Artificial Intelligence and Educational Transformation

WANG Xuenan¹, LI Yongzhi²

（1.Research Institute of Digital Education，China National Academy of Educational Sciences, Beijing 100088; 2.China National Academy of Educational Sciences, Beijing 100088）

Since the official release of ChatGPT on November 30, 2022, generative artificial intelligence has achieved rapid breakthroughs in less than two years, transitioning from open-ended text dialogue interaction to text-to-image, text-to-video, and then to multi-modal interaction technologies that facilitate human-machine interaction. Its development speed far exceeds human reflection and response speed. This intense uncertainty and unknowns have prompted humanity to pay more attention to the iterations of artificial intelligence and the social transformations it triggers. Education, traditionally viewed as a slow variable and the most stable field, is now widely recognized as one of the areas most directly and rapidly impacted by artificial intelligence. Therefore, maintaining a rational, open, objective, and rigorous attitude, combined with a historical materialism perspective, to examine and analyze the real development level of artificial intelligence and educational transformation, as well as the intrinsic correlation and mechanism between the two, will provide foundational viewpoints and perspectives for enabling high-quality development in education through artificial intelligence.

Represented by ChatGPT, generative artificial intelligence relies on its powerful natural language processing capabilities to complete complex tasks such as information retrieval, question answering, content creation, and code generation through open-ended dialogue. Its capabilities are increasingly approaching human intelligence, even extending and partially replacing human intelligence. The current emergence of “intelligent” results from the rich data, enhanced computational power, active open-source environment, and optimization of multi-modal large models. However, essentially, it is not a new technology but a phased product in the development process of artificial intelligence, which has not yet undergone a breakthrough qualitative change at the technical level. Although the history of artificial intelligence development is not long, to objectively evaluate the true level of current artificial intelligence development and future trends, one must place the current technological explosion’s singularity within the historical context of artificial intelligence development and technological revolutions to understand the impact and challenges of artificial intelligence on education.

(1) Three Stages of Artificial Intelligence Development

In 1950, the famous Turing Test marked the beginning of artificial intelligence. In 1956, the Dartmouth Conference formally proposed the concept of “artificial intelligence,” marking the birth of the discipline. After nearly 70 years of development, the research fields within artificial intelligence have undergone multiple differentiations and integrations. After experiencing two notable “AI winters” due to insufficient applications, limited computational power, and lack of funding, this discipline has once again welcomed a rapid growth phase.

According to the degree of intelligence, artificial intelligence can be divided into three major stages: computational intelligence, perceptual intelligence, and cognitive intelligence^[1-2]. The first stage is the computational intelligence stage (1950-2000), where machines store and compute information. The second stage is the perceptual intelligence stage (2000-2021), where machines capture signals from the physical world through sensors, understand some intuitive aspects of the physical world, and efficiently complete tasks related to “seeing” and “hearing.” The third stage is the cognitive intelligence stage (2022-present), where machines possess thinking and learning abilities similar to humans and can autonomously make decisions and take actions. This stage is primarily marked by the release of ChatGPT. However, the scientific community generally believes that artificial intelligence has not yet reached this stage and is still in the exploratory early phase.

(2) Three Trends in Artificial Intelligence Development

Throughout the development of artificial intelligence, two main paths exist: one is through symbolic reasoning, driven by model learning-based data intelligence, known as “symbolicism,” advocating that artificial intelligence should mimic human logical processes to acquire knowledge. The other is through neural networks, driven by cognitive bionics, known as “connectionism,” which promotes learning knowledge based on big data and training, advocating the imitation of human neurons to achieve artificial intelligence through neural network connection mechanisms. Throughout the development of artificial intelligence, the two schools of symbolism and connectionism have experienced cycles of rise and fall, each flourishing due to different technological routes and development models that played important roles in shaping the theoretical foundation and technical implementation of artificial intelligence, reflecting scientists’ relentless efforts to understand and simulate human intelligence. As people’s understanding of artificial intelligence matures, the progress of the “connectionism” development path will gradually slow, while “symbolism” will flourish again. Even the representatives of connectionism, such as Yang Likun, Li Feifei, Geoffrey Hinton, have expressed that the current technological route cannot produce AI with perceptual capabilities^[3]. Based on this, this article preliminarily judges that the future development of artificial intelligence will have the following three major trends:

1. Transitioning from cognitive large models to multi-modal large models. Traditional AI models focus on processing information from a single modality, mainly emphasizing understanding and generating natural language. Multi-modal large models can process various data types such as text, images, audio, video, and code to facilitate content synthesis tasks and integrate multiple information sources. Human intelligence and learning evolution are inherently multi-modal; humans possess eyes, ears, mouths, noses, tongues, and limbs, and artificial intelligence’s learning can also better replicate the real situations triggered by human multi-sensory learning^[4].

