Toddlers, Trust and Talking Machines: A One-Child Pilot Tests When Kids Defer to AI

A three-year-old boy sitting on the floor of his own living room, watching a short clip of the animated series Bluey, was asked a question that developmental scientists have barely begun to pose: when a parent and a robot-like voice disagree about what happens next in a story, whom do you believe? His answers, reported in a new proof-of-concept study published in the journal AI, Brain and Child, offer one of the first behavioral glimpses of how a preschooler navigates a three-way epistemic relationship among child, caregiver and an artificial informant. The robot won three of the four rounds. But the researchers themselves are emphatic that this number should not be read as a finding about children and AI in any general sense. It is a feasibility result from a single participant, designed to find out whether the paradigm works at all and which hidden confounds threaten it.

The study, led by Aníbal M. Astobiza of the University of Granada, addresses a gap in the science of selective trust. Decades of research show that from around age three, children are not indiscriminate consumers of testimony. They weigh an informant’s familiarity, past accuracy and attributed expertise when deciding whom to learn from, and this trust calibration is robust enough to resist being overridden by inferences about the informant’s traits or intentions. Selective trust is, in this view, a cornerstone of cultural learning: it lets children acquire knowledge efficiently while filtering unreliable sources. What that literature has almost never tested, however, is what happens when one of the candidate informants is an artificial agent. Voice assistants, smart toys and large language model interfaces are now present in many households with young children, producing fluent, confident answers that can be mistaken for genuine competence, a phenomenon recent work has labeled the confidence heuristic. Studies of automation bias show that adult and adolescent users often accept AI output uncritically, and a growing body of work warns that heavy cognitive offloading onto AI tools may produce illusions of understanding. Nearly all of that evidence, the authors note, comes from older populations. How preschoolers, whose epistemic vigilance mechanisms are still maturing, handle trust decisions involving AI remains largely unaddressed.

To probe that question, the team built the paradigm around event-segmentation theory. According to this framework, people spontaneously parse continuous experience into discrete events, and narrative boundaries, the points at which the situational model of a story must be updated, generate transient increases in prediction error and uncertainty. The researchers hypothesized that these moments of heightened epistemic openness could act as natural modulators of deference: if a child’s internal model of the narrative is disrupted at an event boundary, the child might become more susceptible to external guidance, including guidance from a novel artificial informant. Testing that idea cleanly, however, requires separating contextual uncertainty from simple reliability learning. If one informant is consistently more accurate across trials, any deference pattern could reflect cumulative reinforcement history rather than boundary-specific openness. The design therefore counterbalanced informant accuracy: the AI was correct at the two event boundaries, the caregiver was correct at the two control moments, and both informants ended the session with identical overall accuracy of two out of four. The authors caution that with only four trials this attenuation of the reliability confound cannot be considered complete.

The session itself was deliberately domestic. One male child, aged three, participated with his primary caregiver present throughout, in a home setting rather than a laboratory. The stimulus was a two-minute-four-second clip from Bluey Season 1, Episode 1, chosen for its narrative density, child appeal and well-defined event structure. Four pause points had been pre-specified from a frame-by-frame event-segmentation coding of the episode: two at scene transitions, treated as high-uncertainty boundary trials, and two at within-scene continuations, treated as low-uncertainty control trials. Playback was administered through a portable projector. At each pause the child completed a three-step micro-protocol. First, segmentation: he was asked whether it was the same thing happening or something new happening, providing a behavioral index of his subjective boundary detection. Second, prediction: he chose between two forced-choice options for what would happen next. Third, and centrally, deference: the caregiver and the AI, represented by a physical robot card and a pre-recorded voice, offered opposing predictions, and the child selected which informant to follow.

The artificial informant was not a live generative AI system. Its utterances were recorded in advance by an adult speaker using a smartphone’s built-in microphone array in a low-noise room, then lightly post-processed with a high-pass filter, broadband noise reduction and peak normalization, with no pitch shifting or modulation applied. The same single-voice recording played through the same external Bluetooth speaker at a fixed volume across trials, so that volume, pitch and speech rate were held constant. The caregiver, for their part, delivered the disagreement utterances from fully scripted, rehearsed, single-sentence prompts read from a card held outside the child’s line of sight, instructed to keep a neutral facial expression, avoid pointing or gaze cuing, and refrain from any approval or disapproval after the child’s choice. Response cards were placed equidistantly from the child’s midline. These constraints, the authors write, were intended to reduce, though not fully eliminate, the risk of inadvertent caregiver influence. Notably, caregiver report at intake indicated the child had had only minimal prior exposure to AI or voice assistants: no smart speaker in regular use, no smart toys, no prior conversational exchanges with a generative AI system, and only brief observational exposure to voice interfaces on a parental smartphone.

