LLMs and other GenAI approaches have demonstrated promising potential in enhancing ASD screening and diagnostic processes. This category encompasses studies utilizing GenAI models to support early ASD screening, assisting clinical decision-making, or derive meaningful insights from clinical datasets. Specifically, it includes diagnostic classifiers that use behavioral or linguistic features, intelligent rating systems to streamline or automate standard assessment tools, and predictive models that estimate ASD risk from diverse data sources. Additionally, GenAI can augment limited clinical samples by producing synthetic datasets or simulate diagnostic scenarios (43).

Mukherjee et al. (2023)’s study explores the application of AI models, specifically Bidirectional Encoder Representations from Transformers (BERT) and ChatGPT, in identifying early signs of ASD through parents’ narratives about their children’s behaviors (38). Texts were collected from social networks and ASD support communities, then labeled as positive (ASD-related) or negative. BERT, a transformer-based model, was fine-tuned for binary classification, achieving 83% accuracy with precision scores of 0.84 (negative) and 0.87 (positive), and F1-scores of 0.85 and 0.79 respectively. Although the study does not provide explicit numerical performance metrics for ChatGPT’s ‘text-davinci-003’ model, which was trained using reinforcement learning from human feedback, it reports that the model demonstrated consistent performance on new data. For positively labeled sentences, the system calculated cosine similarity with a curated library of ASD symptom statements to identify specific challenges, such as speech delay or poor eye contact, enabling more personalized intervention recommendations. This approach offers a scalable, non-invasive alternative to traditional diagnostic methods, particularly valuable in resource-limited settings. However, the authors acknowledge several limitations, including a small dataset size (although the article does not specify the exact number of examples, the reported evaluation metrics suggest that only about 80 sentences were used); data-collection procedures and class distribution, which limit reproducibility and transparency; reliance on subjective parent input; a lack of cultural diversity; and potential inaccuracies introduced by AI-generated outputs. Despite these challenges, the study demonstrates the potential of GenAI to enhance early ASD detection. Carefully curated datasets drawn from parents’ lived experiences are essential, as they may capture subtle behavioral cues overlooked by formal screening.

Deng et al. (2017)’s study explored the use of GANs for diagnosing ASD based on speech patterns (34). This method leverages machine learning to identify speech anomalies characteristic of ASD, such as echolalia, atypical prosody, and repetitive phrasing. However, it is widely recognized that the performance and reliability of such systems are constrained by the amount of available data for model training since annotated samples of autistic behavior or language can be scarce, and building large, representative ASD datasets is challenging and costly. To overcome this limitation, GANs were used to synthesize additional data to augment training sets. The authors evaluated the classifier’s performance on the database which includes over 6,380 utterances from 102 children across four categories (autistic disorder, pervasive developmental disorder-not otherwise specified, specific language impairment, and typically developing children) and compared the GAN-based method against three representative conventional models: a linear Support Vector Machine (SVM), an SVM with a Radial Basis Function (RBF) kernel, and a Multi-Layer Perceptron (MLP) with four hidden layers. Results show that the GAN-based approach achieved the best unweighted average recall (UAR) metrics of 44.06%, corresponding to a relative improvement of 10.4% over the Linear SVM baseline. The study concludes that GAN-based representation learning is a promising approach for ASD detection from speech, especially in contexts where data scarcity limits the performance of conventional models.

Another recent study demonstrates that modern LLMs can serve as “data generators” to produce realistic examples of autistic behaviors, which may be invaluable for data-hungry algorithms in diagnosis research. Woosley et al. (2024) investigates whether LLMs can generate realistic textual examples of autistic behaviors to enhance the performance of BERT-based neural networks (41). The study utilized a small corpus of real clinical observation data derived from CDC surveillance reports in the state of Arizona, comprising free-text entries labeled by trained experts. The dataset is highly imbalanced, with only 14.3% of sentences containing a diagnostic label, and the distribution of examples across autism symptom labels is similarly unequal. To address these limitations, the researchers presented a compelling proof-of-concept by prompting GPT-3.5-Turbo and GPT-4 to generate thousands of hypothetical child behaviors consistent with DSM-5 criteria for ASD. Specifically, they created 4,200 synthetic text snippets (e.g. “child repeats phrases from cartoons without understanding”) to represent ASD diagnostic features. A clinician review of a sample of the AI-generated cases found 83% were valid and correctly matched ASD symptoms. When the synthetic data were added to train a BioBERT model (44), the model’s recall improved by 13%. This indicates LLMs can help overcome data scarcity by providing additional examples of ASD manifestations, potentially boosting the sensitivity of screening tools. However, a precision drop in the augmented model (-16%) was noted, highlighting that synthetic data may introduce noise or irrelevant patterns.