2. Deepening applications from general large models to “big-small linkage.” The growth of AI model computational power and optimization of algorithm efficiency presents a new “Moore’s Law,” where model performance improves with the increase in model scale, data scale, and computational scale, showing a power-law distribution characteristic, which has become a practical obstacle for large models to achieve deep application in industries and create value. Small models can learn from large models through knowledge distillation. At the same time, small models can also feed back to large models, improving their training accuracy. Therefore, the synergy between large and small models is an effective way to reduce training and application costs while enhancing flexibility, applicability, and efficiency.

3. Transitioning from language intelligence to embodied AI. In existing large model applications, AI tools are often embedded in existing processes to improve efficiency, without generating innovative value at the underlying logic and native level. The shift from virtual to real provides opportunities for developing and applying autonomous and adaptive artificial intelligence agents. To create an artificial intelligence agent capable of functioning in the real world, training solely in a textual environment is insufficient; it must possess the ability to perceive the physical properties of the real world^[4]. Generative artificial intelligence technologies represented by GPT-4o can not only achieve human-machine interaction in digital and physical spaces but also provide emotional value, indicating that emotional computing is one of the key research directions for the future of artificial intelligence.

(3) The Reality of Artificial Intelligence Development: Qualitative Differences Between General Artificial Intelligence and Human Intelligence

Currently, people’s astonishment at the interactive level of artificial intelligence to provide emotional value, cognitive mechanisms, and collaborative value mainly stems from their initially low expectations, remaining at the stage of fixed, mechanical robot dialogue or the human-machine Go battle of Alpha Go. In fact, the current real development level of artificial intelligence is far from general artificial intelligence, still significantly different from humans, especially in higher-order cognition and social emotions.

Director Zhu Songchun pointed out in an interview during the 2024 National Two Sessions that the “general” in general artificial intelligence has a specific academic meaning. Generally, in daily physical and social contexts, artificial intelligence must meet three basic conditions: first, it must be able to complete infinite tasks rather than just a few limited tasks defined by humans; second, it must proactively and autonomously discover tasks in the scene, achieving “awareness of work”; third, it must have autonomous values to drive it rather than being passively driven by data^[5]. Currently, although generative artificial intelligence applications such as ChatGPT, Claude-3, Wenxin Yiyan, and iFlytek Spark are recognized as relatively successful both domestically and internationally, they still do not fully meet the standards of general artificial intelligence, nor do they possess capabilities equivalent to humans. They outperform humans in basic cognitive areas such as data processing, memory, combinatorial creativity, speed, and accuracy but lack human emotional rationality, value systems, cognitive and reasoning abilities, and the creativity to innovate from 0 to 1. Large models exhibit significant shortcomings in simulating the real world, whether through external information coding or relying on internal first principles (i.e., scaling laws), showing strong dependence on data, lack of model interpretability, and insufficient common-sense understanding^[6]. If these issues can be resolved in the coming years, the intelligence level of large models is expected to improve further, better integrating into social applications.

Currently, although generative artificial intelligence still falls under the category of weak artificial intelligence, its iteration speed and performance level have far exceeded our original expectations. Analyzing the technical limitations of generative artificial intelligence from an educational perspective will break the limitations of previous grand narratives or micro-arguments regarding technological development and educational transformation, scientifically analyzing and rationally questioning the current and future impact of artificial intelligence on education through complexity thinking. The development of artificial intelligence, training of large models, and educating children exhibit isomorphic characteristics. This article will focus on the elements and links of education as the logical thread, discussing six aspects: structural comparison, logical comparison, symbolic coding, content analysis, interaction mechanisms, and training modes.

(1) Structural Comparison: Large Models and the Brain

Artificial intelligence is essentially a mimicry of the organizational structure and operational mechanisms of the human brain, materializing human intelligence. Reproducing human cognition within computational systems is key^[7]. The GPT-3 large language model has 175 billion parameters, while GPT-4 reaches 1.8 trillion parameters, with a training cost of $63 million^[8]. In the development of language intelligence, the model’s functions have become increasingly powerful, generalization capabilities have improved, and problem-solving abilities have strengthened. Large models attempt to simulate human brain neurons as closely as possible by continuously increasing the number of parameters, aiming to replicate human intelligence. However, the human brain contains billions of neurons, interconnected by synapses, with approximately 80 to 100 billion neurons and over 100 trillion synaptic connections^[9]. Neurons communicate through electrical signals, forming complex networks, and even today, humans do not fully understand their operational principles. Von Neumann proposed in “Computer and the Brain” that “neurons of the same volume can perform more computations than artificial components, process more information simultaneously, and have much greater memory capacity; each neuron’s accuracy may be lower, but its overall reliability is higher”^[10]. This means that if the human brain is organically connected, then artificial intelligence is mechanically connected, with its inherent richness and complexity incomparable. Following the trend of computer science development, in a few years, the parameters of large models may reach the scale of hundreds of trillions, akin to the human brain. According to the power-law principle, appropriately distributing model parameters and training data sizes can achieve effective models within limited budgets or expected computational speeds. However, the relationship between model parameters and model intelligence is not a simple linear relationship; the mechanisms of perception, cognition, reasoning, and innovation between large models and the human brain are different. Therefore, blindly pursuing model parameters cannot achieve a performance that fully simulates human intelligence and may not be the future development trend of large models.

