The author developed the protocol a priori and reported the review in line with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidance to ensure transparent identification, screening, eligibility, and synthesis procedures ([40]). The PRISMA 2020 statement represents a revised guideline for reporting systematic reviews that adds more transparency and completeness to reporting. Since the study could not do a quantitative meta-analysis because of the heterogeneity of interventions, outcome measures, and the populations of participants, the author adhered to the Synthesis Without Meta-analysis (SWiM) guidelines in organizing narrative and thematic synthesis and specifying study selection, grouping, and weighting criteria ([8]). The PICOS approach was used to frame the eligibility criteria, which is a method of supporting transparent and reproducible selection reasoning in health and education reviews ([49]). The PICOS framework has been confirmed as a useful tool in enhancing the searching of clinical databases to conduct systematic reviews. The review protocol was developed a priori but was not prospectively registered in a public database (e.g., PROSPERO).

The selection of the study was based on a preschool-special education PICOS framework. The author selected peer-reviewed studies published in English that reported child-level outcomes among primarily preschool-aged children with developmental disabilities, with inclusion of studies extending into early school-age where relevant, published between January 2010 and January 2026. The significance of early intervention on communication, social, and cognitive development was the reason why autism spectrum disorder, language delay, and global developmental delay were the target conditions. Interventions also demanded an AI element that was directly connected to child development, such as socially assistive robots, adaptive or intelligent tutoring systems, speech-recognition-based applications, augmented or virtual environments, or intelligent agents. Comparators that were eligible included usual instruction, standard therapy, or low-technology assistance. The outcomes were required to respond to at least one of the target domains: communication, social interaction, or cognition and were measured with standardized instruments or structured observational scales. Single-group pre-post designs were acceptable provided that they employed a set of quantitative outcome measures that were capable of tracking a change over time. The studies that only provided qualitative feedback on the feasibility or usability without quantitative results at the level of children were eliminated. These parameters represent the protocol that has been completed regarding screening.

An interdisciplinary search was conducted in ERIC, PsycINFO, Scopus, Web of Science, PubMed, and IEEE Xplore databases. Search strings were a combination of intervention terms, population terms, and outcome terms with adaptations to databases in line with PRISMA-S extension requirements of reporting literature searches in systematic reviews ([43]). The terms used as illustrations were artificial intelligence, machine learning, social robots, intelligent tutoring, early childhood, preschool, kindergarten, social skills, social communication, cognition, and learning outcomes. The database was narrowed down to English-language publications published since January 2010. The final database search was conducted in January 2026. The author reported the plan to achieve good reporting practices of information retrieval of reviews. The search strategy and indexing changes were normalized across the databases to achieve sensitivity with maximum preservation of precision. The preliminary search in the six databases had 1472 records. Complete database-specific search strings are provided in Appendix A. Appendix A contains the complete database search of all the six databases.

All records were exported into a Microsoft Excel spreadsheet for reference management purposes. The duplicates were determined by sorting by authors, year, and titles and eliminated manually (n = 186), so the final record was 1286 unique records. Study selection was conducted by the author using predefined PICOS criteria. To enhance methodological rigor, a second independent reviewer verified a randomly selected subset (20%) of records at both the title/abstract and full-text screening stages. Discrepancies were resolved through discussion, with consensus achieved in all cases. Inter-rater agreement for the subset was high (percentage agreement = 91%; Cohen’s κ = 0.84), indicating strong reliability of the screening process.

Following title and abstract screening, 412 records were excluded. The remaining 874 articles were retrieved for full-text review. Of these, 843 articles were excluded due to wrong population (n = 287), wrong intervention (n = 198), wrong outcomes (n = 156), wrong study design (n = 112), duplicate reports (n = 54), and lack of data (n = 36). A total of 31 studies met the inclusion criteria and were included in the final synthesis.

To enhance clarity and reproducibility of the synthesis, key coding terms were defined as follows:

AI Modality: The form of artificial intelligence technology used within the

Figure 1. PRISMA 2020 flow diagram for study identification, screening, eligibility, and inclusion.

study (e.g., socially assistive robots, conversational agents, adaptive tutoring systems, augmented/virtual reality platforms).

