In August 2025, China’s State Council issued the Opinions on Deepening the Implementation of the “Artificial Intelligence Plus” Initiative, explicitly calling for philosophy and social science research to move toward human–machine collaboration. As large language models (LLMs) make rapid inroads into these fields, however, their expanding role raises a fundamental question: Can artificial intelligence (AI) serve as a universal research tool, equally suited to every branch of the social sciences? CSST recently spoke with scholars across a range of disciplines to examine where its limits lie.

Experiments in paradigm transformation

At the Institute for Studies on Artificial Intelligence and Law at Tsinghua University, scholars in law, computer science, and sociology have used judicial documents to build standardized legal knowledge graphs and independently developed “Shuimu Zhifa,” an intelligent legal service platform offering the public case-similarity searches, assisted document generation, and intelligent legal consultation.

Shen Weixing, director of the institute, explained that law has been among the first disciplines to undergo AI-driven paradigm innovation because it possesses inherently computable foundations: Statutory provisions, judicial rulings, and case classifications form a highly standardized body of texts; legal reasoning rests on rigorous deductive logic; adjudicative outcomes are governed by relatively clear, quantifiable criteria; and the vast body of publicly available judicial documents provides stable, high-quality annotated data for model training.

Meng Tianguang, Party secretary of the School of Social Sciences at Tsinghua University, is leading a different line of experimentation: using LLMs to generate “silicon-based samples” for social research and policy simulation. Meng noted that LLMs and intelligent-agent simulations are further extending the reach of big data analysis. Three modes of human–machine collaboration are gradually taking shape in the social sciences, opening new avenues for addressing long-standing research challenges: in-house collaboration, in which social science teams develop their own AI capabilities; external collaboration between social science and computer science teams; and direct human–machine collaboration, in which researchers use intelligent tools to conduct analysis.

Avoiding one-size-fits-all judgments

While computational law and computational political science have achieved intelligent transformation, many academic institutions have sought to extend standardized computational logic to more speculative disciplines such as ethics, philosophy, and literary studies, yet these efforts have encountered practical obstacles. “Computational ethics,” in particular, has found a small number of applications in China and abroad, but remains largely at the basic auxiliary-tool stage.

Sun Xiangchen, a professor from the School of Philosophy at Fudan University, observed that law has relatively unified statutory norms, clear boundaries of argumentation, and broadly shared evaluative logic, allowing stable classification systems to be built for machine computation. By contrast, ethical theories are diverse, and moral judgments are highly dependent on cultural context and specific circumstances. Across different groups and situations, value choices often admit no single correct answer. Furthermore, LLMs cannot fully explain to humans how such moral decisions are reached, making systematic computation difficult.

Li Yaming, a research fellow from the Institute of Philosophy at the Chinese Academy of Social Sciences, noted that AI can identify facts and extract contextual parameters according to pre-set rules, allowing it to assess whether conduct conforms to a given regulatory framework. It does not, however, possess the underlying capacity required to make genuine value judgments.

Scholars interviewed argued that the application of “AI +” across the social sciences should follow a tiered, category-specific approach, weighing factors including the degree of knowledge structuring, core research objectives, and the extent to which value-based reasoning is involved. Xu Xiaoke, a professor from the School of Journalism and Communication at Beijing Normal University, warned that “techno-utopianism” tends to overestimate both AI capabilities and the pace of technological progress, promoting wholesale intelligent transformation without considering whether particular research tasks are suited to it. Conversely, “techno-rejectionism” sets AI in rigid opposition to qualitative humanities research, dismissing the potential of technological development on the basis of the limitations of current models.

Zhan Weidong, a professor from the Department of Chinese Language and Literature at Peking University, cautioned against rigid, discipline-based stratification. Instead, the division of labor between humans and AI should be determined at the level of concrete research workflows. The degree of permissible AI intervention hinges on a discipline’s knowledge production paradigm: How research questions are posed, explanations developed, evidence verified, and research outcomes produced. Moreover, “AI + social science research” must observe three fundamental red lines: AI-generated evidence must be fully traceable; researchers must retain ultimate responsibility for value judgments; and synthetic simulated data cannot simply replace genuine fieldwork and social survey samples.

Interviewees stressed that the integration of AI and the social sciences should not pursue indiscriminate, across-the-board intelligent transformation. Instead, it should foster a sustainable ecosystem with clear boundaries, centered on human researchers, and supported by AI. For these scholars, AI remains a tool for expanding cognitive horizons and enriching research methods; it cannot replace scholars in posing research questions, constructing theories, making value judgments, or providing humanistic interpretation.

Wu Chao, an associate professor from the School of Public Affairs at Zhejiang University, called for ethical requirements to be embedded throughout the technology R&D process, with model bias constrained during algorithm optimization to reduce risks such as group discrimination at the foundational technical level. He added that generative multi-agent simulations could be used to model diverse social scenarios and explore new mechanisms for social simulation and prediction.