AI Coach Helps Parents of Autistic Children Boost Communication During Storytime

An artificial intelligence system designed to coach parents in real time has shown promising results in a new experimental study of families raising minimally verbal autistic children. The system, called Parent’s AI Coach, or PaiCoach, was evaluated by researchers at the National AI Institute for Exceptional Education at the University at Buffalo, working with colleagues at the University of Texas at San Antonio, Duzce University, and Pennsylvania State University. In a study published in the Journal of Autism and Developmental Disorders, the team found that when parents began receiving AI-generated, expert-validated feedback on their shared book reading sessions, both the quality of their teaching strategies and their children’s communication improved substantially. The findings offer some of the first experimental evidence that AI can meaningfully augment, rather than replace, the human professionals who support families of autistic children.

The research addresses a stubborn gap in early autism intervention. Decades of work have shown that parent-mediated interventions, in which parents learn to embed evidence-based communication strategies into everyday routines, can dramatically increase the intensity of intervention a child receives and help skills generalize across settings. But those interventions live or die on implementation fidelity, the degree to which parents actually deliver the strategies as intended. Traditional coaching models depend on scheduled sessions with specialists, and between those sessions parents are often left without feedback, performance monitoring, or support. The delays between a parent’s performance and a professional’s correction can blunt learning, and the demands of daily family life frequently erode strategy use over time. The research team, led by Yusuf Akemoglu, set out to test whether AI could close that feedback loop.

PaiCoach is built on the Parent-implemented Communication Strategies-Storybook program, an intervention with a strong empirical track record in which parents use three naturalistic strategies during shared reading. Modeling involves demonstrating target words or communicative responses for the child. Mand-model combines prompts or questions with language models to elicit a response. Time delay creates opportunities for children to initiate communication by intentionally pausing before offering help. In the PaiCoach architecture, parents record short videos of their reading interactions and upload them to the platform. A backend pipeline combining automatic speech recognition, multimodal video analysis, and large language model processing then identifies instances of each strategy, estimates fidelity scores, and drafts time-stamped feedback. Crucially, no AI output reaches a parent until trained human reviewers verify it, making PaiCoach a human-in-the-loop system rather than an autonomous one.

The underlying AI was trained on the ASD-HI benchmark dataset, which contains 478 expert-labeled parent strategy-use instances drawn from 48 real-world shared reading sessions. The dataset was partitioned into training, validation, and test sets for model development. For the version evaluated in this study, the researchers replaced the original Whisper speech recognition component with Qwen3 and upgraded the classification and feedback model from GPT-4o to GPT-5.1. Internal evaluation showed the strategy detection model achieved roughly 82 percent recall, while the multimodal fidelity assessment model reached 68.6 percent accuracy. Strategy classification relied on observable interaction sequences, such as recognizing a time delay when a parent created a communication opportunity and deliberately paused, rather than attempting to judge which strategy a parent should have chosen.

To test the system rigorously, the team used a concurrent multiple-baseline across participants single-case experimental design, a method that allows researchers to demonstrate functional relations at the individual level by staggering the start of intervention across participants. Four mother-child dyads took part, each with a minimally verbal autistic child between 45 and 56 months of age. After stable baselines of three to seven sessions in which parents read as they normally would, families completed a one-hour Zoom training on the three communication strategies and then began using PaiCoach for six to seven intervention sessions, followed by maintenance probes roughly three weeks later. All procedures were conducted remotely across approximately 10 to 12 weeks.

The results were striking. Visual analysis showed immediate and sustained increases in parent implementation fidelity for all four mothers following the introduction of PaiCoach. Parent 3, for example, went from 0 percent fidelity across every baseline session to 76.4 percent in her very first intervention session, eventually stabilizing between 74.3 and 88.2 percent. Tau-U effect sizes, which quantify nonoverlap between baseline and intervention data, reached 1.00 for parent fidelity across all four dyads, indicating complete separation between phases. Child communicative responsiveness, measured as the percentage of parent-provided opportunities that drew a verbal or nonverbal response, rose in parallel, with Tau-U values ranging from 0.72 to 1.00 and a mean of 0.84, representing large to very large effects. Gains generally held during maintenance, even after feedback features were switched off.

The human validation data offer a candid look at the current state of AI in this domain. Across 53 reviewed sessions and 879 coded instances, about 69 percent of AI-generated outputs were accepted without modification while 31.4 percent required expert editing. Error rates varied by task: time detection required editing most often, followed by strategy detection, fidelity scoring, and feedback generation. The system occasionally missed gesture-based child responses when a book obscured the camera view, and overlapping speech or natural variation in parental praise sometimes caused fidelity scores to be downgraded incorrectly. Even so, the efficiency gains were considerable. Manually coding a five-minute session typically takes an experienced coder 20 to 30 minutes and a newly trained coder up to an hour. PaiCoach produced transcripts, preliminary coding, scores, and draft feedback within minutes, letting reviewers verify and correct results in roughly 5 to 10 minutes per session.