Whereas previous studies have highlighted the strengths of GenAI models in ASD screening and diagnosis, recent study have shown that general-purpose generative models like GPT-4V perform worse than ASD-specific models in diagnostic tasks. Zhao et al. (2025) developed AV-FOS, a deep learning model for automatically recognizing interaction styles in ASD children using audio-visual data annotated with the Family Observation Schedule-Second Version (FOS-II) (42). The FOS-II is a validated tool for analyzing parent-child interactions but requires manual coding by trained observers, which is time-consuming and labor-intensive. AV-FOS was designed to address the limitations of manual annotation, while GPT-4V lacks domain-specific tuning. Compared to AV-FOS, GPT-4V (tested with Prompt Version 1 and 2), SlowFast Networks based on the Convolutional Neural Network (CNN) structure, and Vision Transformers based on the transformer structure showed lower accuracy, F1 score, and efficiency. Although the AV-FOS study excluded some rare annotations due to limited data, most omitted annotations had limited clinical significance. Overall, GPT-4V fell short in clinical applicability for behavioral assessment compared to the specialized AV-FOS model.

ASD affects communication, social cognition, and emotion recognition—domains where practice and feedback are crucial. AI technologies are therefore being designed to deliver scalable therapeutic assistance (10, 45, 46). The studies in this section examine GenAI tools that either assess the condition of individuals with autism or scaffold social-behavioral interventions.

Because emotional cues span facial expressions, body language, vocalizations, and even heart-rate signals, multimodal GenAI models can improve recognition accuracy—especially given autistic children’s challenges in expressing and interpreting emotions. Kurian et al. (2024) illustrate this with m_AutNet, a personalized framework that fuses facial and vocal data to recognize emotion in autistic children (36). A CNN-based visual embedder clusters images by similarity, while a transfer-learned CNN extracts audio embeddings; a Wasserstein-tuned GAN then aligns these domains for classification. The performance of the model was tested on the dataset comprising video recordings of 75 children who exhibit stereotypical behaviors such as head-banging, stimming, and spinning. In real-time emotion recognition tests, m_AutNet reached 88.25% accuracy, outperforming prior affect-recognition systems for the ASD children. Although the modest dataset limits generalizability, the results spotlight the value of personalized multimodal modeling in autism-focused emotion recognition.

Across these applications, the unifying aim is to strengthen autistic individuals’ social-behavioral skills through personalized and often interactive GenAI. Approaches range from LLM-based chatbots and socially assistive robots to virtual agents—tools increasingly aligned with digital-therapeutic trends in clinical practice (10, 25).

A clear example is EmoEden (40), which blends LLMs with text-to-image generators to help children with high-functioning autism (HFA) enhance their emotional recognition and expression skills. Parents input preferences so EmoEden can tailor story difficulty and dialogue flow. With personalized conversations and visuals that target common emotional challenges, children could practice identifying emotions, responding to others, and sharing their own feelings. Over 22 days, six HFA participants (8–12 years) showed marked gains in emotion recognition, emotional richness, and context-appropriate vocabulary. Observation records noted high completion rates and active engagement. Parents also described post-training gains, such as heightened empathy and willingness to express discomfort. GenAI allowed EmoEden to expand user input into richer expressions, though occasional unrealistic details emerged. Overall, EmoEden’s blend of personalized storytelling, supportive real-time feedback, and conversational models offers an accessible pathway for emotional learning, with broader implications for at-home training. Despite the promise of automated adaptation and reduced reliance on human labor, the authors stress that such AI cannot supplant therapists and must be carefully supervised to mitigate “hallucinations” or misdirected responses.

