Society is inherently a nonlinear and openly evolving complex network shaped by coupled interactions. How to achieve accurate measurement and prediction of diverse social elements, such as institutions, behavior, and psychology in a limited observation period has long confused the humanities and social sciences. The emergence of generative artificial intelligence exemplified by large language models such as ChatGPT marks the accelerated arrival of a general AI era and is quietly reshaping the epistemic rules of social inquiry. Therefore, more and more scholars have begun to turn to a research strategy that uses generative actors or AI agents to carry out social simulacra. Here the so-called “Social simulacra” refers to conducting social experiments within temporally accelerated simulacra worlds constructed by AI agents. Since the rise of agent-based modeling, this field has continuously negotiated the balance between simplifying reality and explaining it. When AI agents enhanced by large language models can not only reproduce collective behaviors but also generate digital subjects endowed with memory and learning capacities, social research becomes a process of computation, simulation, prediction and generation. Social realities are no longer merely represented but are produced as emergent simulacra through repetition and variation.

The “Smallville” AI town developed by Stanford’s Artificial Intelligence Laboratory offers a compelling illustration of this shift. By embedding AI agents into a dynamic simulation environment that maps the real world, social simulacra not only reconstructs the research paradigm, but also reconstructs the social interaction itself. Personalized agents can transform complex social phenomena into adjustable parameters, enabling researchers to conduct controllable experiments on human behavior, institutional interaction and so on. Unlike real-world experiments that rely on long-term accumulation, the simulacrum world allows researchers to directly manipulate time flows to quickly test multiple possible futures. Although this methodology has significant advantages, generative social simulacra also brings profound ethical challenges. Existing research has discussed the governance of generative artificial intelligence (Stahl / Eke 2023), the relationship between AI agents and ethnographic methods (Jaton / Sormani 2023 ; van Voorst / Ahlin 2024), and the advantages of agent-based models (Epstein, 2006) and large language model-driven agents (Piao et al. 2025). However, the experimental ethics behind such applications is still underdeveloped. When researchers shape the simulacrum world through parameter adjustment, they may transform complex social mechanisms into governable, computable objects, and embed cultural assumptions and historical biases into the model itself. With the help of theoretical resources such as “media evocation” (Van der Goot / Etzrodt 2023) and “materializing morality” (Verbeek 2006), this paper revisits the development of AI agent-driven generative social simulacra research through the combination of deductive speculation and reflective balance, and based on a representative case review. At the same time, this paper examines how this research path fits and promotes social science research, and systematically examines the simulacrum world constructed by agents in generative simulation, with particular attention to the deep ethical issues that may emerge during its experimental process.

From the perspective of post-media, the emergence of AI agents means the individualized turn of media and the epistemological shift from “human-machine” to “human-agent-human”. This change echoes the emphasis of actor-network theory on the initiative of non-human actors. AI agent is no longer just a passive research tool. Through continuous co-construction with human beings and institutional actors, they are actively participating in the production of social reality narratives. Therefore, the ethical implications of its experimental application deserve more in-depth examination. First of all, when the simulacrum environment generated by AI dominates the social science experimental framework, the authority of social knowledge will be challenged. In order to enhance the sense of reality, AI agents tend to simulate emotions and memories, and rely on data sets that reflect specific cultural norms and historical power structures. Therefore, researchers’ expectations may be encoded by algorithms, thus blurring the boundary between analytical modeling and subjective preferences. Compared with traditional digital ethnography, generative social simulacra is easier to compress social complexity into executable variables, and to obscure how power relations are naturally transformed into objective truth through the black box mechanism. Secondly, agent-based simulacra are deeply influenced by the changing power relations among human beings, AI systems and institutional actors. When using human data to train non-human agents, the issue of informed consent emerges. At the same time, whether the emergent collective behavior generated by AI groups has moral status also raises ethical disputes. In addition, the technical basis of contemporary social simulation is still deeply embedded in Western modernist epistemology. The individualistic hypothesis in the training data may conflict with the ethical framework of collectivism. The improvement of large-scale simulation ability also further aggravates the risk of normalization of behavioral manipulation and monitoring research practice. Thirdly, the existing ethical framework is being impacted by the fluid time structure of the simulacrum world. Ethical norms such as the beneficial principle and the justice principle are difficult to apply to those digital subjects that can be infinitely copied and modified. Acceleration time and adjustable parameters distort the traditional risk assessment mechanism, and the replaceability of agent identities and memories also complicates the existing privacy concepts. At the same time, the feedback loop between simulation-based results and real policy interventions further indicates that hyperreal models are increasingly reshaping the social reality they were intended to study. When social simulacra attempts to construct social phenomena that originally require long-term observation, the institutional recognition of the narrative of AI generation behavior may solidify some explanatory frameworks that weaken free will and cultural subjectivity. Furthermore, commercial simulacra platforms may also create new cognitive inequalities by focusing knowledge authority on technologically advantageous institutions. At the same time, the physical infrastructure required to maintain large-scale simulations, including servers, energy, and water resources, has also triggered environmental justice issues that have long been neglected in digital research practices (Crawford 2021: 43-44). Moreover, in this paradigm shift, we must also be wary of researchers’ tendency to “play God”.