(2) Logical Comparison: Probabilistic Reasoning and Conceptual Reasoning

Probabilistic reasoning and decision theory provide important thinking methods and decision-making bases for artificial intelligence systems. By establishing Bayesian networks and using reinforcement learning techniques, artificial intelligence systems can make decisions based on past experiences and observational results, improving decision accuracy and efficiency. Therefore, current artificial intelligence based on probabilistic reasoning has inherent technical limitations.

On one hand, artificial intelligence is based on probabilistic reasoning, while human intelligence is based on conceptual reasoning; there is a qualitative difference between the two. Probabilistic reasoning calculates based on existing information and data to obtain the maximum likelihood. Conceptual reasoning, belonging to formal logic, is based on concepts—abstract symbolic products of human thinking activities—expressing understanding, induction, or classification of certain entities or phenomena through language, reflecting people’s higher-order thinking forms regarding cognition and understanding. When computer language has not broken through von Neumann structure and binary logic, all calculations stored ultimately become a relationship of addition and subtraction, still infinitely expanding in low-dimensional space. Generative artificial intelligence has not yet broken through the computational model of probabilistic reasoning; it merely operates under the support of big data, high computational power, and large models, along with human feedback reinforcement learning (RLHF), enabling machines to make decisions based on uncertain information, achieving maximum probability and results closest to human thinking. Large models cannot apply a single algorithm to solve various problems; artificial intelligence can only respond to deterministic instructions. However, the human brain can face different problem scenarios, execute different tasks simultaneously, and switch freely to cope with uncertainties. Thus, it is evident that artificial intelligence currently remains at the low-level thinking stage of logical reasoning, probabilistic reasoning, and causal reasoning and cannot exhibit the high-dimensional human intelligence.

On the other hand, generative artificial intelligence finds it challenging to break through linear, fragmented causal logic chains and cannot generate real, specific practical content in real-time based on diverse social cultures and ethics. However, this does not mean that the content it generates lacks creativity; rather, due to the lack of constraints from logical systems and ethical norms, it often leads to situations of “knowledge fantasy.” From the perspective of engineering practice, generative artificial intelligence can indeed produce unexpected wisdom. However, from the effectiveness of knowledge generation, the knowledge created by generative artificial intelligence is achieved through training on past big data using probabilistic reasoning, akin to “driving while looking in the rearview mirror.” McLuhan vividly explained the “rearview mirror effect” as “using inherent experiences to solve problems, looking at the present through the rearview mirror, we walk backward into the future”^[11]. This presents an essential difference from real educational scenarios. The education of children and human learning occur in real teacher-student interactions or practical labor contexts, cultivating qualities through action and construction, integrating the scientific knowledge system of human wisdom with new experiences generated in current real life.

(3) Symbolic Coding: Language Coding and Implicit Knowledge

Language is a unique symbolic system of humanity, consisting of phonetics as its material shell and semantics as its meaningful content, along with lexical materials and grammatical organization rules^[12]. Language itself is a form of coding. Therefore, whether the content of education can be coded and decoded becomes a key distinction between what can be “said” and what cannot be “said.” In 1958, Michael Polanyi first proposed in “The Study of Man” that human knowledge is divided into explicit and implicit knowledge (also known as tacit knowledge), typically described as knowledge represented in written text, charts, or numerical formulas, which is only one type of knowledge; the other type is knowledge that cannot be systematically expressed, like the knowledge we possess while performing an action^[13]. He pointed out that compared to explicit knowledge, implicit knowledge’s important characteristics include: first, it can be logically explained through language, text, or symbols; second, it cannot be transmitted through school education, mass media, etc.; third, it cannot be subjected to “critical reflection”^[13].