Koegel et al. (2025) similarly investigated Noora, a GPT-4-driven tool that delivers real-time feedback on empathetic responses (35). In a four-week randomized controlled trial (RCT) with 30 autistic adolescents and adults (11–35 years), each of the experimental and the waitlist control group completed 200 empathy trials. Each trial consisted of an empathy-inducing prompt, a participant-generated response, and immediate feedback. Findings revealed significant improvements in empathetic communication skills among participants, observed both within Noora practice sessions and generalized to social interactions; 71 percent of the participants in the experimental group showed an improving trend from their first 50 responses to their last 50 responses, with an average improvement of 13.2%. The experimental group showed a mean delta change score (pre- and post-intervention) of 37.67%, whereas the waitlist control group showed a mean improvement of 2.53%. Participants also reported increased confidence and high satisfaction levels regarding Noora. However, several limitations were noted. These included the absence of direct comparisons with traditional face-to-face empathy interventions, occasional discrepancies between Noora’s AI-generated evaluations and human raters’ judgments especially related to grammar or phrasing issues, a relatively small sample size, and the lack of long-term follow-up to assess sustained effectiveness. Despite these limitations, the study provides initial evidence supporting the potential of GenAI-based interventions in effectively enhancing empathetic conversational skills.

Traditional emotion-training methods often force children to stare directly at faces—an aversive task for many autistic learners. VR/AR environments can ease this discomfort (47). However, VR/AR interventions typically offer only pre-scripted scenarios with limited, multiple-choice responses that do not adapt to individual users. Integrating GenAI can bridge this gap between VR/AR practice and real-world application by introducing free-form dialogues that are tailored to each user, thereby assisting individuals with ASD in developing generalized social skills across diverse settings.

Building on this integration of immersive technologies and GenAI, a study by Lyu et al. (2024) introduced EMooly, a tablet-based game integrating GenAI and AR (37). This application aims to enhance social-emotional learning for autistic children through the active involvement of their caregivers. The system comprises five phases: 1) a customization phase, utilizing GPT-3.5/4 to generate personalized social stories for each child; 2) a comprehension phase, where the child read and understands these social stories, accompanied by the caregiver; 3) an observation and imitation phase, involving turn-taking exercises between the child and caregiver to practice mimicking facial expressions; 4) a recognition phase, which prompts children to engage in a dynamic AR activity to identify the target emotion from emotional expressions overlayed in the real-world environment; and 5) a reinforcement phase designed to consolidate the learned concepts with reflective questions. The researchers conducted a between-subjects controlled study with 24 autistic children (mean age: 6.0, 3 girls and 21 boys) and their caregivers, who participated in single-visit sessions at special education centers. The study compared EMooly to a traditional slide deck-based intervention method, serving as a baseline. Results indicated that EMooly significantly enhanced children’s emotion recognition skills and outperformed the baseline method in terms of usability and overall user experience. Specifically, children with EMooly showed larger improvements on quizzes assessing emotion recognition abilities (scored out of 10) from pre- to post-intervention, with an average increase of 1.5 points, whereas the baseline group exhibited a mean decrease of 0.41 points. In sum, GenAI is poised to make VR/AR social skills training more immersive and personalized, but ongoing evaluation will be needed to measure whether these AI enhancements using LLM translate to better real-world social outcomes.

Extending beyond virtual environments, GenAI is also being integrated into physical interaction platforms such as social robots. In particular, social assistive robots (SARs) have long been explored in ASD therapy as engaging and judgement-free interaction partners (48). These robots can generally be divided into two types: retrieval-based and generation-based. Retrieval based models select responses from a predefined repository, limiting their interactions to pre-recorded dialogues. However, when relevant conversational scenarios are absent from the database, autistic children may quickly lose patience or interest. In contrast, generation-based models leverage GenAI, allowing robots to create responses dynamically beyond predetermined scripts, resulting in more natural and flexible conversations. Recently, researchers have explored combining LLMs with social robots, where the LLM manages language interactions and the robot provides visual and interactive engagement, enhancing user experience (46).

One such implementation of this LLM-augmented social robotics approach is demonstrated in the study by She et al. (2021) (39). This study aimed to improve the conversational abilities of the humanoid robot NAO when interacting with autistic children by developing a deep learning-based dialogue model. Using sequence-to-sequence architecture enhanced with attention mechanisms and GloVe word embeddings, the researchers trained the model through transfer learning—initially on dialogues from typically developing children, then fine-tuned with dialogues from autistic children. This model was integrated into the NAO robot and evaluated against several baseline models, including a standard Seq2Seq model and a GAN-based conversational agent (GCA), with and without BERT embeddings. The proposed model outperformed all baselines across multiple automated metrics, including BLEU score (0.23 vs. 0.15 for Seq2Seq), Greedy Matching, Embedding Average, Vector Extrema, and Skip-Thought similarity. It also achieved better semantic coherence and word distribution similarity with real conversational data, as reflected in lower KL divergence and Earth Mover’s Distance. Human evaluations conducted by 12 autism-experienced raters confirmed these results, with the proposed model scoring highest for single utterance quality (3.05) and overall script quality (3.23), compared to GCA (2.82 and 2.87) and Seq2Seq (1.89 and 1.83). These findings demonstrate that the model substantially enhances the naturalness, contextual relevance, and appropriateness of robot-generated conversations for children with ASD.