Parents themselves responded enthusiastically. In semi-structured interviews, all four mothers described wanting their children to communicate more and to engage more deeply with books, and several noted that the AI’s constant availability and nonjudgmental nature made it easier to seek and absorb feedback. One parent described the system’s comments as instructive criticism, while another emphasized the flexibility of uploading videos late at night before bedtime, when traveling to appointments would have been impossible. Quantitative measures backed up the interviews: mean ratings for training clarity, procedure feasibility, and perceived usefulness hovered between 4.3 and 5.0 on a 5-point scale, and the System Usability Scale yielded a mean score of 77.5, indicating good usability. Parents did report friction with video upload navigation, a reminder that user interface design matters as much as the underlying models.

What distinguishes PaiCoach from most AI tools in autism research is its indirect pathway of influence. Rather than delivering therapy directly to children, the system improves the learning environment by sharpening the adults around them. That design choice aligns with human-centered AI frameworks, which hold that effective systems should augment human decision-making rather than supplant it. The authors are careful to frame their findings as preliminary evidence for the feasibility of AI-augmented coaching, not proof of an autonomous intervention, noting that child communication behaviors proved harder for the AI to classify reliably than parent strategies.

Future work, the researchers say, should focus on expanding training datasets to capture greater variability in real-world interactions, improving multimodal models that integrate visual, auditory, and linguistic data, and exploring confidence-based thresholds or hybrid review models that apply human oversight selectively. Larger and more diverse populations, additional daily routines, and head-to-head comparisons with traditional coaching models are all on the agenda. If those efforts succeed, systems like PaiCoach could help solve one of early autism intervention’s most persistent problems: making expert-quality coaching available to every family, at the moment it is needed, inside the routines of everyday life.

Shared book reading offers a particularly strategic setting for this kind of intervention. The routine is already familiar to most families, occurs naturally several times a week, and provides repeated, predictable opportunities for children to request, label, and respond within a low-pressure context. Embedding communication strategies into an activity parents already perform reduces the burden of adding new obligations to crowded family schedules, which has long been a barrier to sustained strategy use outside formal sessions.

The single-case methodology deserves note as well. Unlike group designs that average outcomes across many participants, a concurrent multiple-baseline design evaluates each dyad as its own control, staggering intervention onset while baselines remain stable. This approach is well suited to early-stage feasibility research because it documents individual change patterns and demonstrates that improvements track the introduction of the system rather than maturation or practice alone. Tau-U effect sizes complement visual inspection by quantifying the degree of nonoverlap between phases, with values approaching 1.0 indicating that intervention-phase data points consistently exceeded baseline performance.

The maintenance findings carry practical weight. When feedback features were disabled for follow-up probes, fidelity gains largely persisted, suggesting that the AI-supported coaching period was sufficient for parents to internalize the strategies rather than becoming dependent on continuous prompting. Whether such durability extends to longer intervals and to other routines remains an open question.

The study also illustrates a broader trend in assistive technology research: benchmark datasets of expert-labeled behavioral instances are becoming foundational infrastructure for building and evaluating multimodal models in autism intervention. Because these systems must interpret subtle, context-dependent human behavior, transparent reporting of detection accuracy, editing rates, and failure modes, as this study provides, will be essential for building trust among clinicians and families as AI-supported coaching matures.

Subject of Research: An AI-supported parent coaching system for improving communication outcomes in minimally verbal autistic children

Article Title: Parent’s AI Coach (PaiCoach): A Single-Case Experimental Study of an AI-Supported Parent Coaching System for Autistic Children

Article References: Akemoğlu, Y., Li, Z., Zheng, Q., Roberts, E. B., Singh, S. P., & Xiong, J. (2026). Parent’s AI Coach (PaiCoach): A Single-Case Experimental Study of an AI-Supported Parent Coaching System for Autistic Children. Journal of Autism and Developmental Disorders. https://doi.org/10.1007/s10803-026-07519-6

Image Credits: AI Generated

DOI: 10.1007/s10803-026-07519-6

Keywords: artificial intelligence, autism, parent-mediated intervention, PaiCoach, shared book reading, human-in-the-loop, single-case experimental design, large language models, early intervention, communication strategies, implementation fidelity, multimodal AI