Implementation Features: Characteristics related to how the intervention was delivered, including session duration, frequency, setting (e.g., home, clinic, classroom), and involvement of adults (e.g., teacher, therapist, caregiver).

Human-AI Orchestration: The manner in which human agents interact with and support AI systems during intervention delivery, including levels of prompting, scaffolding, supervision, and co-participation.

A structured data extraction form captured bibliographic details, country, participant characteristics, disability categories, intervention characteristics, AI modality, comparator details, outcome domains, instruments, and key statistics. The data extraction form is provided in Appendix B. Data extraction was conducted by the author using a structured extraction form. To ensure reliability, a second reviewer independently verified a subset (20%) of extracted studies, with discrepancies resolved through discussion. The extraction fields and processes adhered to the Cochrane Handbook of Systematic Reviews of Interventions in the generation of evidence that is reproducible ([14]). The summary tables were developed to present the evidence by outcome domain (Table 1) and evidence quality by domain (Table 2). Sample sizes were recorded as reported by study authors; where sample sizes were not reported, this was indicated as “NR” rather than imputed.

Methodological quality was assessed using design-appropriate tools. For randomized controlled trials (RCTs), the author applied the revised Cochrane Risk of Bias 2 (RoB 2) tool, which assesses bias in 5 domains, including randomization process, deviations in intended interventions, missing outcomes data, measurement of outcomes, and reporting results ([53]). The RoB 2 tool suggests a different evaluation of risk of bias of randomized trials. In non-randomized trials, such as quasi-experimental designs, single-case experimental designs, and cohort studies, the author relied on Joanna Briggs Institute (JBI) Critical Appraisal Checklists, which touch upon the aspects of confounding, comparability of the participants, measurement validity and reliability, and the appropriateness of the analysis ([1]). In the single-case experimental design, standards of the What Works Clearinghouse (WWC) were also used. General scores on low, moderate (some concerns), or high risk of bias were determined. The credibility of effects in synthesis was informed using domain-level judgments. A single, unified master risk of bias table was developed to provide all studies with a single, consistent rating with the same rating that was applied to all tables, appendices, and narrative text. All domain-level quality appraisal results for the 31 studies are presented in Appendix C.

For randomized controlled trials, the RoB 2 domains were defined as follows:

D1 = bias arising from the randomization process;

D2 = bias due to deviations from intended interventions;

D3 = bias due to missing outcome data;

D4 = bias in measurement of outcomes;

D5 = bias in selection of the reported result.

To enable comparison across study designs, a unified overall risk-of-bias rating was applied: studies were classified as low risk when all domains were rated low; moderate risk (“some concerns”) when at least one domain raised concerns; and high risk when one or more domains indicated high risk or when studies lacked sufficient methodological rigor (e.g., feasibility or framework studies without outcome data).

Given the substantial heterogeneity across AI modalities, implementation parameters, and outcome measures, a meta-analysis was not feasible. Thematic and narrative synthesis was done to combine the findings of communication, social interaction, and cognition. This methodology entailed the use of iterative coding to extract descriptive and explanatory themes associated with effectiveness and context of implementation with special focus on patterns by modality, session structure, and setting. Where the effect sizes and statistical significance were reported, they were extracted to enhance interpretation. The synthesis expressly utilized the methodological quality in balancing the evidence and the other identified consistencies, divergences, and gaps to inform future research ([14]). Synthesis of evidence was used to determine patterns within AI technologies, such as social robotics, adaptive tutoring systems, augmented reality platforms, speech recognition tools, and conversational AI in every outcome domain. Table 1 provides the outcome domain summary findings.

Table 1. Summary of evidence by outcome domain.

Note. RJA = responding to joint attention; IJA = initiating joint attention; ToM = Theory of Mind; SAR = socially assistive robot; AR = augmented reality; PAND = percentage of all non-overlapping data; MBLR = mean baseline reduction; Tau-U = non-parametric effect size for single-case designs. Consistency ratings: High = findings consistent across ≥ 75% of studies; Moderate = consistent across 50% - 74%; Mixed = inconsistent findings; Limited = insufficient studies. Quality ratings based on study designs, risk of bias, sample sizes, and replication across studies. Some studies appear in multiple domains if they measured outcomes across categories.

