A scholar interpretes a tablet inscription. In the AI era, historical imagination and embodied fieldwork remain critical. Photo: TUCHONG

The growing capacity of large language models (LLMs) to generate texts has begun to make itself felt within historical scholarship, attracting sustained attention from the academic community. Historical writing has always unfolded along two intertwined axes: the pursuit of truth and the work of interpretation. However strongly the discipline may aspire to objectivity, it cannot dispense with judgment, selection, and evaluative choice.

Humanity in embodied experience

A central concern regarding historical texts produced by LLMs lies in their lack of “a humanistic dimension.” This absence is not simply a technical shortcoming, but reflects a broader tension that frequently arises when technical instruments are introduced into the humanities, often resulting in an imbalance between instrumental rationality and humanistic values.

Digital humanities, which has developed rapidly in recent years, exhibits comparable limitations. This, in turn, compels us to confront a fundamental constraint of artificial intelligence (AI). Algorithm-based systems are unable to generate a genuinely humanistic spirit, as contemporary AI remains bound by its technical architecture and lacks higher-order cognitive capacities. The emergence of such capacities may depend less on ever-expanding computing power or storage capacity than on lived, embodied experience. In historical research, scholars are expected to set aside rigid dogmas and preconceived bias, to integrate ideas through reflective engagement, and to apprehend the intrinsic connections between thought and lived experience, along with the value implications embedded within ideas themselves. Reaching this level of understanding constitutes a central aspiration of historical scholarship and underlies what many scholars describe as empathy.

Systems such as ChatGPT and DeepSeek represent remarkable technological breakthroughs, yet they lack this experiential foundation. Their knowledge is derived entirely from datasets. Regardless of parameter scale or the apparent sophistication of their outputs, the absence of lived experience makes it difficult for them to possess genuinely humanistic qualities. At the present stage of technological development, LLMs are unable to perceive the subjective intentions of historical actors, fully grasp the origins and trajectories of ideas, or register non-rational factors through algorithmic means.

The importance of lived experience in historical research also finds support within the philosophy of science. As early as the 1950s, the American scholar Norwood Hanson argued that all observation is inevitably shaped by the observer’s prior theories, assumptions, beliefs, cultural background, and knowledge structures. Historians, as observers of historical materials, draw upon diverse life experiences that furnish them with multiple perspectives on the real world, thereby enriching historical interpretation and writing. By contrast, LLMs lack personalized experiences and a sense of “presence.” Reliant on unified core algorithms, they struggle to produce scholarship marked by intellectual particularity. In extreme cases, this limitation risks approaching the form of technological alienation cautioned against by Martin Heidegger, reducing historical writing to “cookie-cutter” standard content, like industrial products produced on an assembly line.

Ecosystem mindset

Despite these constraints, AI in historical research should not be understood merely as a technical tool. When attention shifts to emerging directions in historical scholarship—particularly environmental history—additional insights come into view. Environmental history situates human history within nature itself, conceiving it as one element within a complex and evolving ecosystem. If this conception of an “ecosystem” is expanded, AI may also be understood as part of it.

In the era of LLMs, concerns are widespread that large numbers of jobs may disappear. In practice, however, while AI may displace certain forms of labor, it also generates new roles, such as prompt engineers. The development of AI both depends upon and reshapes its social environment, a reciprocal process that reflects the complexity, diversity, and vitality characteristic of the “ecosystem.”

To grasp the future direction of historical research, it is necessary to adopt an ecosystem mindset—one that allows historical inquiry to assume more diverse forms, extend across longer temporal horizons, and operate over broader spatial ranges. Human history and historical writing are themselves components of such an ecosystem and will inevitably be shaped by the diffusion of LLMs. As social structures and cognitive frameworks continue to evolve, LLMs may prompt historians to reflect more deeply on the nature of history, reassess their values and modes of thinking, and reconsider established methods and theoretical approaches. When ChatGPT was first released, its wide versatility and low threshold for use sparked extensive public discussion and generated a palpable sense of pressure within academia. Given their operating principles and the biases inherent in their training data, it is unsurprising that LLMs often perform poorly when confronted with specialized historical questions. Even so, their value should not be dismissed outright. Rather, their outputs warrant a cautious and critical approach, particularly in light of the fact that they represent a compressed synthesis of accumulated human knowledge.