Another relevant work involved integrating the Pepper robot with ChatGPT to facilitate natural, open-ended dialogues with autistic individuals in real time (49). This study presented two different scenarios to leverage the robot’s capabilities to enhance communication, social skill development, and problem-solving abilities: 1) an informal interaction scenario focused on building rapport within a relaxed, comfortable environment; and 2) a structured interaction mimicking a psychoeducational setting, in which the robot posed problems for the users to solve. Although promising, this study primarily served as a feasibility demonstration, highlighting possibilities rather than confirming effectiveness through comprehensive evaluation.

Given the vast amount of online discourse surrounding ASD, ranging from scientific findings to misinformation, parents often struggle to find reliable resources about ASD. In response to this challenge, an LLM-based assistant has recently been proposed to provide on-demand, accurate information and guidance tailored to caregivers’ questions about ASD.

A recent study by He et al. (2024) assessed the effectiveness of LLM chatbots in addressing ASD-related questions (33). The researchers collected a total of 239 consultation queries posted by 100 randomly selected autistic individuals or their families from a web-based medical consultation platform in China. The answers were newly generated using OpenAI’s ChatGPT-4 and Baidu’s ERNIE Bot, which were then compared to answers previously written by human physicians. A panel of three chief physicians conducted evaluations of each answer. Evaluators preferred physician responses over those from chatbots. Also, physician responses achieved higher Likert scores in relevance, accuracy, and usefulness. The only exception was empathy, in which ChatGPT surpassed physician responses. This suggests LLM chatbots showed the possibility to augment patient or caregiver psychoeducation with a more empathic tone, although they may need further fine-tuning to match enough precision.

Despite the aforementioned optimism surrounding the application of GenAI in ASD care, current research highlights several critical gaps and challenges that warrant careful attention in future work.

While language is a crucial component of ASD evaluation and intervention, an ideal GenAI system should combine cues from tone of voice, facial expressions, eye gaze patterns, motion data, and even biological markers to form a complete picture. Developing such multimodal GenAI systems capable of seamlessly combining linguistic, visual, behavioral, and physiological data is an essential future direction, and recent work has started moving in this direction. For instance, advanced AI models could simultaneously analyze a child’s speech transcripts, social interaction videos, and wearable sensor measurements to generate individualized risk assessments or therapeutic recommendations (17, 50, 66). Achieving this goal necessitates large-scale, diverse datasets and innovative model architectures capable of effectively processing and integrating heterogeneous data streams (16).

In particular, ensuring data diversity is critical; future models must be trained on representative datasets encompassing various ages, cultural contexts, and functioning levels (67). Expanding multimodal databases and establishing shared data repositories accessible to the international research community represent concrete steps forward. Global collaboration and standardization of data collection methods and formats will help overcome current limitations related to small sample sizes and data scarcity (21, 75, 76).

Additionally, integrating AI tools across multiple domains through hybrid or ensemble approaches could significantly enhance diagnostic and therapeutic capabilities. For example, combining an LLM specializing in language processing with computer vision models proficient in emotion recognition can yield more accurate, robust, and generalizable systems (49, 52). Prior work in other neurodevelopmental contexts, such as comprehensive gait analysis in pediatric cerebral palsy, has demonstrated the value of leveraging multimodal data and deep learning to capture complex, clinically relevant patterns (77). Ultimately, multimodal GenAI has the potential to simulate multidisciplinary evaluations, synthesizing diverse information similarly to clinical expert teams, thereby enabling nuanced diagnostics, personalized interventions, and real-time adaptive support for individuals with ASD based on a wide range of real-time inputs.

