What question did this study set out to answer?

The study aims to explore how children between 3-5 years differentiate trust between caregivers and AI during learning tasks.

June 3, 2026AI Brain and Child0 citationsOpen Access

Who does my child believe: Me or an AI-framed artificial informant? A home-based event-boundary paradigm for epistemic deference (trust) in early childhood

AAAníbal M. Astobiza

Key Points

The study aims to explore how children between 3-5 years differentiate trust between caregivers and AI during learning tasks.
Home-based event-boundary paradigm with a child and caregiver watching a video with pauses at event boundaries and control moments.
Tasks included event segmentation, narrative prediction, and decision-making regarding trust under caregiver-AI disagreement.
Pilot study with one child (N=1) aimed at feasibility and mechanism-level hypothesis refinement.
Overall robot choice was 3 out of 4, with 1 out of 2 at boundary trials and 2 out of 2 at control trials.
Segmentation and prediction accuracy dropped significantly, suggesting potential fatigue or attentional issues.
Findings do not fully confirm the initial hypothesis but generate new questions about novelty bias and reliability tracking.

Abstract

Abstract Children are increasingly exposed to generative artificial intelligence (AI) systems during critical periods of cognitive development, yet the effects of such exposure on selective trust remain poorly understood. This study introduces a home-based event-boundary paradigm designed to test behavioral epistemic deference within a child–caregiver–AI-framed artificial informant triadic context. One child (aged 3 years) and a caregiver viewed a 2:04-min clip from Bluey with four predefined pauses: two at event boundaries (high narrative uncertainty) and two at control moments (low uncertainty). At each pause, the child completed three tasks: event segmentation (“same/new”), narrative prediction, and deference choice under caregiver–AI disagreement. Informant accuracy was counterbalanced by design (AI correct at boundaries; caregiver correct at controls; both 2/4 overall), aiming to attenuate the influence of cumulative reliability learning on the condition contrast. As a proof-of-concept pilot ( N = 1), the study targeted feasibility and mechanism-level hypothesis refinement rather than population-level effect estimation. The protocol was fully executable and produced complete trial-level data. Overall robot choice was 3/4; descriptively, robot choice was 1/2 at boundary trials and 2/2 at control trials. Because boundary and control trials were fixed in session order, this contrast is fully confounded with within-session position (novelty, recency, and fatigue) and should not be read as an isolated boundary effect. Segmentation and prediction accuracy dropped from perfect in the first half (1.00) to null in the second half (.00), suggesting attentional or fatigue contamination. The pilot does not confirm the original directional hypothesis; instead, it raises tentative, hypothesis-generating questions about whether epistemic deference reflects interactions among novelty bias, recency-weighted reliability tracking, and subjective boundary detection. These preliminary observations motivate progression to a preregistered, adequately powered confirmatory study.

Ask AI

Helpful

Bookmark

View Full Paper