During this training session, Qiaoyu Cai, a postdoctoral fellow at the Institute for Advanced Study in Humanities and Social Sciences at Tsinghua University, offered an overview of developments in the global field of AI for Humanities.
I. Foundations and National-Level Initiatives
Policy Guidance and Ecosystem Building
One especially representative example is the “Humanities Perspectives on Artificial Intelligence” initiative launched by the U.S. National Endowment for the Humanities (NEH). Rather than focusing on technological innovation itself, the program centers on AI-related social issues such as democracy and public trust, privacy and civil rights, national security, and economic competitiveness. To date, it has provided more than $6 million in support for humanities projects at multiple levels, including individual research, collaborative teams, and curricular training. The NEH has also supported universities in establishing interdisciplinary research centers. Examples include the Center for AI and Experimental Futures at the University of California, Davis, which emphasizes the democratization of AI and the practice of civil rights; Barnard College’s Wihanble S’a Center, which focuses on the intersection of Indigenous protocols and AI; and North Carolina State University’s EASE Center, which concentrates on the ethics of AI agents. Together, these efforts have created a comprehensive ecosystem linking research, teaching, scholarship, and public engagement.
Another landmark initiative is the Human and AI Virtual Institute (HAVI), launched by Schmidt Sciences, the foundation established by former Google CEO Eric Schmidt and his wife. This project moves beyond the one-way model in which AI merely assists humanities research and instead proposes a framework of “two-way empowerment.” On the one hand, AI can support humanities research; on the other hand, insights from humanities scholars can help advance AI’s capacities for multilingual, multimodal, and cross-cultural understanding. The program encourages applicant teams to include both humanities scholars and researchers from science and engineering so that humanistic values can be meaningfully embedded into technological implementation. Individual projects may receive up to $800,000 in funding, and the program is open to universities and nonprofit institutions around the world.
Europe, by contrast, has placed greater emphasis on infrastructure building. A representative example is the Dutch project “HAICu: Using AI to Access, Connect and Analyze Heritage Collections.” This initiative connects researchers, research groups, national libraries, and cultural heritage institutions. Its goal is to uncover deeper relationships among heritage datasets and to help users generate explainable, traceable, fact-based, and coherent narratives, thereby elevating AI application into a broader project of knowledge governance for public institutions.
II. University Research
Interdisciplinary Collaboration and Paradigm Innovation
At leading universities around the world, research at the intersection of AI and the humanities has taken multiple forms and progressed along several fronts at once. Princeton University’s “Humanities for AI” project emphasizes embedding humanistic values and methods throughout the entire process of AI development and use. It rejects the view that the humanities are merely a “patch” for AI and instead argues that humanistic principles should be incorporated into the design of the parent model at the pretraining stage, so that humanistic commitments and AI technology are integrated from the outset.
Brown University’s “Model-Scale-Context” project takes model, scale, and context as its core analytical framework in order to deepen critical AI studies. The project argues that AI operates differently and produces different effects at the levels of the individual, community, society, and nation-state, and that it is always embedded in specific social structures and cultural contexts. AI therefore should not be treated as a transcendental abstraction; instead, research must begin from concrete situations.
Washington University in St. Louis has adopted a dual-track strategy. On the one hand, it uses AI to support traditional humanities research in literature, media, history, film, and related fields. On the other hand, it uses the strengths and limitations of AI revealed during that research process as a basis for reflecting on AI itself, thereby fostering a productive interaction between tool use and critical inquiry into AI as an object of study.
Particularly aligned with the goals of this competition are relevant research centers at the University of Chicago and the University of California, Irvine. At the University of Chicago, AI research focuses on the reconfiguration of humanities inquiry in the age of generative media. Using AI as a point of entry, it promotes deep integration among literature, linguistics, philosophy, sociology, and computer science, while examining how generative models are reshaping modes of knowledge production and methodology. UC Irvine, meanwhile, has advanced a humanities-centered AI strategy at the school level, extending the significance of AI for humanities research into the realms of policy-making and resource allocation, and thereby achieving a systematic breakthrough from teaching and research to institutional design.
