Embodied cognition in the age of AI should be further researched. Photo: TUCHONG

As one of the most significant developments in generative artificial intelligence (AI) in the 21st century, large language models (LLMs) have had an extremely wide-ranging influence on daily life and work, steadily promoting an epistemological shift in how children learn. LLMs function as nodal technologies in the ongoing transformation of digital learning. The cognitive shifts they introduce differ from earlier phases of digitalization, further accelerating the move from bodily perception toward machine-mediated perception and bringing both opportunities and challenges to education in the intelligent era.

Digital & analog modes

Geoffrey Hinton, the 2018 Turing Award laureate known as the “father of deep learning,” outlined two modes of knowledge sharing within an “agent community” in his May 25, 2023 lecture at the University of Cambridge, titled “Two Paths to Intelligence:” the “digital mode” and the “analog mode.” LLMs represent the digital mode, while humans represent the analog mode. LLMs consist of numerous terminals, each of which can be regarded as an agent capable of instantly accessing the learning outcomes of other agents, creating an information environment characterized by seamless sharing and exchange. By contrast, humans share knowledge far less efficiently. Differences in the structure of individual neural networks make lossless data transmission—akin to weight-sharing in machine learning—impossible. Although the analog mode is far less efficient in data transmission, it embodies higher-dimensional value. As Einstein observed, the assumption of a strict causal explanation of nature does not originate from the human spirit. It is the result of the long-term adaptation of human reason.

The widespread application of tools operating in the digital mode compels deeper reflection on the continuing role and value of the analog mode. Put differently, in an environment where tools are capable of lossless transmission, sharing, and autonomous learning, it becomes essential to articulate the advantages humans retain in cognition, learning, and labor. Echoing David Ricardo’s theory of comparative advantage, identifying these human strengths is fundamental for designing education that looks to the future. At this point, while generative AI has effectively closed the chapter on “cramming-style” learning, it has also ushered in a new stage of transformation for analog-mode cognition based on the digital-mode tool environment.

The transformation of the digital environment has, in turn, altered the cognitive approaches of learners. Under the influence of the digital-mode technological environment, children exhibit greater curiosity and higher cognitive plasticity. Their analog-mode thinking has undergone noticeable change: They are moving from interactions rooted in linear, singular, and immediate physical environments toward interactions defined by nonlinearity, multiplicity, and remote presence. Human understanding has always emerged through a sense of “spatial” interaction, but in the digital era, this spatiality is no longer confined to the physical world—it increasingly extends into digital space.

Machine perception

In digital environments, machine perception has emerged as a new way of understanding the world. The way humans perceive the world has progressed through several stages: direct bodily perception; perception aided by physical tools such as eyeglasses; perception mediated by virtual tools such as search engines; and now perception mediated by intelligent machines like ChatGPT. With the rapid development of AI and virtual reality technologies, AI can now interact with humans as an independent “other,” learning and iterating on its own and translating human behavior into data parameters. At the level of interaction with the world, technology can circumvent human subjective perception, transform the real world into a world of data, and, through the recombination of vast datasets, construct multiple virtual worlds that differ markedly from physical reality. In this context, learning—an essential means of understanding the world—must rely not only on direct bodily perception and contact with things, but also on gradually mastering machine perception in order to make sense of a world increasingly mediated by technology. Therefore, as the era of generative AI unfolds, future-oriented education must anticipate these developments and adjust its strategies in a timely manner.

The learning environment shaped by LLMs imposes new expectations on children’s learning. The growing prevalence of digital-mode tools urges educators to re-examine the importance of analog-mode thinking. In responding to the cognitive challenges posed by generative AI, children’s future learning should emphasize interdisciplinary thinking, the capacity for competent human-machine interaction, and the ability to reflect critically on symbolic systems.

