A four-family AI-supported coaching study showed promising shared-reading changes under continuous expert review

AI may help organize feedback about a parent-child learning interaction, but interpreting that interaction requires care. Yusuf Akemoğlu and colleagues evaluated PaiCoach as a support for parents during shared book reading with young autistic children. The system analyzed short videos and drafted feedback connected to specific moments. Trained reviewers checked the outputs before families could access them. The study therefore concerns an AI-assisted professional workflow, with human interpretation built into the intervention rather than added only when something appeared to go wrong.

Four mother-child pairs participated. The children were 45 to 56 months old and used limited expressive language, while continuing their existing support services. A concurrent multiple-baseline design staggered the introduction of the intervention across families. Each pair completed several baseline sessions, six or seven intervention sessions and one or two maintenance observations, over approximately ten to twelve weeks. During baseline, parents recorded ordinary shared reading. The intervention began with an hour of live online training and continued with video-based coaching through PaiCoach.

The researchers measured how consistently parents implemented the targeted communication-support strategies and how children responded during the observed reading routines. Visual analysis showed increases after the coaching package was introduced, with corresponding improvements in child responsiveness. Maintenance observations occurred about three weeks later with feedback disabled, and offered preliminary evidence that some changes remained. Staggering intervention onset strengthens an individual-level analysis compared with a single before-and-after observation, but four families still provide a narrow basis for generalization.

The human-review results are central to interpreting the technology. Across 53 reviewed sessions and 879 coded instances, 31.4 percent of AI-generated outputs required editing. Reviewers corrected problems involving timing, strategy identification, fidelity scoring and feedback. Child communication was especially difficult to classify when gestures were obscured or speech overlapped. These are consequential details in a setting where a mistaken interpretation could change the advice a family receives. The published outcomes were based on the reviewed system, not on uncorrected AI output delivered automatically to parents.

The study cannot separate the effects of initial live training, continued expert-validated feedback, the platform and the AI analysis itself. It also does not establish superiority to conventional coaching, benefits across other routines or sustained changes beyond the brief maintenance period. All participating parents were mothers, and the children and contexts represented only a small part of the diversity of autistic communication. Reported workflow efficiencies came from implementation experience rather than a randomized comparison of staff time. The appropriate conclusion is promising preliminary evidence for a supervised coaching package with substantial ongoing review requirements.

For AIEDHK, the broader research lesson is to report the whole human-AI workflow and make its review burden visible. Future evaluations could compare coaching approaches, include more families and communication styles, and test whether observed changes extend to meaningful everyday participation. They should also examine which AI errors persist and how corrections affect feedback. These are proposed research directions beyond the study, not guidance to replace a family's existing support. The opportunity lies in assisting professional observation and family learning while preserving human judgment about each interaction and keeping claims proportional to the evidence.