Thus, it is evident that the development of artificial intelligence, centered on natural language understanding and processing and machine learning, relies on the corpus and data information that can be recorded, coded, and logically constructed. The intelligence of large language models is based on explicit knowledge that can be recorded, coded, and disseminated through language, but the existence of implicit knowledge, as another type, is often overlooked. Various types of coding have certain limitations in connotation expression and meaning construction; the limitations of textual expression restrict the development of multi-modal large models’ intelligence. Multiple codings and transformations can lead to significant filtering and attenuation of information. As Wittgenstein stated, “Language dresses thoughts; from the outer form of this dress, one cannot infer the form of the thought it covers”^[14]. Language is a tool for human thinking and communication, but its expressive capacity is limited, unable to fully capture and describe the complexity of the real world. Language serves as both a scaffold of thought and a shackle of thought. In human learning and evolutionary development, implicit knowledge often occupies a larger proportion, holds greater significance, and presents more challenges, such as discerning colors in the spectrum or feeling the texture of materials through touch.

Artificial intelligence struggles with principle-based knowledge, procedural methods, and value-based knowledge; it is powerless in generative teaching, emotional teaching, and practical teaching^[15]. Such knowledge and teaching are not easily “transmitted”; they are more suited for “demonstration.” Learning occurs through doing, in rich, complex, and precise multi-sensory interactions, establishing true connections between body, mind, brain, and physicality. Furthermore, even if experiences are expressed in language, they lose much contextual and background information for the recipient. When the recipient interprets from their perspective, it loses all the nuances (relative to the expresser). Therefore, artificial intelligence, based on large language models as its core technology, merely injects explicit knowledge that humans can encode, calculate, or express through language, and its self-supervised language models cannot obtain knowledge about the real world; its essence is “compression.”

(4) Content Analysis: Massive Data vs. High-Quality Data

Although there is no clear consensus among scientific researchers and industry personnel regarding many issues in the field of artificial intelligence, there seems to be a consensus that data quality is key to the emergence of large model capabilities in the next phase. In the production relationships of large models, data is the production material, computational power is the productive force, and algorithms are the production tools.

Represented by ChatGPT, generative artificial intelligence combines labor-intensive, technology-intensive, and capital-intensive technologies and industries. This is because the vast majority of computational power is used in pre-training, primarily for data collection and cleaning; in addition, fine-grained, high-quality data labeling is also a significant human-intensive task, and much foundational work is aimed at obtaining high-quality data.

Regarding the impact of data volume (Training Tokens) and model parameter volume (Parameters) on the model, OpenAI in 2020 enhanced the intelligence level of large models by expanding model parameters^[16]. However, the conclusion drawn by DeepMind later changed this, indicating that under limited computational resources, more and better training data is more important than merely increasing model parameter scale^[17].

In our traditional cognition, it is generally believed that China has a comparative advantage in massive data during the new wave of artificial intelligence development. However, the reality is otherwise, especially in the education sector, where the issue of high-quality usable data is more prominent. Although we have the largest number of teachers and students globally, continuously generating new data in daily education and teaching management, the currently available high-quality data mainly comes from static, sedimented professional texts such as books, news, and scientific papers. This data is far from sufficient for optimizing and deepening applications of large models, such as transitioning from large models to industry-specific models. The free public data available on the internet lacks depth and precision, failing to meet the strong professionalism and high accuracy required for education-specific models. Although China already possesses massive educational big data, including multi-modal teaching data, the amount of high-quality, structured, and computable effective data is limited. The main issues include incomplete and non-unified data standards, narrow data collection coverage, insufficient professionalism in model construction, single and mechanical application services (mainly focusing on adaptive teaching and question bank types), lack of open sharing, and the need for improved privacy protection. Particularly, the lack of standards and data in teaching environments and processes greatly restricts the development and accumulation of educational big data. Therefore, excavating the value behind existing data, strengthening future data management, clarifying industry standards, establishing data usage rules, and ensuring that large models are trained with sufficient and accurate professional data are the basic prerequisites for empowering education with generative artificial intelligence.

(5) Interaction Mechanism: Reinforced Feedback and Teaching Interaction

In information processing, human feedback is key to enhancing the “intelligence” of large models. Human feedback reinforcement learning is a new training paradigm in the field of generative artificial intelligence, guiding intelligent systems’ behavior through human feedback. Over the past few years, various large language models (LLM) have generated diverse texts based on human input prompts, primarily relying on contextual logic and probabilistic reasoning, which can lead to certain biases. However, through RLHF, language models trained on general text data can align with complex human values, making generative artificial intelligence more “humanized.” It is precisely the feedback and tuning of human wisdom that brings artificial intelligence closer to human intelligence.

Classroom teaching is also a complex information transmission system that is purposeful, directional, and orderly. Teaching feedback, as a necessary part of the teaching process, allows teachers to adjust and optimize teaching strategies through real-time feedback to adapt to students’ learning behaviors. Regarding teaching feedback, accuracy, relevance, guidance, motivational quality, timeliness, diversity, and interactivity are its core characteristics^[18]. It can be seen that teaching feedback and RLHF share the same execution mechanism.