New generative AI technologies require learners not simply to train their thinking and accumulate knowledge from a particular vantage point, but to approach problems in a comprehensive, cross-domain manner. In traditional industrial settings, education prioritizes the cultivation of one-dimensional, highly instrumental, and stable workers—skills that reinforced production efficiency. Consequently, traditional learning relies heavily on the repetitive transmission of basic knowledge and technical skills. But as AI, which excels in knowledge storage and combination, increasingly replaces learning activities based mainly on knowledge transmission, it is necessary to re-examine the value of learning and embrace meaningful transformation. The knowledge economy demands integrated, applicable knowledge rather than the fragmented, context-free knowledge emphasized by older pedagogical models. In an LLM-mediated learning environment, treating disciplinary boundaries as rigid boundaries of thought will detach learners from the way knowledge is produced in the future—and from genuine understanding of the world. Learning objectives in this context should focus on disciplinary integration and guide children away from merely “answering a question in a certain subject” toward “exploring solutions to real-world problems.”

Embodied learning

As a learner, a child’s world is self-centered yet inclusive, and the ability to represent material reality accurately depends on the gradual coordination of observing, listening, and touching schemas. However, with the proliferation of AI and remote-presence technologies, the concept of the “surrounding world” has changed. A form of “deterioration of distance” has emerged, changing the way the world is understood from a “human-world” to a “human-technology-world.” In environments shaped by artificial general intelligence, children’s sensory schemas now shift rapidly from the physical world to the digital world, making human-machine collaborative cognition an urgent epistemological issue.

In such contexts of human-machine interaction, children now inhabit two overlapping worlds. The first is the tangible world—as it truly exists—that provides direct physical feedback. In epistemological terms, this world differs little from that of previous generations. The second is the world projected through technological intermediaries, which is the direct object of technological change. Early interpretive tools—such as books—had already made it possible to map distant places and experiences, offering both descriptive accounts and analytical interpretations. For this reason, people have long grown accustomed to understanding the world through mediated technologies and relying on them as primary avenues for learning. What LLMs alter is the descriptive and analytical attributes of these mediating tools through which the world is understood. While books merely serve as symbolic carriers with human-constructed language systems, LLMs and other generative AI technologies actively assume analytical functions, demonstrating the characteristics of other hetero-relational technologies and thus possessing a form of “quasi-otherness.” Today’s digital technologies are already capable of supporting personalized learning, serving as virtual instructors, and even facilitating forms of emotional interaction between humans and machines. In this context, the essential competencies required of children in an LLM-mediated learning environment include the ability to adapt to human-machine collaboration, the capacity to develop cognition grounded in machine-mediated perception, and openness to emerging technologies.

From early childhood onward, nearly all meaningful teaching relies on language as a ”signification system.” Before the advent of LLMs, the signification produced in different teaching scenarios remained largely discrete. However, digitalization and intelligent technologies have accelerated the unification of these systems. Technology has, in essence, constructed a single, unified virtual intermediary world. In this digital world, people experience something akin to Jaques Lacan’s “mirror stage,” identifying themselves with the visual gestalt of their own bodies. Every technological design choice reflects the cultural background and conscious or unconscious biases of its creators. Since many of the most influential platforms are operated by well-funded companies, it is foreseeable that children using these corporate-controlled platforms will engage in machine-mediated perception shaped by norms created for detection and regulation rather than personal empowerment. This makes it essential to cultivate children’s critical thinking regarding information flows, their ability to filter information, and their capacity to reflect on symbolic systems within LLM information cocoons.

Learning in the intelligent era requires consideration of the relationship between “virtual situations in the real world” and “real situations in the virtual world.” As digital technology, AI, and networked environments converge, machines endow human beings with a new kind of “body,” extending physical presence into virtual space. With the combined influence of biotechnology, digital technology, and intelligent systems, the human body can no longer be understood solely as a biological organism; it also functions as a cultural body in the social sense, and on this basis becomes a technological body formed through new modes of embodiment in digital settings. Therefore, “machine perception” in the digital context, though seemingly “disembodied,” is actually a new form of perception in the virtual context. It is not opposed to “embodied cognition,” but instead an expansion of the research on “embodied cognition” learning based on a new definition of “body,” offering directions for understanding “embodied” cognition in the age of intelligent technologies.

Zhang Jingwei is an associate professor from the Faculty of Education at Northeast Normal University. This article has been edited and excerpted from Journal of Educational Studies, Issue 3, 2025.