Universities in China have also been acting actively. Fudan University led the publication of The Future Is Already Here: Blue Book on the Intelligent Development of the Humanities and Social Sciences, while also hosting a conference on intelligent humanities and social sciences. Together, these efforts provide a systematic review of AI-driven theoretical innovation and paradigm transformation, while attempting to consolidate scattered research practices into a shared academic agenda. This marks an important step in the more systematic exploration of the field within China.
Qiaoyu Cai emphasized that these global developments show that research at the intersection of AI and the humanities has moved beyond the early stage of mere tool application and is now entering the deeper waters of paradigm reconstruction. This competition seeks to build on that intellectual trajectory by encouraging participants to move beyond the superficial question of “how to use AI” and instead to explore the more fundamental question of “how AI is reshaping humanities research.”
Zhang Guangwei, a lecturer at the School of History and Culture and the Institute of Area and Country Studies at Shaanxi Normal University, as well as Director of the Silk Road History and Culture Virtual Simulation Experimental Teaching Center, then shared his personal experience and reflections as a practitioner working at the intersection of AI and historical research, offering participants a highly valuable practical model.
I. From Tool to Partner
The Evolving Role of AI in Historical Research
Zhang Guangwei divided his research trajectory into two phases—the pre-large-model era and the large-model era—clearly illustrating how AI has evolved from an auxiliary tool into an academic partner.
From 2016 to 2023, during the pre-large-model era, Zhang’s team focused primarily on applying computer-vision-based deep learning to historical research. Working with sources such as Tangut texts, ancient steles and epitaphs, and local gazetteers, the team developed OCR systems that enabled full-text digitization of historical materials. In particular, in the area of ancient map recognition, the team used OCR to identify individual characters and then piece them together into correct place names, making possible the intelligent retrieval and visualization of place names on ancient maps. On their intelligent ancient-map digitization platform, users need only enter a keyword to locate ancient maps containing that term and automatically pinpoint it on the image. This has made ancient maps searchable and has greatly improved the efficiency of historical geography research.
A representative achievement from this phase was the team’s work on the digitization of the Tangut dictionary Wenhai, based on its Tangut OCR system. Using network analysis, the team modeled and systematically studied the dictionary’s explanatory structure. Since most Tangut characters in Wenhai are explained by two other Tangut characters, this method transformed a flat dictionary designed for human reading and learning into a networked structure. With the help of algorithms, the team could rapidly identify circular explanatory chains and strongly connected components, enabling a layered study of Tangut semantics. The results were published in the journal Digital Humanities. Zhang noted that at this stage AI still functioned primarily as a tool, but it laid a solid data foundation for more advanced research later on.
At the beginning of 2023, with the sudden arrival of ChatGPT, researchers across many fields began experimenting with possible applications. Zhang’s team was among the first to use it in the processing of historical documents. At first, AI’s performance in punctuating classical Chinese, translation, and entity recognition was imperfect—intelligent, but not yet sufficiently specialized. However, after carefully studying the GPT-3 paper, the researchers discovered that large models possess in-context learning abilities. By providing just a small number of examples, AI could quickly acquire knowledge in a specialized domain. For instance, after being given several examples of punctuated Tang-dynasty documents, the model’s ability to process subsequent texts improved dramatically, reaching or even surpassing that of untrained students.
As the team’s understanding of large models deepened, their research began moving into deeper waters. In a paper presented at DH2024, the flagship international conference in digital humanities, Zhang’s team attempted to use large models to reconstruct ancient road and transportation networks from historical documents. The model successfully extracted place names and distance information from the sources, and when combined with the team’s intelligent ancient-map digitization platform, it enabled the construction and direct visualization of the road network of Qing-dynasty Jiangxi Province on ancient maps. The paper received a perfect score from one reviewer. Building on this work, the team continued its research and, using a locally deployed DeepSeek-R1 32B model, implemented functions for both “searching images through text” and “searching texts through images.” This work allows AI not only to identify place names and the distances between them from the vast body of local gazetteers, but also to assist in constructing road networks and visualizing them on ancient maps.