(6) Cultivation Mode: Multi-modal Input and Holistic Development

In information input, multi-modal information types are prerequisites for input effectiveness and richness. By combining different types of data, large models can better understand and predict complex real-world problems. Currently, most models are trained as separate modules, converting different modalities into language text and then stitching them together to achieve approximate multi-modality; the shortcoming lies in the inability to conduct deep complex reasoning in multi-modal space. Native multi-modality technically advances further, having the ability to process different forms of data (language + audio + visual) and initially pre-training on different modalities, fine-tuning with additional multi-modal data to enhance effectiveness.

Just as the embodied learning theory in the field of education emphasizes activating multiple brain regions through input from visual, auditory, sensory, and tactile multi-sensory information and the multi-modal interactions among learners, technology, and environment, achieving optimal learning outcomes through deep learning^[19]. The training of large models also reflects the value orientation of cultivation modes, which raises the question of whether to choose quality education or exam-oriented education. If a single-dimensional, single-modal “drill question” reinforcement training is chosen, the intelligence of large models may rapidly improve in some areas in the short term, but it will quickly reach a bottleneck. If a holistic development and multi-modal quality education is chosen, the iteration speed of large models may be slower than the former, but the upper limit of intelligence will be higher. Because general knowledge is the foundation of specialized knowledge, developing general cognitive abilities first is essential to developing specialized cognitive abilities, and the same applies to large models. The education sector must also guard against the entry of “book-smart” large models with high scores but low capabilities into the application market.

Artificial intelligence is not only a scientific issue but also an educational and social issue. If human civilization aims to inherit and develop, proactively facing artificial intelligence is a necessary step. However, overall, people tend to overestimate the short-term effects of artificial intelligence while underestimating its long-term effects. Therefore, measures must be taken from the current situation to objectively and rationally view the development of artificial intelligence and make judgments. Currently, the rise of the third wave of artificial intelligence does not stem from academia but is driven by the urgency and market demands from the business sector. Essentially, this is not a new technological breakthrough in the field of artificial intelligence but a result of the inevitable trend and strong demand driven by the popularization and transformation of educational digitization.

(1) The Impact of Artificial Intelligence on Education

In the long run, the impact of artificial intelligence on educational development should prioritize the consideration of the following three aspects:

First is value rationality. Today’s educators may not be able to accurately predict the complex intertwining factors of future impacts, especially the rapidly changing factor of artificial intelligence, which has deeply integrated the influences of collective wisdom, artificial intelligence, and social networks into our lives. The capabilities of artificial intelligence primarily stem from the vast data learned from humans, which contains key clues and facts that can help us solve problems, as well as biases, discrimination, hostility, and hatred present in human society. When artificial intelligence learns human data without ethical safeguards or moral frameworks, it also learns human weaknesses. When artificial intelligence serves humanity, it subtly contains biases and other issues^[20]. Therefore, consciously cultivating the ability of learners to form values and moral judgment capabilities that adapt to future society, enabling them to make independent judgments with firm value rationality in the face of complex and unpredictable circumstances, is essential.

Second is ethics and morality. It is crucial to construct a social ethical and moral system after the advanced development of machine intelligence. Currently, the roles that large models of artificial intelligence will play in the future can be categorized into three types: tools, partners, or adversaries, with different social cultures having varying positions on them. Japan’s “AI Principles” emphasizes that future artificial intelligence may play the role of a quasi-member of society or even a human partner; it stipulates that if AI develops to the stage of quasi-member or human partner, it must adhere to the ethical and moral norms of human society as well as the ethical and moral norms established for artificial intelligence^[22]. In contrast, Western science fiction films and novels often portray artificial intelligence as antagonists, enemies of humanity. What role large models of artificial intelligence will ultimately play, how they will coexist harmoniously with humans and nature, and better assist humanity, should be prioritized for consideration. The decoding of the mysteries of carbon-based life and the establishment of artificial intelligence agents (silicon-based life) based on these principles should also become a focus for the future of artificial intelligence. We should maintain an open attitude, uphold the original intention of artificial intelligence serving human societal development, and establish it under the constraints of human ethical and moral guidelines. Simultaneously, the human ethical and moral system must progress in tandem with changes in the forms of civilization. The primary responsibility of education is to cultivate qualified citizens for future society and play an important role in building an ethical and moral system for an intelligent society.