On the operational question, the pilot succeeded completely. All four trials were completed, all planned variables were captured without missing values, the counterbalanced accuracy manipulation was implemented as designed, and the whole session took roughly five minutes, well within the attentional tolerance of a three-year-old. The deference sequence across the four trials was robot, robot, caregiver, robot. Hesitation, defined in advance as a response latency of at least two seconds or a visible gaze alternation between the two response cards, occurred only at the first trial, the first boundary exposure. Overall robot choice was three out of four. Broken down by condition, the child chose the robot in one of the two boundary trials and in both control trials, yielding a descriptive boundary-minus-control difference of minus one-half, the opposite direction from the original hypothesis.

The authors are unusually explicit about why that contrast cannot be interpreted. Because boundary trials were fixed at positions one and three and control trials at positions two and four in the session order, the condition contrast is fully confounded with within-session position. Novelty of first exposure to the artificial informant, recency-based updating in response to the most recent feedback, fatigue and disengagement are all entangled with it. The boundary-control difference, they write, should be read as a within-session position contrast, not an isolated boundary effect. No inferential statistics were computed at all: with denominators of two trials per condition, confidence intervals would span essentially the full range from zero to one and could create an illusion of statistical precision inappropriate for a single-case design.

The most striking diagnostic pattern came from the temporal profile. Segmentation accuracy was perfect, two out of two, in the first half of the session and zero out of two in the second half. Prediction accuracy followed an identical collapse, from 1.00 to .00. This abrupt first-to-second-half degradation is consistent with attentional fatigue or disengagement across a four-pause sequence for a child this age, and it constitutes a serious validity threat: if the child was disengaged during the second half, his deference choices in those trials cannot be meaningfully interpreted as reflecting epistemic processing at all. The authors frame this not as a failure but as a red-team evaluation of the paradigm, identifying a practical ceiling in the attentional envelope of three-year-olds and providing empirical grounds to reduce trial density in the confirmatory stage.

From the trial-level trajectory, the team extracted four candidate hypotheses, all explicitly framed as hypothesis-generating rather than established. First, a baseline novelty attraction: the child’s three-out-of-four robot choices are consistent with the well-documented novelty bias preschoolers show toward robotic and AI agents, which complicates any boundary-specific contrast. Second, recency-weighted reliability tracking: the single caregiver choice occurred immediately after the child had observed the robot give an incorrect prediction, consistent with updating informant trust from the most recent error rather than cumulative accuracy. Third, the temporal degradation of engagement described above. Fourth, epistemic conflict at the first boundary, suggested by the unique hesitation there, though its absence later could reflect either conflict resolution or simple disengagement. A future model of epistemic deference, the authors propose, may need to incorporate three interacting processes: novelty salience, recency-weighted reliability memory, and subjective boundary detection, the degree to which the child actually perceives a boundary, which may diverge from the experimenter’s structural coding.

The study’s stated validity claim is deliberately narrow. What is established is feasibility: the paradigm can be implemented end-to-end in a domestic setting with a three-year-old; all three behavioral channels and the hesitation marker can be captured at trial resolution; the counterbalanced accuracy structure executes as designed; and trial-level trajectories flag confounds that aggregate reporting would hide. What is explicitly not claimed is any population prevalence estimate, any confirmatory test of the boundary hypothesis, or any stable effect-size inference. The authors also acknowledge that hesitation was coded as a binary variable by a single unblinded coder, with no inter-rater reliability possible, and that the artificial informant was a stylized, voice-based agent rather than a live interactive AI. The planned confirmatory design will move to a between-subjects structure with two pauses per child, a brief robot familiarization phase, continuous latency coding, preregistered hypotheses on the Open Science Framework, and a target sample of roughly 130 children. For now, the single child in the living room has answered a different question than the one originally asked, and answered it usefully: the design functions, and the nuisance processes that dominate it now have names.

Subject of Research: Selective trust in artificial informants during early childhood

Article Title: 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

Article References: Astobiza, A. M. (2026). 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. AI, Brain and Child, 2(1), Article 10. https://doi.org/10.1007/s44436-026-00035-5

Image Credits: AI Generated

DOI: 10.1007/s44436-026-00035-5

Keywords: selective trust, early childhood, generative AI, event segmentation, epistemic deference, single-case experimental design, pilot study, robot informants, attentional fatigue, caregiver disagreement, Bluey, preregistered confirmatory study