At this stage, a landmark “aha moment” emerged. When Zhang Guangwei uploaded his paper on the network analysis of the Tangut dictionary Wenhai into Google NotebookLM, the AI generated an extremely professional audio dialogue discussing the paper. The conversation not only captured the article’s central argument with remarkable accuracy, but also pointed out a potential connection to anthropological theory that the author himself had not noticed. This suggested that AI had already acquired a fairly sophisticated capacity for understanding, and could serve as a qualified interlocutor capable of offering feedback and expanding one’s thinking.
Yet when confronted with historical corpora running into the millions of characters, a simple question-and-answer model was no longer sufficient. In the course of addressing real research problems, the team developed and proposed the concept of Knowledge Protocol Engineering (KPE), which aims to embed the methodological expertise of human scholars into the construction of multi-agent systems, enabling AI to participate meaningfully in reasoning through complex historical questions.
KPE is not merely a matter of programming; it is a logical framework. If RAG equips AI with an encyclopedia, then KPE provides it with a domain-specific operations manual. In humanities research, a great deal of expert knowledge is tacit and internalized. The goal of KPE is to make that tacit knowledge explicit and formalized, turning it into protocols that AI can execute.
Although large language models are powerful, their probabilistic nature makes their outputs inherently unstable. For rigorous historical research, such instability can be fatal. Through KPE, researchers can solidify disciplinary methodology into a structured workflow. When AI encounters uncertainty, the protocol can require it to return to the human researcher for clarification, thereby creating a human-in-the-loop model of collaboration.
In an age when AI can efficiently handle factual recall and basic reasoning, what remains the value of the historian? Zhang argued that scholars should shift from being mere repositories of knowledge to becoming constructors of meaning. AI may remember more and calculate faster than we can, but asking profound questions, making value judgments, and constructing the larger meaning of history remain uniquely human tasks. KPE serves precisely as the bridge between the two: humans design the protocols (defining questions and methods), AI executes the protocols (processing data and reasoning), and together they produce knowledge.
Within a multi-agent system, different AI agents assume different roles: some interpret user intent, some retrieve evidence from a knowledge base, some summarize materials, and others assess the logical gaps between conclusions and evidence. In effect, this design simulates the research workflow of the historian. When working on one unattributed letter among the tens of thousands of documents in the Sheng Xuanhuai Archives held by the Art Museum of the Chinese University of Hong Kong, Zhang’s multi-agent system demonstrated a document-processing capacity strikingly similar to historical scholarship: in the first round it made a preliminary judgment; in the second, it introduced new evidence and identified contradictions; in the third, it revised its own conclusion, ultimately arriving at a highly credible result. This case suggests that AI can already intervene in two of the most central operations of historical research: textual criticism and reasoning.
II. The Role of the Researcher
From “Questioner” to “Architect”
Drawing on years of practice, Zhang Guangwei shared three central insights.
First, traditional deep learning and rigorous digitization work are the foundation of AI-humanities research. The impressive performance of large models depends on high-quality data; without the prior digitization of Tangut texts, ancient maps, and other materials, later advanced research would not have been possible. The core value of humanities research lies in the problem itself. AI is only a means of addressing that problem. One should not use AI for its own sake; rather, the problem should determine how the technology is used.
Second, researchers need to increase the “bandwidth” of their communication with AI. At present, most people still interact with AI through a chat box, but this low-efficiency mode of exchange cannot fully unlock AI’s potential. It is like using dial-up internet in the age of gigabit broadband: it limits AI’s ability to understand context and mobilize knowledge. Through programming, the construction of multi-agent systems, and similar approaches, one can significantly increase the density and efficiency of human-machine interaction, allowing AI to better grasp the core needs of humanities research.
Third, the concept of KPE is key to achieving a deeper symbiosis between AI and the humanities. KPE does not require everyone to become a programmer. Rather, it translates the tacit research methods and thought patterns of humanities scholars into computable, executable protocols, thereby becoming an operations manual for AI in this field. Through the closed-loop model of “humans design protocols, AI executes protocols, humans revise protocols,” it enables a healthy human-in-the-loop relationship. As the technology matures, this may gradually evolve into the more efficient model of human-on-the-loop.