Table 2. Evidence quality summary by domain.

Note. Evidence quality ratings integrate study design rigor, risk of bias assessments, consistency of findings across studies, magnitude of effect sizes, and sample sizes. the most consistent evidence was observed for requires multiple well-designed studies with consistent findings and low risk of bias. RCT = randomized controlled trial; SSED = single-subject experimental design.

The initial database search retrieved 1472 records (ERIC = 312; PsycINFO = 287; Scopus = 298; Web of Science = 245; PubMed = 198; IEEE Xplore = 132). The title and abstract screening was done on 1286 unique records after eliminating 186 duplicates. Out of them, 412 were eliminated due to irrelevance to the review scope, and 874 records were fully assessed in terms of eligibility for full-text review. After careful consideration of the articles in terms of PICOS criteria, 843 articles were excluded due to the following reasons: wrong population (n = 287), wrong intervention (n = 198), wrong outcomes (n = 156), wrong study design (n = 112), duplicate reports (n = 54), and inadequate data (n = 36). Out of the total number of studies, 31 articles were included in the final synthesis because they all fulfilled the inclusion criteria. The flow diagram of PRISMA (Figure 1) is the example of the entire screening process. Verification: 1472 − 186 = 1286; 1286 − 412 = 874; 874 − 843 = 31. The further exclusion categories and counts are given in Appendix C.

The 31 included studies were published between 2013 and 2026 and conducted across diverse countries and geographical regions including the United States, Chinese Mainland, Japan, Italy, Spain, Greece, Iran, Kazakhstan, France, Hong Kong SAR, Taiwan Region, South Korea, Belgium, Serbia, Australia, Algeria, Chile, Colombia, Lebanon, and the United Kingdom. The sample sizes were also quite different, including single-case designs consisting of 4 - 6 participants ([25]); [52]; [41] and quasi-experimental cohorts of 193 participants ([30]). The target group was mostly preschool-aged children with autism spectrum disorder (ASD), but some studies used children with attention-deficit/hyperactivity disorder (ADHD), developmental delays, language disorders, or mixed groups of typically developing (TD) and neurodiverse children. Detailed study characteristics are presented in Table 3.

Table 3. Characteristics of included studies (N = 31).

Intervention characteristics, including technology type, comparator, and duration or dosage, are summarized in Table 4.

Table 4. Intervention characteristics.

The outcome measures used across the included studies are presented in Table 5.

Table 5. Outcome measures.

Note. ASD = autism spectrum disorder; TD = typically developing; ADHD = attention-deficit/hyperactivity disorder; ID = intellectual disability; DD = developmental delay; DSD = delayed speech development; GQ = general quotient; DA = developmental age; NVIQ = nonverbal IQ; NR = not reported; M = mean; RCT = randomized controlled trial; SSED = single-subject experimental design; ATD = alternating treatments design; UCD = user-centered design; ML = machine learning; SAR = socially assistive robot; AR = augmented reality; FER = facial expression recognition; ARIA = adaptive robot-mediated intervention architecture; NORRIS = non-contact responsive robot-mediated intervention system; RISTA = robot-mediated imitation skill training architecture; RSE = robotic social environment; CARS-2 = Childhood Autism Rating Scale-Second Edition; NEPSY-II = Developmental Neuropsychological Assessment-Second Edition; AF = Affect Recognition; ToM = Theory of Mind; IN = Inhibition; CBCL = Child Behavior Checklist; TRF = Teacher Report Form; SDQ = Strengths and Difficulties Questionnaire; ESCS = Early Social Communication Scales; SDS = standard difference score; RSDS = restrained standard difference score; RDDS = restrained duration difference score; RT = response time; ADOS-2 = Autism Diagnostic Observation Schedule-Second Edition; SCQ = Social Communication Questionnaire; MSEL = Mullen Scales of Early Learning; VABS-II = Vineland Adaptive Behavior Scales-Second Edition; WIAT-II = Wechsler Individual Achievement Test-Second Edition; RJA = responding to joint attention; IJA = initiating joint attention; LCS = longest common subsequence; LTM = least-to-most; ADI-R = Autism Diagnostic Interview-Revised; ABAS-II = Adaptive Behavior Assessment System-Second Edition; MB-CDI = MacArthur-Bates Communicative Development Inventories; ESES = Educators’ Sense of Efficacy Scale; MLU-M = mean length of utterance in morphemes; NDW = number of different words; TNW = total number of words; CLAN = Computerized Language Analysis; FSM = finite state machine; FPFT = first positive feedback time; KBIT-2 = Kaufman Brief Intelligence Test-Second Edition; TOLD = Test of Language Development.