This paper holds that ethical guidelines designed for the interaction of real physical individuals are not enough to cope with the simulacrum conditions in the social experiment of AI-generated social experiments. Therefore, we need to promote the transformation of ethical paradigm, like extending the protection of human subjects to data, intelligent derivatives and behavioral trajectories, establishing an ethical review mechanism suitable for generative social simulacra, and ensuring that the simulacrum world is built on controllable, safe and sustainable infrastructure. Instead of pursuing perfect digital replication of society, researchers should retain intentional incompleteness in agent design to maintain the possibility of critical reflection and accidental discovery.

The rise of Artificial Intelligence-Generated Worlds(AIGW) has made ethical reflection and forward-looking governance indispensable. Ethical research in generative social science must go beyond procedural compliance and face the problem of simulation sovereignty. In the post-media era, who controls the simulacra on which society understands itself has become an unavoidable core issue. The ultimate goal of ethical AI agent design should not be to continuously improve the prediction accuracy, but to protect people’s living conditions. In the era of algorithmic determinism, it should retain meaning for those incalculable experiences and defend the freedom of human beings to start again.

Bibliography

Baines, James T. / Taylor, C Nicholas / Vanclay, Frank (2013): “Social impact assessment and ethical research principles: ethical professional practice in impact assessment Part II”, in: Impact Assessment and Project Appraisal 31, 4: 254–260.

Crawford, Kate (2021): Atlas of AI: Power, politics, and the planetary costs of artificial intelligence. New Haven: Yale University Press.

Epstein, Joshua M. (2006): Generative social science: Studies in agent-based computational modeling. Princeton: Princeton University Press.

Jaton, Florian / Sormani, Philippe (2023): “Enabling ‘AI’? The situated production of commensurabilities”, in: Social studies of science 53, 5: 625–634.

Piao, Jinghua / et al. (2025). AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society. in: arxiv: 2502.08691[cs.SI]. https://doi.org/10.48550/arXiv.2502.08691

Stahl, Bernd Carsten / Eke, Damian (2024): “The ethics of ChatGPT- exploring the ethical issues of an emerging technology”, in: International Journal of Information Management, 74, 102700.

van der Goot, Margot J. / Etzrodt, Katrin (2023): “Disentangling two fundamental paradigms in human-machine communication research: Media equation and media evocation”, in: Human-Machine Communication 6: 17-30.

van Voorst, Roanne & Ahlin, Tanja (2024): “Key points for an ethnography of AI: an approach towards crucial data”, in: Humanities and Social Sciences Communications 11: 1-5.

Verbeek, Peter-Paul (2006): “Materializing morality: Design ethics and technological mediation”, in: Science Technology, & Human Values 31, 3: 361–380.