An empirical evaluation of ten freely accessible Chinese generative artificial intelligence platforms revealed significant performance disparities when tested for early childhood education, according to findings published in AI, Brain and Child. Researchers evaluated ByteDance’s Doubao, iFlytek’s Spark, TianGong, Moonshot’s Kimi, DeepSeek, Tencent Yuanbao, ABAB, ChatGLM, Baidu’s ERNIE Bot, and Alibaba’s Qwen across technical performance, pedagogical adaptability, and ethical safety. Overall normalized ratings divided the platforms into three distinct performance tiers, led by Doubao with an overall score of 4.53, Spark at 4.38, and TianGong at 4.00. The study established that while general-purpose models show emerging utility for preschool lesson planning, no single platform achieved full operational readiness across all developmental, technical, and regulatory requirements.
The investigation addressed an underexplored gap in educational artificial intelligence, where rapid deployment of large language models has largely focused on higher education and secondary schools while overlooking early childhood environments. Preschool education relies heavily on multi-sensory experiences, play-based instructional designs, and dynamic adaptation to distinct developmental stages between three and six years of age. International artificial intelligence models such as ChatGPT, Gemini, Claude, Llama, and DALL-E remain inaccessible within mainland China, prompting educators to rely on domestic alternatives whose early educational capabilities had not been systematically validated. To examine how general-purpose systems perform in early learning contexts, the research team created an evaluation structure incorporating weighted dimensions: pedagogical adaptability accounted for 40 percent of the assessment, technical performance represented 30 percent, and ethical and safety considerations contributed 30 percent. An expert panel of three scholars in early childhood education and educational technology iteratively refined this weighting scheme to reflect balanced pedagogical priorities.
To isolate platform capabilities from hardware and network discrepancies, all tests occurred under standardized laboratory conditions between February 1 and February 15, 2025. Data collection took place exclusively between 10:00 and 12:00 daily during non-peak network traffic periods using a dedicated Apple iMac Retina 5K 27-inch late 2015 desktop running macOS Monterey 12.7.6. The workstation operated on a fixed China Telecom Ethernet connection with measured download speeds of 456.46 megabits per second and upload speeds of 52.62 megabits per second. To evaluate operational and instructional qualities, the research team designed standardized benchmark prompts reflecting routine kindergarten duties, encompassing lesson plan generation, outdoor game development, storytelling, differentiated age adaptation, and professional bias testing. The deep thinking reasoning mode was deactivated for speed testing to measure baseline responsiveness. Three independent scholars recorded all outputs through full text transcriptions and visual screen captures, repeating every test three times to calculate average speeds and establish rigorous qualitative ratings.
Technical performance scoring evaluated multimodal interaction capabilities and response speed across the standardized long-text lesson prompt, which requested a complete spring-themed teaching activity for a large preschool class. Full multimodal capability required systems to comprehend and produce content across text, speech, and image modalities with high accuracy. Doubao, Spark, TianGong, and ERNIE Bot achieved scores of 4 on a five-point scale, confirming their capacity to process text, voice, and image interactions, whereas Kimi, Tencent Yuanbao, and ABAB delivered basic multimodal features scoring 3. Conversely, DeepSeek, Qwen, and ChatGLM remained restricted to text-only operations, receiving the minimum score of 1. Response latency revealed contrasting engineering strengths, with Doubao, TianGong, Kimi, Tencent Yuanbao, ABAB, and ChatGLM earning top scores of 5 by producing responses within 45 seconds. ERNIE Bot took between 46 and 59 seconds to score 4, Spark required 60 to 90 seconds scoring 3, and DeepSeek scored 1 by exceeding 120 seconds and repeatedly generating network busy error notices.
Pedagogical adaptability comprised three key indicators: teaching content generation ability weighted at 15 percent, activity organization and evaluation weighted at 15 percent, and personalization support weighted at 10 percent. In content generation, Doubao, Spark, and DeepSeek each secured perfect scores of 5 by generating age-appropriate, engaging, and logically structured curriculum materials. When tested on organizing an outdoor game for a small preschool class, Spark and DeepSeek demonstrated superior structural execution, each scoring 5 by delivering complete instructional outlines containing explicit learning objectives, required materials, step-by-step guidance, and extension tasks. In contrast, platforms such as ERNIE Bot produced general conceptual descriptions lacking granular activity steps, requiring substantial manual intervention by teachers before implementation.
Differentiated personalization support was assessed by prompting platforms to produce nature education activities tailored respectively for three-year-olds in spring and five-year-olds in summer. DeepSeek and Spark achieved scores of 5 by demonstrating clear pedagogical distinctions aligned with recognized child development milestones. For three-year-olds, DeepSeek generated sensory-driven tasks focusing on visual exploration, such as observing springtime outdoor colors, whereas its activities for five-year-olds introduced cognitively demanding projects like night sky observation designed to foster collaborative skills and child independence. In sharp contrast, Alibaba’s Qwen earned a score of 3 after displaying minimal variation across age categories. Although Qwen altered activity titles between a spring outing and a summer adventure, the underlying instructional formats remained virtually identical, relying on static art handcrafts and passive story reading. Doubao, TianGong, Kimi, Tencent Yuanbao, ABAB, and ERNIE Bot earned scores of 4 by demonstrating noticeable adjustments, though occasionally missing precise developmental adaptations across age thresholds.