As previously discussed, the limited interpretability of GenAI models presents a significant barrier to their adoption in clinical ASD care. Therefore, enhancing transparency and explainability of AI systems is a critical future research direction. Specifically, developing Explainable AI (XAI) methods tailored for generative models in ASD diagnosis and intervention could significantly improve clinician and family trust. Ideally, such models would not only provide diagnostic or therapeutic recommendations but also clearly highlight the underlying rationale, such as specific behavioral cues or language features, that informed their decisions (66).

While preliminary efforts in interpretable AI for ASD have begun, applying these techniques to large generative models remains challenging (66). Achieving progress in explainability for ASD-focused GenAI will require adapting broader XAI methodologies to the unique developmental and behavioral complexities characteristic of ASD. Future research should explore advanced explainability methods, including attention visualization which identifies the specific input elements influencing model outputs, and counterfactual explanations, demonstrating how different inputs could alter the model’s predictions (21). Additionally, integrating domain-specific clinical knowledge directly into model architectures could enhance interpretability without compromising performance (78).

GenAI amplifies long-standing ethical challenges in autism research—fairness, privacy and informed consent—while introducing novel threats such as hallucinated content, bias amplification and cloud-based data-security vulnerabilities (76). To manage these risks, we advocate an embedded-ethics interface that couples mental-health practice with computing.

Figure 4 illustrates a quadripartite interface in which (i) clinicians define therapeutic goals and outcome metrics, (ii) engineers translate these requirements into model architectures and validation pipelines, (iii) clinical ethicists perform real-time algorithmic-risk audits and regulatory alignment, and (iv) autistic patients and caregivers contribute lived-experience feedback. This continuous, bidirectional collaboration is intended to accelerate innovation while safeguarding patient safety and social equity.

Complementary scholarship substantiates the value of such an interface. McLennan et al. (2022) propose an embedded-ethics model in which ethicists join AI teams “from the workbench,” participate in sprint meetings and co-author methodological papers, thereby operationalizing real-time ethical scrutiny in medical-AI development (79). Cartolovni et al. (2022) map ethical, legal and social issues across algorithm, physician, patient and organizational layers and recommend an “ethics-by-design” workflow, conceptually aligned with our proposed interface (69). These precedents underscore the feasibility and necessity of embedding ethics expertise directly within ASD-GenAI development teams.

While GenAI technologies have the potential to address shortages in healthcare professionals and reduce labor-related costs, such benefits must not come at the expense of exploiting vulnerable populations or exacerbating inequalities (80). Ensuring that GenAI serves as an equitable tool in ASD care requires verifying that training data accurately reflect the realities of target communities and that models are fine-tuned for specific cultural, linguistic and socioeconomic contexts (81). Ethically robust AI development for ASD must therefore incorporate comprehensive bias assessments and fairness evaluations, proactively include diverse linguistic and cultural representations, and pursue dataset diversification strategies—such as targeted data collection in underserved regions—to ensure equitable performance and avoid amplifying historical disparities (15, 73). Comprehensive international guidelines are therefore needed to address the distinct challenges of deploying these technologies in low- and middle- income countries (LMICs) and other resource-constrained environments (82).

Future work should quantify model performance across gender, ethnicity, culture and socioeconomic status in both diagnostic and interventional tasks. Where disparities emerge, corrective measures—dataset diversification, algorithmic debiasing or fairness-constrained architectures—must be implemented and transparently reported (9, 18, 83). In parallel, participatory co-design sessions with autistic self-advocates from diverse backgrounds can surface implicit assumptions, improve usability and foster trust (78). Ultimately, robust bias-mitigation pipelines and context-specific fine-tuning will be pivotal to realizing the promise of GenAI without accentuating existing inequities in ASD diagnosis and intervention.

Many GenAI-based tools for ASD (e.g., chatbots, robots, virtual reality programs, and mobile apps) have primarily undergone only preliminary pilot studies or evaluations in controlled or simulated environments. Before broader adoption, these interventions must demonstrate reliability, efficacy, and safety through rigorous clinical trials and secure appropriate regulatory approvals. Therefore, future studies should involve larger-scale RCTs evaluating meaningful clinical outcomes, including social functioning, adaptive behavior, and academic achievement, and directly comparing AI-enhanced interventions against standard care models (9, 18). Additionally, research must carefully assess possible adverse effects, such as increased screen dependency or unintended behavioral changes outside therapeutic contexts. Longitudinal studies are particularly crucial in ASD research, given the dynamic and evolving nature of developmental trajectories (9, 11). Long-term investigations can reveal whether improvements from AI interventions are sustained, generalized beyond training contexts, and positively influence life outcomes such as independence or employment in adulthood.

