accuracy levels: grammar error detection at 30%, travel planning at 60%, and visual question answering at 90%.
• Use pre-scripted AI outputs to simulate task-specific accuracy and measure belief updating against a Bayesian benchmark (Beta–Binomial model) and conservatism in updating.
• Operationalize delegation as a binary choice per trial (delegate to AI vs. answer oneself) and track trial-by-trial beliefs, confidence, and trust dynamics.
• Collect baseline dispositional measures (TiA, AI literacy, Need for Cognition) and post-task trust to predict initial priors and modulation of updating and delegation.

(a) Grammar Error Detection

Experimental results

Research questions

• RQ1 RQ1: How do users form, update, and transfer beliefs about an AI system’s accuracy when interacting with the same system across multiple tasks with different accuracies?
• RQ2 RQ2: How do users’ beliefs about AI accuracy and their self-confidence jointly influence delegation decisions in human–AI interaction?
• RQ3 RQ3: How do dispositional trust and related individual differences shape users’ initial beliefs about AI accuracy before observing performance?

Key findings

• Priors about AI accuracy do not reset across tasks but carry over from the previous task, showing cross-task belief inertia.
• Within tasks, belief updates move in the Bayesian direction but are about half as strong as Bayesian updating (conservatism bias).
• Delegation is primarily predicted by lagged beliefs about AI accuracy rather than self-confidence once beliefs are controlled.
• Higher dispositional trust predicts higher initial priors about AI accuracy, with AI literacy providing an independent boost.
• The study provides systematic evidence of belief updating and delegation in a multi-task, multi-accuracy setting, highlighting cross-task spillovers and bounded rationality in belief dynamics.

(b) Travel Planning