From early OCR and ancient-map recognition based on deep learning, to the use of large models to reconstruct ancient transportation networks, and then to multi-agent systems for historical archival analysis, Zhang especially emphasized that in the age of AI, the value of historical scholarship lies not in memorizing historical facts—an area in which AI already far surpasses humans—but in asking profound questions, making value judgments, and constructing historical meaning. Humanities researchers should cultivate a relationship of mutual learning and mutual teaching with AI, deepening their understanding of their own disciplines through collaboration with it and thereby transforming the very mode of knowledge production. In this process, AI no longer simply answers questions; it is incorporated into complex research workflows and guided, through protocols, structures, and rules, to carry out an entire set of humanities methodologies.
During the discussion session, participants raised a number of incisive questions about technology, ethics, and submission details, and the organizing team responded to them one by one.
III. Questions on Scholarly Practice
1. Choosing Between Traditional Machine Learning and Large Models
Q: Compared with the instability of large models, is traditional machine learning more reliable for applications in the humanities?
Zhang Guangwei: Because large models generate results probabilistically, instability is indeed a real issue. But traditional machine learning can serve as a verification tool. For example, by vectorizing and embedding texts, researchers can check whether information extracted by a large model—such as character relationships—matches the original text above a certain similarity threshold (for example, 0.95), thereby enabling traceable validation of the result. The two are not opposed; they can complement one another.
2. Copyright and Data Sources
Q: How should copyright issues be handled in the application of RAG technologies?
Zhang Guangwei: In response to concerns about copyright in RAG applications, Zhang pointed out that using a RAG system in academic research is essentially similar in nature to an individual reading the relevant literature, and therefore does not in itself create a copyright problem. However, one must still be careful to avoid hallucinated information generated by AI. Data sources may include university-subscribed databases, publicly available academic articles, and journal literature. Individual researchers can use large models on their personal computers by calling cloud-based APIs, without needing high-performance hardware for model training.
3. Do the Humanities Need “Double-Blind Experiments”?
Q: A participant from the medical field asked whether humanities research could adopt double-blind or triple-blind experimental designs to reduce the influence of AI hallucinations.
Zhang Guangwei: The core of the humanities lies in critical reflection and the construction of meaning, which differs fundamentally from the goals of medical experimentation. It may therefore be unnecessary simply to transplant the double-blind model into the humanities. That said, the risk of hallucination can be reduced through multiple rounds of human-machine interaction and cross-validation using different AI tools, while the researcher’s own critical judgment remains indispensable.
4. Entry Paths for Beginners and Interdisciplinary Researchers
Q: How should undergraduates from a purely humanities background, or science and engineering researchers crossing into the humanities, get started?
Zhang Guangwei: First, they need to move beyond the “good student” mindset and learn to become a mentor to both themselves and AI, proactively setting research goals and collaborative directions. Second, they can use AI to lower the barriers to programming—for instance, by using natural-language instructions to have AI generate code—and gradually build interdisciplinary competence. Third, they should begin with concrete problems, first using AI to handle foundational tasks such as document processing and data organization, and only then moving gradually toward deeper levels of paradigm innovation.
5. AI and Dead Scripts or Variant Characters
Q: How can AI accurately translate dead scripts such as Tangut, as well as variant and vulgar characters found in Dunhuang manuscripts?
Zhang Guangwei: Multimodal large models cannot directly solve the recognition and translation of low-resource dead scripts; dedicated OCR models still need to be developed. As for variant characters, if the goal is bibliometric analysis, a certain degree of recognition error may be acceptable, provided that later verification ensures the reliability of the research. But if the work concerns specialized philological study, then human proofreading must be combined with AI assistance.
6. AI in Qualitative Research
Q: What are the possible pathways for applying AI to theoretical innovation and qualitative research?
Zhang Guangwei: Large models have tremendous potential in qualitative research. By constructing intermediate representational forms such as character relationship networks, they can help bridge qualitative and quantitative research. Multi-agent systems can also participate in the construction of complex theories, generating unexpected sparks of thought through iterative reasoning and thereby serving as catalysts for theoretical innovation.During