Evaluating user perceptions and engagement is essential for successfully implementing AI interventions in real-world ASD care settings. Existing research indicates that autistic individuals often respond positively to AI-based tools, appreciating their predictability and nonjudgmental interaction style. However, future research should systematically address ongoing user engagement, adherence, and potential dropout rates, identifying factors influencing sustained use (25). Understanding long-term user experiences from the perspectives of patients, families, and clinicians will inform the development of AI tools that are engaging, practical, and effective beyond controlled research settings.

GenAI technologies can potentially support not only direct ASD interventions but also training of caregivers and professionals. Successful adoption of these technologies will require targeted training and capacity-building among therapists, special educators, pediatricians, and caregivers to enhance their confidence and competence in utilizing AI effectively. While some clinicians may initially fear displacement by AI, a more realistic and beneficial scenario positions AI as assistive technology complementing human expertise (9). For example, LLMs could generate preliminary clinical reports from session notes, increasing efficiency (19, 44, 54); robotic assistants could handle repetitive therapeutic tasks, allowing professionals to focus on nuanced clinical decisions (11, 39, 45, 48, 57); or virtual patient simulations powered by GenAI could provide realistic training scenarios for novice practitioners (37). Clearly defining the complementary roles of AI and human professionals and providing adequate training and support will be critical steps toward successful integration into clinical ASD practices

Currently, most AI-based assistive technologies for ASD remain at the research stage, developed primarily within laboratories and not yet widely implemented into clinical practice (50). To facilitate clinical adoption, critical issues such as regulatory approval processes, insurance reimbursement policies, and clear evidence of cost-effectiveness must be addressed (15). Additionally, successful integration of GenAI tools into clinical environments poses pragmatic challenges. Practitioners will need training to use new technologies effectively, adapt existing workflows, and build trust in AI-generated recommendations. Furthermore, maintaining and updating these AI systems, including managing data privacy, applying software updates, and ensuring reliable Internet connectivity, requires resources and technical capabilities that may be limited, particularly in low-resource settings. Addressing these barriers through careful planning and infrastructure investment is essential for achieving meaningful real-world impact from AI tools in ASD care.

Another critical research priority involves optimizing interactions between autistic individuals and AI systems. Future studies should investigate which interaction modalities (e.g., text-based chatbots, voice assistants, robot embodiments) are most effective, comfortable, and accessible for ASD users. Research could explore methods for refining AI systems to better interpret ambiguous or minimal user inputs without requiring autistic individuals to adapt their communication style to technology. Additionally, user-centered prompt engineering specifically tailored to neurodivergent communication styles represents an important area of investigation (9, 50).

Studies such as the Pepper robot (49) suggest that tailoring human–robot interaction paradigms specifically to ASD users (e.g., employing simplified language, visual supports, or predictable robotic behaviors) can enhance usability and effectiveness. Collaborative research involving ASD specialists and user-experience experts can identify design principles most conducive to user engagement and therapeutic effectiveness. Enhancing human–AI interaction in these ways aims to strengthen user engagement, improve therapeutic alliance, and ultimately improve clinical outcomes for ASD individuals.

Due to inherent limitations in ASD-related datasets, generative approaches that create synthetic data or augment existing samples play an increasingly important role. Such approaches include generating synthetic behavioral and neuroimaging data representative of autistic populations, simulating social environments or virtual individuals for AI model training, and applying transfer learning techniques to adapt general-purpose AI models to ASD-specific contexts (41). Evaluating the quality, bias, and practical utility of these generative data augmentation methods is critical, particularly in addressing challenges related to small sample sizes and dataset biases common in ASD research.

Additionally, future research directions include developing specialized AI models optimized for ASD-specific data. This involves designing neural network architectures and explainable AI frameworks specifically tailored to ASD datasets, thereby improving model performance, interpretability, and clinical relevance (10). Investigating the underlying infrastructure of AI, such as effective data augmentation strategies, feature generation techniques, and model refinement methodologies, will further advance ASD-focused AI applications (8). Thus, this research area encompasses not only the applied use of AI tools in practice but also the foundational methods by which AI systems for ASD diagnosis and intervention are developed, trained, and continually improved.