Third is talent cultivation. In the future intelligent society, whether artificial intelligence can coexist harmoniously with humans, nature, and society depends not on artificial intelligence itself but on whether humanity’s cognition and attitude towards artificial intelligence can evolve rapidly. Therefore, education should shift towards cultivating higher-order abilities such as innovative thinking in learners. The future society will require a large number of high-level talents with human-machine collaborative capabilities, with innovative thinking, computational thinking, and emotional abilities becoming key competitive strengths for humanity. To address the new challenges of the artificial intelligence era, countries should re-examine the value of the school education system and reflect on the issues of “what kind of people to cultivate” and “how to cultivate them.” People realize that, more than in any historical period, there is a greater need to highlight human value and consolidate human power to resist anxiety and fear, distinguishing between humans and machines, and between humans and artificial intelligence^[23]. In the face of an uncertain post-truth world, education should not only focus on teaching students what to learn but also help them escape from being “tools” to shape “complete individuals,” stimulate their subjectivity and intrinsic motivation, and cultivate their independent thinking and sustainable learning abilities. The simultaneous promotion of various aspects of education and comprehensive development is closely related to human emotions^[24]. Therefore, nurturing social emotional skills that machines cannot possess is a key goal of future education.

In the short to medium term, artificial intelligence brings six aspects of impact on education. First is the impact on cultivation goals. To address the long-term challenges posed by artificial intelligence, education must adjust talent cultivation goals based on future societal needs, focusing on developing students’ core competencies and nurturing the correct values, essential qualities, and key abilities required for lifelong development and adaptation to social progress. Second is the impact on learning methods. Artificial intelligence can facilitate personalized learning paths, provide intelligent learning assistance, and create more realistic learning scenarios for learners through virtual reality and augmented reality technologies, simulating scientific experiments that cannot be presented in the real world, among other things. Third is the impact on teaching methods. Through artificial intelligence, humanity can resolve the dilemma of large-scale teaching and individualized instruction in practice, promoting educational equity and enhancing educational quality, achieving better teaching and learning outcomes. Fourth is the impact on teacher-student relationships. Previously, teachers were the academic authorities in the classroom, but now students can access knowledge instantly using tools like ChatGPT and Sora, which may exceed what teachers can provide. When teacher-student relationships are no longer solely built around knowledge transfer, how to better play a guiding, motivating, and exemplary role, and how to reinterpret the importance of teaching by example while maintaining the dignity of the teaching profession, becomes a challenge for teachers. Fifth is the impact on educational content. Mechanical memorization content in textbooks will significantly decrease, leaving space for deep learning, cognitive innovation, and experiential learning. Additionally, potential ideological risks of general artificial intelligence must be guarded against. The ideological biases embedded in pre-training data will subtly influence learners. Sixth is the impact on educational management. The application of artificial intelligence in educational management has matured relatively, and technology has promoted the efficiency, refinement, and scientific nature of educational management, forming many excellent cases across the country and accumulating rich experience. At the same time, it is essential to continue exploring the integrated application of educational management data, improving data governance levels, and strengthening data security supervision^[21].

(2) How Education Can Actively Respond to the Challenges of Artificial Intelligence

Currently, the emergence of generative artificial intelligence has shifted the object of technological influence from human physicality to human cognition, from physical beings to human wisdom and consciousness. The unique characteristic of humanity as a subjective existence—ways of thinking—will also be challenged. We must re-think education, transforming it to promote the awakening of human consciousness and the enhancement of skills, thereby preserving human value and freedom. It remains to be confirmed whether the data-driven implementation methods of generative artificial intelligence are the optimal path. The inherent technical flaws of large models based on probabilistic reasoning and the constraints of resource consumption render the pursuit of increasing parameters and enlarging models without value. When the data-driven dividends are exhausted, whether there exists a third path or whether new research paradigms or technological routes will emerge remains to be seen. On this, we should maintain a questioning and rational attitude.

Taking a broad perspective while focusing on specific issues is crucial. Among the complex elements of the educational ecosystem, addressing the following three questions is most urgent and guiding. First, as China’s traditional educational advantages are significantly weakened by artificial intelligence, what competencies and abilities should be prioritized in student cultivation? Second, as generative artificial intelligence technology develops, how should new teacher-student relationships be managed? Third, how has artificial intelligence changed the ways knowledge is produced and disseminated, and what qualitative differences exist in teaching modes compared to the era of educational informatization?

1. Emphasizing the Cultivation of Students’ Higher-Order Thinking

In the era of artificial intelligence, the goals and models of education shift from knowledge-based and subject-based to competency-based, meaning that the acquisition of immediately applicable knowledge will be increasingly assisted by artificial intelligence. Students are exposed not only to vast amounts of certain information but also to generative content that is difficult to discern as true or false, requiring the enhancement of students’ digital literacy and skills, which are essential competencies for the future. If, in the information age, we required students to have the ability to identify and solve problems, in the era of artificial intelligence, we require students to have the ability to pose questions, even high-quality, logical, and open-ended questions. Formulating good questions is the beginning of effective collaboration between humans and artificial intelligence. Currently, the content generated by generative artificial intelligence is approximately at the average level of human common sense; to bring it closer to or reach peak levels, effective prompts are necessary. This involves higher-order thinking skills such as comparison, analysis, application, transfer, synthesis, and evaluation, while traditional lower-order thinking skills such as memorization, retrieval, and calculation are gradually being replaced by artificial intelligence.

