In recent years, with the rapid advancement of artificial intelligence (AI) technologies—particularly the substantial improvement of foundation models, agent-based simulation, and big data analytics—the academic community has begun to reflect on and debate an important question: Can AI replace investigative studies□ This question concerns not only the technological evolution of research methods but also the fundamental logic underlying knowledge production in the social sciences.

At present, AI is gradually evolving from an instrumental application in investigative studies into a cognitive tool capable of simulation, and may ultimately drive a paradigm shift in social science research through its own form of “collective intelligence.” AI systems can process volumes of data far exceeding human capacity in extremely short periods of time, endowing investigative research with near-real-time capabilities for macro-level scanning. Traditional statistical methods typically rely on explicit theoretical assumptions and predefined variables, whereas AI can detect latent structures under conditions of weak assumptions—an evident advantage in contexts where social behavior is highly complex and interactions are frequent.

Nevertheless, fundamental differences remain between AI and traditional investigative studies. The latter engage with real individuals in actual social settings, whose behavior, expressions, and choices take shape within concrete social relations and institutional environments. By contrast, the “individuals” in AI simulations are essentially generated from existing data and model assumptions. Their sociality is encoded rather than lived. Biases exhibited by AI systems trained on real-world data, in a certain sense, reflect social reality and are not inherently “erroneous.” The problem, however, is that algorithmic biases are often embedded within model structures and parameters, making it difficult for researchers to distinguish between genuine social facts and outcomes amplified by models.

More importantly, policymaking concerns not only behavioral prediction but also identity and value coordination. Methods such as in-depth interviews and participant observation often reveal modes of social understanding that statistical patterns cannot capture. This is precisely the irreplaceable value of investigative studies at the level of meaning. Today, investigative research is no longer merely a question of how to obtain data, but of how to define real society.

Although foundation models are already capable of internally simulating interactions, conflicts, and integration among multiple agents, such “agent societies” remain fundamentally different from real society. The “sociality” within these models is a cognitive structural simulation rather than an experiential form of social generation. Such simulations are neither directly embedded in real mechanisms of incentive and punishment nor subject to the structural constraints shaped by social strata, organizations, institutions, and historical trajectories in actual society. They are closer to thought experiments or ideal-type deductions in the sense of social sciences than to direct representations of how real society operates.

This is precisely why agent simulations based on foundation models have clear boundaries when applied to social investigations. They offer significant advantages in theoretical exploration, mechanism inference, and scenario analysis. They can efficiently generate research hypotheses, test the internal consistency of different logical pathways, and simulate multi-actor cognitive interactions in policy analysis and complex decision-making. Even so, these advantages cannot be automatically extended into a capacity to replace empirical investigation. Agent-based simulations cannot measure social facts, accurately reflect the distributional characteristics of social behavior, or fully reveal the profound influence of power relations, institutional constraints, and structural inequalities on actors’ choices. In particular, they struggle to substitute for the empirical observation of silent groups and informal practices.

Overall, AI has not replaced investigative studies, nor is it likely to do so entirely. In the age of AI, the significance of investigative research lies not only in providing a factual foundation, but also in sociologically calibrating, interpreting, and interrogating algorithmic outputs. The future is more likely to bring a hybrid research paradigm grounded in AI and validated through traditional investigative research, rather than the wholesale replacement of investigative studies by AI.

Gao Yuning is a tenured associate professor and deputy dean of the School of Public Policy and Management at Tsinghua University.