Various AI-mediated interventions were conducted, most of which employed social humanoid robots e.g., NAO ([4]; [9]; [50]); CommU ([28]); Pepper ([7]); and Kebbi) and specialized adaptive systems (e.g., ARIA, NORRIS, and RISTA). The average sessions were 25-30 minutes, with the majority of the studies providing 6-21 sessions. The frequency of interventions ranged from single assessment sessions to 30-day daily sessions. Half of the studies involved caregivers, therapists, or teachers. The three studies published in 2026 also used the latest technological innovations, such as conversational AI that can better implement peer-mediated intervention, socially assistive robots to improve attention, and the use of AI-based recommenders of educational activities.

3.3.1. Communication Outcomes

Across studies targeting communication, AI-based tools improved joint attention, imitation, and related foundational behaviors ([41]; [50]). Numerous robot-mediated methods showed some impressive improvements of joint attention, such as the experiments with autonomous humanoids, which minimized caregiver prompting and maximized triadic interaction ([48]; [62]). According to [62] and [48], the results indicated an increase in orienting, shared attention, and early communicative behaviors in preschoolers.

Imitation was a frequent target. Controlled comparisons had indicated greater improvement in imitation and attention under conditions where the practice was mediated by robots as compared to that of human instructors ([60]), and closed-loop or semi-autonomous systems promoted better gesture learning and maintaining attention on tasks ([6]). An RCT study by [28] (n = 68 children, 30 ASD, 38 TD) found significant Time x Group interaction in the robot group (F (1, 26) = 11.45, p < 0.01), with half of the children showing an improvement in joint attention with human partners after interacting with the robot—a carryover effect (F). [6] revealed that children with ASD glanced at the robot 52.76 percent of the session time compared to 25.11% when seen by a human therapist (p < 0.005), and the ARIA system attained a success rate of 95.83% in the tasks attempted.

Language and speech outcomes were addressed in newer studies. [18] tested the RASA robot, which supports speech therapy using lip-syncing, with 12 children who have language disorders. There was also a much greater improvement in oral vocabulary (p = 0.007), syntactic understanding (p = 0.007), and overall language ability (p < 0.001) in the robot-assisted group, and the effect size of engagement was d = 1.66. [20] established that conversational AI-enhanced peer-mediated intervention promoted conversational initiations (MBLR 223% 598) and responses (Tau-U = 1.00) and skills transferred to new interaction partners. The implementation of the augmented reality intervention among six children by [41] revealed that PAND = 98% in regard to gaze following, and the skills were maintained at a one-month follow-up.

3.3.2. Social Outcomes

Social outcomes were the most extensively studied domain ([23]). Humanoid robots supported turn-taking, emotion identification, and broader psychosocial engagement ([34]). [23] conducted an RCT with 51 children (44 completed) and found that both robot-assisted and human-only therapy yielded significant improvements in CARS-2 scores, NEPSY-II social perception, and inhibition. The only significant between-group difference favored the robot condition on SDQ Prosocial Behavior (t (42) = 2.457, p = 0.014). [34] documented significant gains in emotion recognition, perspective-taking, and social situation understanding across 10 sessions with 14 children.

Two studies suggested maintenance and generalization. [52] used a fully robotic social environment where six children observed affective exchanges between two robots. All participants achieved 80%-100% accuracy in emotion recognition, with skills maintained and generalized at one-month follow-up. [3] employed a NAO robot for gesture training in 10 preschoolers with ASD; the subgroup advancing through training (n = 4) showed significant improvement in ABAS-II Social Adaptive Domain at six-month follow-up (p = 0.022). [51] directly compared robot-based versus human-based interventions for 38 children with high support needs, finding that only the robot-based intervention produced significant and durable improvements in joint attention.