From a practical standpoint, LLMs may function as a kind of co-pilot for historians. In tasks centered on information retrieval and large-scale data processing, historians often find themselves at a disadvantage. Yet within the field of digital humanities, technical tools cannot replace scholarly interpretation, nor can LLMs displace historians from the research process. Instead, the AI era points toward the possibility of a model of human–machine collaboration grounded in their respective strengths. Under such an arrangement, the division of labor between the main driver and the co-pilot assigns LLMs an explicitly instrumental role, thereby forming the practical foundation for human–machine collaboration. AI systems take on mechanical and routine tasks, while historians remain responsible for judgment, interpretation, and decision-making. Through a high degree of integration, the two can complement one another’s strengths and jointly advance historical research. This configuration is not accidental, but resonates with Marshall McLuhan’s classic proposition that media function as extensions of the human body. Through sustained human–machine interaction in historical writing, the role of LLMs as extensions of the scholarly mind may become increasingly apparent.

The emergence of LLMs, together with major advances in generative AI applications, is likely to accelerate cycles of technological change. When considering the relationship between historians and LLMs, it is therefore necessary to leave room for future transformation. The possibility that AI may become an integral component of human life, propelling humanity toward a new stage of development, can no longer be dismissed as purely speculative. Indeed, the science-fiction inspired idea of human–machine symbiosis has long existed as a conceptual prototype within philosophical reflection.

Symbiotic relationship

With technological advances—such as gene editing—continuing to alter the human living environment, theorists have begun to engage more seriously with the possibility of posthumanism. As a theoretical orientation, posthumanism challenges traditional anthropocentric views, emphasizing how sci-tech progress reshapes understandings of the body, consciousness, and society. The intellectual origins of this line of thought can be traced to discussions of cybernetics in the 1950s, a school of thought that tended toward technological determinism and held that, once information technology reaches a certain stage, human–machine symbiosis would emerge as a natural outcome. To treat AI as an entity capable of coexisting with humans is, in a sense, to acknowledge a form of technological subjectivity. In a future where a “technological singularity” is widely expected to occur, human social development may therefore reach a critical turning point.

Within this context, interaction between humans and machines should not be understood as an effort to “biologize” machines or “engineer” life to achieve a “marriage between artificial and natural,” nor as a scenario in which AI replaces historians. Rather, it suggests the possibility of positioning AI as a companion species, enabling a transition toward posthuman historiography.

Viewing AI as a companion species of historians emphasizes that “the history written by humans need not be solely human history.” The cognitive premise of this approach lies in the fact that historical writing is a comprehensive record of the past, focusing on the relationship between humans and non-human organisms. Introducing AI as a companion species into the research process will not only expand methodological approaches but also open up new research questions. However, technical barriers to LLM use in historical research remain, and the extent to which such a symbiotic relationship can be realized remains uncertain.

Historical imagination & scholarship

While the future may inspire a degree of optimism, the challenges of the present are no less pressing. The evolution implied by human–machine symbiosis does not necessarily unfold in a uniformly positive direction. The convenience of automated text generation enabled by LLMs may foster intellectual complacency. In the teaching of historical writing, a key challenge lies in determining how students can be introduced to LLMs while still cultivating independent thinking and originality. Commercial enterprises have already developed AI-assisted academic writing systems aimed at standardizing scholarly production, a trend that has generated considerable anxiety among university historians. When ChatGPT first appeared, many universities responded with outright bans, fearing that such technologies would undermine higher education. For emerging technologies, however, passive restriction is rarely as effective as active guidance. From the perspective of historical research, several measures may help ease anxiety surrounding AI. First, historians should continue to emphasize the use of multiple sources of evidence, thereby reducing overreliance on LLMs. Second, it is essential to remain clearly aware of the limitations inherent in these systems. Finally, scholars must recognize that the selection of training data reflects the biases of those who design and deploy LLMs. Although history is fundamentally a discipline oriented toward truth-seeking, historical research also depends on imagination, which remains an indispensable path toward historical understanding. Rapid technological progress may exert negative effects on imagination and modes of thinking. Therefore, in this context, safeguarding historians’ imaginative capacities has become an urgent task.

For society as a whole, LLMs represent a significant leap in the presentation of knowledge, lowering barriers to information acquisition and academic writing. At the same time, the increasing similarity of training datasets and algorithms across models undermines the spirit of academic innovation that advocates for diversity and free exploration. Thus, in the age of LLMs, what most truly reflects the core competitiveness of historians is no longer their grasp of historical knowledge or their monopoly on materials, but their scholarship itself. Without a strong grounding in historical thinking, users cannot effectively guide these tools and may instead find themselves guided by them. Therefore, for historians to truly master LLMs, sustained training in historical thinking and the preservation of disciplinary professionalism remain essential.

Wang Tao is a professor from the School of History at Nanjing University. The article has been edited and excerpted from Historical Research, Issue 5, 2025.