Technology enhances and extends certain human functions, which correspondingly leads to the weakening and atrophy of other functions, resulting in intellectual laziness among humanity. Neuroscience and related experiments repeatedly demonstrate that historical technologies and tools continuously shape the human brain; synaptic connections between neurons will reorganize based on our thinking habits. The internet era leads to information overload, while generative artificial intelligence continuously creates knowledge, potentially leading to shallow thinking among humans. The rich media of the internet and diverse stimuli can excite the brain’s prefrontal cortex, but the hippocampus, responsible for deep thinking, is not activated in this process^[25], encouraging intellectual laziness among individuals. Curiosity and the desire for exploration need to be encouraged and rewarded, while the tendency to “take shortcuts” is inherent to humanity, which may result in collective intellectual decline. The significant threat posed by artificial intelligence is not that it will replace human jobs but that humans may fall into the trap of the powerful functions of artificial intelligence, becoming accustomed to machine-provided solutions and abandoning independent thinking. If humanity becomes accustomed to easily obtaining solutions without engaging in independent thought, fully delegating thinking to machines and artificial intelligence, that will represent the greatest threat to humanity.

Therefore, teachers need to return to the original intention of education, effectively utilizing interactive and heuristic teaching methods, placing greater emphasis on the question-and-answer interactions between teachers and students and among students themselves, focusing on the development of students’ thinking, emotional, and moral growth, rather than merely improving efficiency in classroom processes or increasing the volume of teaching content, to avoid misuse that exacerbates educational involution. This requires teachers to continually enhance their digital literacy and skills, understand the basic principles of content generation and output by generative artificial intelligence, and approach and apply it rationally and objectively in education and teaching.

Currently, when we discuss the “interactive heuristic teaching method,” it represents a return to teaching methods in the intelligent era. This is an innovative practice of the educational philosophies of both Eastern and Western cultures. Socrates advocated the “Socratic method” of teaching, which entails not directly telling students a particular knowledge point but first posing questions to them and allowing them to respond. If students answer incorrectly, he does not directly correct them but instead poses another question to guide them toward the correct conclusion, which he termed “the art of midwifery,” as it facilitates the birth of correct thoughts^[26]. Confucius emphasized the importance of heuristic teaching in “The Analects,” stating, “If one does not feel distressed, one does not stimulate; if one does not feel hesitant, one does not express. If one does not reflect on one corner, one will not return to the other three corners.” In Zhu Xi’s commentary on “The Analects,” he interprets this as the state of cognitive levels where one feels distressed and hesitant, while stimulation is opening the mind and expression is articulating words. In Zhu Xi’s view, the state of distress and hesitation reflects cognitive levels, while stimulation is a method to open the mind and articulate thoughts. In short, the heuristic teaching prevalent in China’s excellent traditional culture emphasizes question-and-answer teaching based on students’ active thinking.

When generative artificial intelligence enters educational settings, the “interactive heuristic teaching method,” which integrates the essence of both Eastern and Western cultures, places greater emphasis on “stimulation” and “interaction,” that is, through the two-way effective questioning interactions between teachers and students, stimulating students to engage in deep learning and cultivate their higher-order thinking. Its characteristics include problem-based inquiry, strong interaction, and strong feedback; only by triggering questions that genuinely invoke deep thought in students and providing timely positive feedback can the brain’s cortex be stimulated, promoting brain activity. When students achieve a certain expected goal, the brain will activate reward systems, secreting dopamine, norepinephrine, and endorphins, allowing students to experience pleasure and joy mentally. The true occurrence of autonomous learning among students happens when they focus solely on learning, free from material rewards and utilitarian objectives, characterized by three aspects: first, the teacher’s stimulating work must be based on students’ active thinking, which can be reflected in students posing questions; second, the shift from primarily teacher-directed questioning in traditional classroom teaching to multi-actor interactions and multi-round question-and-answer exchanges between teachers and students and among students themselves; third, the teacher’s instructional design goals must be reasonable, adhering to the principle of the “zone of proximal development,” and emphasizing timely positive feedback. The interactive heuristic teaching method is not merely synonymous with a specific teaching method; rather, it is a teaching philosophy and guiding ideology that can manifest as a teaching method or be integrated into various teaching methods.