[46] and [22] conducted a cross-cultural analysis of 36 children from Japan and Serbia and found that engagement was significantly correlated with affective valence-arousal (r = .56-.73). [22] demonstrated that structured robot activities significantly improved initiation of social interaction (β = 4.64, t₂₃ = 2.39, p = 0.026). [42] showed that individualized, adaptive approaches successfully maintained engagement across 7-10 sessions in 11 children with ASD/ADHD, with familiar activities yielding higher engagement than unfamiliar ones.

3.3.3. Cognitive Outcomes

Several studies examined cognitive outcomes including attention, memory, and academic skills. Evidence for cognitive targets such as attention, working memory, problem solving, and early academic skills was positive but more tentative than that for social-communication domains ([13]). [13] provided 17 children with socially assistive robots in daily in-home sessions over an average of 41 days, observing sustained engagement (65% average) and significant WIAT-II gains (d = 0.54, p < 0.01) in numerical operations and math reasoning. [25] demonstrated dramatic improvements in attending behavior: one participant showed an 85-fold increase in eye contact duration (0.4s → 91.4s), with weighted average Tau-U = 0.930 (p < 0.001).

[30] involved 193 preschoolers comparing AI robot teaching to teacher-led instruction for household tasks. Both instructional groups significantly outperformed the no-instruction control, with the robot-assisted group showing slightly better outcomes. A quasi-experimental study involving an AI teaching robot reported better performance on home assignments than on teacher-led activities ([31]). [64] developed an AI-driven educational activity recommender achieving 94% accuracy in matching activities to developmental needs, with 100% of clinical participants agreeing that mood-based recommendations were beneficial.

[4] examined robot action sequencing strategies in 11 children with ASD and found that personalized Therapy mode sequences achieved high task success rates compared to random Explore mode sequences. [9] found that while both ASD and TD children performed better with human partners than robots overall, ADOS severity predicted robot-mediated performance (R² = 0.27), suggesting that intervention effects may vary by symptom severity. Exploratory implementations using virtual agents or chat-enhanced platforms have demonstrated high engagement ([7]). However, they were limited by small samples, scarce autism-specific data, and the absence of rigorous controls, highlighting the need for larger, theory-driven trials ([7]).

3.3.4. Teacher and Program Support

Several studies examined the use of AI to support screening, assessment, and classroom integration ([16]; [32]; [36]; [38]). A robot-administered autism screener produced outcomes aligned with caregiver reports, suggesting the potential for scalable early identification ([45]). Short, structured robot activities integrated into preschool classrooms support social communication without disrupting routines ([22]). Eye-tracking paired with robot-supported testing differentiated joint attention profiles associated with cognition and autism severity ([15]). These tools may alleviate educators’ workload and augment program evaluation, although direct child-level gains were mixed ([44]).

A study-level summary of outcome domains and primary outcomes is provided in Table 6.

Table 6. Summary of findings for included studies (N = 31).

The main findings reported by each included study are summarized inTable 7.

Table 7. Key findings.

As shown in Table 8, the included studies varied in reported effect sizes, statistical indicators, and evidence of maintenance or generalization.

Table 8. Effect sizes and maintenance.

Note. JA = joint attention; RJA = responding to joint attention; IJA = initiating joint attention; MBLR = mean baseline reduction; PAND = percentage of all non-overlapping data; Tau-U = non-parametric effect size for single-case designs; SDS = standard difference score; RSDS = restrained standard difference score; RT = response time; LCS = longest common subsequence; LTM = least-to-most; LTM-RI = least-to-most robot intervention; FPFT = first positive feedback time; RL = reinforcement learning; NN = neural network; CAI-PMI-ID = conversational AI-enhanced peer-mediated intervention by peers with intellectual disabilities; PMI-ID-only = peer-mediated intervention by peers with intellectual disabilities without AI; RBI = robot-based intervention; HBI = human-based intervention; RAST = robot-assisted speech therapy; NVIQ = nonverbal IQ; CARS = Childhood Autism Rating Scale; SAD = Social Adaptive Domain; GQ = general quotient; ESES = Educators’ Sense of Efficacy Scale; MLU-M = mean length of utterance in morphemes; NDW = number of different words; TNW = total number of words; N/A = not applicable.