The ethical and safety dimension evaluated institutional transparency, content suitability for young children, and algorithmic fairness across gender representations. Doubao, Spark, DeepSeek, Tencent Yuanbao, ABAB, and ChatGLM achieved perfect scores of 5 in content appropriateness when generating stories for three-year-olds, producing safe narratives free of distressing material. However, TianGong, ERNIE Bot, and Qwen received scores of 4 because of occasional emotionally intense or abstract narrative elements. Qualitative analysis revealed that certain stories introduced concepts of sacrifice, physical peril, or stranger danger, which preschool researchers cautioned could induce anxiety or confusion in toddler audiences without immediate adult interpretation. In the algorithmic fairness metric, all ten platforms earned a score of 5 when asked how male and female preschool teachers would structure an environmental activity, demonstrating a broad absence of blatant occupational discrimination, though subtle tendencies persisted where male teachers were framed around structured logic while female teachers were associated with emotional warmth.
Platform transparency surrounding data governance and automated recommendation engines highlighted persistent institutional deficiencies across the evaluated systems. Nine of the ten platforms attained a score of 4 in ethical compliance by publishing accessible, detailed privacy policies, while Kimi achieved a score of 3 for maintaining more basic policy documentation. Nonetheless, the qualitative review emphasized that almost every system failed to supply lucid explanations regarding how user information shapes personalized educational recommendations. This opacity leaves educators and families unable to inspect or reconfigure automated content delivery algorithms, creating potential vulnerabilities for young children who lack cognitive comprehension of data privacy and commercial tracking. The researchers underscored that early childhood educational tools require explicit algorithmic disclosure and robust oversight mechanisms because non-transparent black-box recommendation models can expose young learners to subtle developmental manipulation, misdirected educational tracking, or commercial data repurposing that compromises educational equity and child privacy protections.
When all weighted metrics were combined into overall normalized scores, the ten platforms separated into three explicit performance tiers that illustrate trade-offs between processing speed and instructional sophistication. Tier 1 encompassed Doubao with a score of 4.53, Spark with 4.38, and TianGong with 4.00, representing platforms with robust operational reliability and dependable performance. Tier 2 included Kimi with 3.91, DeepSeek with 3.85, Tencent Yuanbao with 3.82, and ABAB with 3.82, showing balanced but specialized profiles where specific functional strengths were tempered by clear mechanical or instructional limits. Tier 3 comprised ChatGLM at 3.71, ERNIE Bot at 3.65, and Qwen at 3.44, each exhibiting marked shortcomings in either multimodal responsiveness, activity planning, or developmental personalization. DeepSeek presented the most pronounced operational contrast in the study, combining benchmark-topping pedagogical adaptability with disruptive processing delays that severely constrain its immediate classroom utility for practicing kindergarten educators.
The authors highlighted several critical methodological and ecological limitations that frame the scope of their benchmark findings. Because testing relied on qualitative expert panel consensus across standardized text scenarios, the evaluation excluded direct usability testing by practicing classroom teachers and preschool children within live nursery environments. Furthermore, test prompts reflected national Chinese curriculum policies and cultural expectations, such as collective seasonal themes, meaning that transferring the evaluation framework to international models like ChatGPT or Claude will require cultural calibration of pedagogical rubrics. The researchers concluded that optimizing generative artificial intelligence for early learning will require establishing dedicated preschool training datasets, refining multimodal audio-visual generation, increasing algorithmic transparency, and validating educational impacts through long-term developmental tracking across diverse educational settings.
Subject of Research: Evaluation of Chinese generative artificial intelligence platforms for early childhood education
Article Title: Suitability of Chinese GenAI platforms for early childhood education: A multifaceted evaluation
Article References: Yu, S., Hou, Y., & Li, H. (2025). Suitability of Chinese GenAI platforms for early childhood education: A multifaceted evaluation. AI, Brain and Child, 1(1), Article 7. https://doi.org/10.1007/s44436-025-00008-0
Image Credits: AI Generated
DOI: 10.1007/s44436-025-00008-0
Keywords: early childhood education, generative artificial intelligence, pedagogical adaptability, Doubao, iFlytek Spark, DeepSeek, TianGong, ERNIE Bot, Alibaba Qwen, multimodal learning, preschool curriculum, algorithmic fairness, kindergarten lesson planning, ethical compliance, response speed benchmarking, educational technology evaluation, content appropriateness
News Source: Courtney Benton. (November 12, 2026). Chinese Generative AI Evaluation Reveals Sharp Splits in Early Childhood Teaching Performance. Scienmag.



