In Depth: AI Foray Is Teaching China’s Edtech Firms a Valuable Lesson

Recent years have seen a surge in the popularity of artificial intelligence (AI)-powered tutoring apps, offering services such as English language training, mathematical problem-solving, and essay-writing assistance. These apps are increasingly being integrated with large language models from both major technology companies like ByteDance and emerging players like DeepSeek, in efforts to enhance the educational experience for users[para. 1][para. 2]. Demand for such AI-edtech solutions is rising swiftly: statistics from ByteDance’s cloud unit Volcano Engine show that the token consumption by clients offering K-12 tutoring grew twelvefold in the six months leading up to May. Despite this growth, the actual efficacy of AI in overcoming the intricate challenges of education depends heavily on the sophistication of foundational models and their adaptability to individual learner needs[para. 3].

China’s private education sector suffered a significant setback in 2021 with the introduction of the government’s “double reduction” policy aimed at lowering students’ homework and after-school tutoring burden. Nonetheless, technology firms have remained committed, pouring investment into digital education platforms to direct more traffic towards their AI-driven tutoring tools. About 20% of queries on AI chatbots like DeepSeek and ByteDance’s Doubao are associated with K-12 curricula, according to company sources[para. 4]. ByteDance, for instance, established an AI-focused division, Flow, in November 2023 to develop a suite of AI educational tools, including Doubao Aixue, which uses advanced reasoning and image recognition to explain problems, generate essay ideas, and grade assignments[para. 5][para. 6]. The company also offers Gauth, a math-solving app initially released in high-pressure East Asian educational markets before being expanded to the U.S. and Europe—markets more open to paying for AI tutoring services. Gauth's early shift to a subscription model enabled financial independence from ByteDance[para. 8].

Other private education firms are venturing into this space as well, often emphasizing AI instructors that simulate human interaction. Doushen Beijing Education & Technology showcased an AI tutor capable of personalized essay instruction, with plans to extend the technology across all K-12 subjects. New Oriental has launched one-on-one classes with Chinese teachers, native English speakers, and AI instructors collaborating for English language teaching[para. 9]. Companies like Gaotu Techedu and Jingzhunxue are unveiling learning machines and classes powered by advanced AI models from cloud giants like Alibaba, promising interactive, personalized learning experiences for children[para. 10].

A critical differentiator in the edtech AI race is the quality of foundational models, which determine how effectively software can respond to individualized questions. While most models handle simple queries well, they still struggle with complex junior-high-level math that demands nuanced, stepwise explanations rather than simple answers—a need amplified by curriculum reforms in major Chinese cities[para. 13][para. 14]. This situation prompts companies like NetEase’s Youdao to develop proprietary models capable of step-by-step reasoning, with Youdao recently open-sourcing its math-focused model, which it claims outperforms competitors in mathematical problem-solving[para. 16]. Firms such as TAL Education leverage their accumulated educational data to tailor models to local exam requirements, while others, like New Oriental, integrate open-source models with their own technology to provide comprehensive solutions without fully developing proprietary foundational models[para. 18][para. 19][para. 20][para. 21]. To safeguard core data, some edtech firms prefer private cloud deployments over public ones[para. 22].

Personalization remains a core challenge and opportunity for AI tutoring. For one-on-one tutoring, AI must not just fill knowledge gaps but also adapt teaching styles and content to individual student needs—diagnosing whether performance issues stem from broad misunderstanding or partial comprehension of subject matter[para. 23][para. 24]. Effective personalization could be enhanced by leveraging data collected from homework and student feedback, possibly in collaboration with supplementary material publishers[para. 25]. Ultimately, building robust, student-tailored AI tutoring will depend on training models with extensive, historic educational data[para. 26].