Methodological quality was heterogeneous. Of the 31 included studies, 11 (35.5%) were rated as low risk of bias, including RCTs by [23] and [28], as well as methodologically rigorous single-case experimental designs meeting WWC standards by [41], [20], [52], [48], and [25]. Fifteen studies (48.4%) were rated at moderate risk (some concerns) due to small samples, limited blinding, quasi-randomization, or single-group designs. Five studies (16.1%) were rated at high risk, primarily technical feasibility or design framework studies lacking clinical outcome data ([7]; [29]; [47]). Common methodological limitations included small sample sizes, lack of long-term follow-up, and heterogeneous outcome measures. The strongest evidence supported improvements in joint attention, imitation, and emotion recognition. Cognitive outcomes, while promising, had weaker evidence due to fewer controlled trials. Complete risk of bias assessments are presented in Table 9.

Table 9. Risk of bias assessment for included studies; study design, tool and overall risk (N = 31).

Table 10. Risk of bias domains.

Notes: RoB 2 = Cochrane Risk of Bias 2 tool; JBI = Joanna Briggs Institute; WWC SSED = What Works Clearinghouse Single-Case Design; N/A = not applicable.

Table 11. Summary of risk bias ratings.

Note. Low risk studies included well-designed RCTs, rigorous single-subject experimental designs meeting WWC standards, or comparative studies with objective measures and adequate controls. Moderate risk studies had methodological limitations such as small samples, lack of blinding, pseudo-randomization, high attrition, or absence of control conditions. High risk studies were primarily technical feasibility or design framework papers without clinical outcome data.

In addition to intervention studies reporting direct child-level outcomes, several included studies focused on screening, assessment, system validation, or framework development. These included eye-tracking diagnostic tools, machine learning validation studies, and conceptual or feasibility frameworks.

While these studies provide important insights into the broader ecosystem of AI applications in early childhood and special education—particularly in areas such as early detection, progress monitoring, and system design—they do not constitute direct evidence of intervention effectiveness. Accordingly, these studies were analyzed separately and were not included in conclusions regarding intervention efficacy.

The strength of this review is its rigorous, prespecified methodology: PICOS-aligned eligibility, dual screening and extraction, and design-appropriate appraisal (Cochrane RoB 2; JBI). To ensure transparency and replicability, these steps, which are described in the Methods section and illustrated in the PRISMA flow (Figure 1), are applicable. The evidence base of mixed methodology included a variety of modalities of AI (social robotics, adaptive systems, AR/virtual agents) and environments, which facilitated a synthesis of practices (Table 1). The detailed appendices (A-D) give complete disclosure of search methods, extraction methodologies, quality review, and decision to rule out.

Limitations mirror the field’s developmental stage: few large RCTs, small samples, brief interventions (often ≈10 sessions), limited to no long-term follow-up, and heterogeneity of measures, which makes synthesis and generalization difficult. The limitation to English-language literature might have left out pertinent studies. These limitations, as indicated by the quality appraisal (Table 4), put the risk of publication bias higher and the overgeneralization of promising initial effects at risk ([7]; [50]).

AI-based interventions seem to be the most effective when specific, clearly defined skills, e.g., joint attention, imitation, or emotion recognition, are targeted. Sessions of 20 - 30 minutes with a regular schedule (23 times a week) ensure maximum activity. The role of adults will not lose its significance; the use of AI tools must complement human teaching. Response to intervention is predicted by individual differences in baseline skills, which imply the necessity of individual assessment ([42]; [51]).

The next phase should prioritize: 1) adequately powered, multi-site RCTs with longer duration and follow-up to test retention and cross-context generalization; 2) multilingual, culturally adapted systems, such as Arabic speech/NLP, and culturally aligned social scripts validated with local communities; and 3) co-designed pipelines linking educators, therapists, and developers to embed pedagogical/therapeutic principles from the outset. These directions follow directly from the gaps identified in our synthesis (Table 4) and selection overview (Figure 1).