Can ‘AI Teachers’ Address the Real Pain Points in Education? (AI Translation)

【周刊提前读】“AI老师”试水学科教育 能否解决真痛点？

Subject education has rapidly become one of the fastest and most accessible application scenarios for artificial intelligence (AI) in China. Since 2025, various AI-powered educational features such as essay correction, math problem explanation, English speaking practice, and even AI assistants for university entrance counseling have permeated the entire educational spectrum—from K12 and higher education to adult learning. These capabilities, particularly complex reasoning and task decomposition, have substantially increased the commercial value of “AI teachers”. Notable recent launches include MiniMax and Gaotu’s AI English class using Daniel Wu’s likeness, which sold 20,000 courses at 398 yuan each on its launch night, as well as the VideoTutor platform by Stanford students, which rapidly attracted $1 million in funding for its agent-generated explanatory videos. [para. 1]

AI’s improved reasoning, especially after integrating models like DeepSeek, has enhanced its ability to solve and explain educational problems, echoing the Socratic teaching method long advocated by educators. Edtech firms are leveraging this upgrade for brand renewal—rebranding apps to emphasize AI functions. The 2024–2025 period saw a 12-fold increase in K12 education token usage, reflecting a surge in platform activity, while the monthly active users (MAU) of AI subject education apps rose to over 22 million in May 2025, driven in part by major firms like ByteDance with apps such as “Doubao AI Learning” and “Gauth”. [para. 2][para. 3]

Despite the regulatory “Double Reduction” policy limiting educational burdens and off-campus training, internet giants are investing heavily, recognizing both the strategic importance of AI entry points and strong underlying demand. For example, 20% of chatbot query volume in ByteDance apps relates to K12 academic questions. Meanwhile, AI’s automation of repetitive educational tasks (e.g., correcting subjective exam answers or providing interactive feedback) allows for significant scaling and cost reductions, exemplified by one teacher’s user capacity increasing from 150 to 800 students. [para. 4][para. 5]

User understanding of generative AI is evolving, and the main logic behind most AI- or “adaptive”-branded courses remains iterative question and content recommendations, not truly individualized learning. However, the popularity of DeepSeek and similar tools has increased AI search frequency, prompting hardware and course providers to highlight interactive “AI 1-on-1” capabilities. Still, replacing real teachers entirely remains contentious for parents, who value human expertise and interaction. [para. 6][para. 7]

Expansion of AI teaching tools also raises risks around over-reliance and potential erosion of independent learning and creativity among students. Official guidelines now urge students to use AI critically and prohibit its uncritical application in homework and exams. During the 2025 national college entrance exam period, leading models like DeepSeek and Doubao suspended their exam-related query functions. [para. 8][para. 9]

From 2024, ByteDance has dominated the ecosystem, updating its Doubao model to version 1.6 and splitting its AI initiatives across multiple user-focused apps, fueling rapid acquisition. Doubao’s monthly active users exceeded 130 million in May, although its AI learning app “Doubao AI Learning” trails content-heavy incumbents like Zuoyebang and Yuanfudao. “Photo search” for answers is a worldwide need, with Chinese applications such as Gauth and Question AI topping US app charts, though regulatory constraints in China require tiered parental controls. [para. 10][para. 11][para. 12]

Fully interactive AI teachers—capable of sophisticated, engaging dialogue and subject mastery—have yet to be realized at scale, but prototypes abound (e.g., AI English, math, writing tutors). Such products emphasize scalability, reproducibility and cost-effectiveness. Yet, personalization, content diversity and accurate simulation of master teachers’ classroom charisma still require substantial research and development. [para. 13][para. 14][para. 15]

Model inference ability has sharply improved, with top AI models now achieving over 80–90% accuracy on most math questions, and in some high school exam scenarios surpassing 130 points out of 150. However, real-world application remains uneven—models struggle with novel or flexible exam questions and detailed step-by-step solutions. As a result, many firms prefer integrating open-source models like DeepSeek and customizing them locally for proprietary educational uses, though the pace and expense of such efforts vary widely. [para. 16][para. 17][para. 18]

AI’s transformative potential in education is evident, especially in automating grading, supporting scalable personalized learning, and democratizing access to expert-level assistance. However, true educational value depends on integrating AI tools with pedagogy that preserves teacher expertise, fosters creativity, and ensures fairness, as well as addressing the longstanding challenge of providing nuanced, individualized feedback within a standardized exam-driven system. [para. 19][para. 20]