2. Focusing on Building New Teacher-Student Relationships

How teachers adapt to their roles in new teaching relationships, how they engage in human-machine collaborative teaching, and how to address digital ethics issues among teachers and students are all important contents in constructing new teacher-student relationships. By deconstructing the quality structure of excellent teachers and empowering machines through pre-trained models, the goal is to create virtual teachers that are “homogeneous” with excellent teachers. The traditional teacher-student relationship, characterized by “teacher-centered” and “knowledge-centered,” will be weakened or even disappear, while a new teacher-student relationship characterized by “student-centered” and “learning-centered” will gradually emerge. The binary subject relationship of unidirectional transmission will shift to a multi-directional interactive “teacher-machine-student” triadic relationship, forming a new educational ecology. The reason for viewing “machines” as new subjects lies in their continuously developing intelligence and interactivity, which iterates the mechanization and programmability of traditional machine teaching.

Teachers will transition from being “gatekeepers of knowledge” to “choreographers of learning.” First, they should place greater emphasis on guiding students’ emotions, attitudes, and values. The future new teacher-student relationship needs to be more emotional and interactive; future human teachers must learn to coexist with machines, utilizing “machine teachers” to their advantage, requiring greater affinity and empathy to connect with students’ inner worlds and transform education into an “art.” Second, they will gradually become producers of knowledge, facilitators of learning, and guides for growth. Teachers will increasingly take on mentorship roles, guiding students to find correct learning objectives, scientific learning methods, and efficient learning paths, reminding or constraining them to form self-disciplined learning habits, and providing emotional support for students’ comprehensive practice and social experiences. The collaboration between human teachers and “machine teachers” will unfold based on the full utilization of their respective strengths. The advantages of human teachers primarily include supporting students’ social emotional skills, influencing and shaping students with their worldviews, life philosophies, and values, and the ability to integrate knowledge across different disciplines. In contrast, the intelligent technology’s unique advantages compared to previous information technologies lie in its ability to address the challenges of recognizing differentiated learning needs, assisting in implicit cognitive barriers, and adapting to diverse learning paths, making precise teaching possible^[27]. Therefore, the advantages of “machine teachers” currently primarily focus on their vast knowledge reserves, near-infinite computational capabilities, and their memory of problem-solving paradigms, treating each student with infinite “patience” and personalization during interactions^[28].

3. Innovating and Exploring the Transformation of Teaching Models in the Intelligent Era

How to scientifically understand the current development of artificial intelligence technology and its impact on education requires in-depth research by the educational community. The current technology has not matured enough to be systematically, comprehensively, and accurately applied in teaching; placing excessive emphasis on the application of artificial intelligence technology in the micro-environment of education and teaching may be premature. Teachers must first recognize the limitations of current technology. Compared to human intelligence, generative artificial intelligence currently lacks the ability to judge “capability boundaries”; for questions it cannot answer, it provides answers based on probability, which often contains erroneous information. Both teachers and students need to use artificial intelligence safely, effectively, and appropriately, and education should prepare every student to utilize generative artificial intelligence technology or other future technologies well. In this context, teachers should focus on guiding students to enhance their understanding of the essence of generative artificial intelligence technology and its preliminary applications, emphasizing rational judgment before “interacting” with new technologies.

Furthermore, artificial intelligence education represents a qualitative difference compared to educational informatization. In teaching, the chain of thought dialogue between teachers, students, and generative artificial intelligence offers a completely different experience compared to past computer-assisted teaching and using digital educational resource platforms^[29]. This transformation involves qualitative differences in educational subjects, resource provision, content production, and interaction methods, representing not merely an improvement in efficiency in certain aspects of the educational process or an enrichment of resources but a systemic leap towards educational digitization and intelligence based on educational informatization, driving innovation in the underlying logic of education and better realizing the essence of education. For instance, teachers can generate necessary graphic stories or videos through generative artificial intelligence technology to conduct exploratory activities, enhancing their instructional design and organizational abilities, and increasing classroom interactivity, but this does not mean that new technology should be treated as a primary teaching tool. Artificial intelligence technology serves as a process-oriented pathway and crucial driving force in deepening the digital transformation of education; thus, it is essential to accelerate the transformation and application of the “five new” system of new educational forms in the digital age^[30], which encompasses equitable, inclusive, sustainable, and lifelong educational philosophies, shaping a high-quality personalized lifelong learning system where “everyone learns, everywhere learns, and anytime can learn”; constructing a teaching model centered on data-driven large-scale adaptive instruction; innovating educational content focused on competency and skills; and promoting refined management, precise services, and scientific decision-making in educational governance^[31]. Through intelligent technology, we can break existing path dependencies and systematically empower educational transformation, achieving high-quality educational development.