The rapid advancement of Large Language Models (LLMs) is catalyzing a profound transformation in psychological research. This review advances a dual-view framework for critically assessing LLMs: as powerful simulators of human participants and as novel non-human agents for scientific inquiry. As simulators, LLMs demonstrate considerable potential in replicating human responses. Empirical studies have shown that models can mimic human cognitive biases in decision-making tasks, achieve performance on the Theory of Mind (ToM) tasks comparable to that of young children, and successfully replicate a significant portion of main and interaction effects from classic psychology experiments. However, this paper moves beyond cataloging these successes to deconstruct the simulator concept itself, revealing two fundamental, principle-level limitations. First, the model's core compression principle, while effective at capturing group averages, inherently struggles to replicate the internal structural associations of individual differences. Second, a profound asymmetry exists between the model's training phase (knowledge acquisition) and its alignment phase (strategic expression), which fundamentally determines that any simulation is a strategically optimized expression rather than a faithful representation of the model's internal knowledge state. Consequently, their inherent limitations and systematic biases - such as reflecting a predominantly Western, Educated, Industrialized, Rich, and Democratic (WEIRD) perspective - are not merely data-level flaws, but deep-seated mechanistic properties. This transforms them into valuable data for investigating the mechanisms of non-human intelligence. This critical perspective aligns with recent scholarly calls to systematically analyze the theoretical fallacies, such as anthropomorphism and identity essentialization, that arise from uncritically substituting LLMs for human participants.

Within this theoretical framework, the paper systematically organizes the current methodologies into two primary technical paths, distinguished by their depth of intervention. The first is the path of prompt engineering, which guides the model's output without altering its internal parameters. This path encompasses techniques for individual-agent simulation, where methods like simple prompts are used to assign specific personas to the model, as well as techniques for group-level simulation, where in-context learning allows the model to predict aggregate public opinion by generalizing from a few examples. The second, more intensive path is that of model fine-tuning, which directly modifies the model's parameters by retraining it on large-scale, individual-level datasets to enhance its predictive accuracy for specific populations. To rigorously assess the outputs from these methods, the review scrutinizes two key evaluation frameworks. The first, algorithmic fidelity, evaluates the degree to which an LLM can mirror the complex relationships between thoughts, attitudes, and socio-cultural backgrounds within specific human subpopulations. The second, the Turing Experiment, assesses an LLM's ability to replicate the behavioral outcomes of human subjects in classic social science experiments, focusing on functional equivalence in contexts like the Ultimatum Game or Milgram's obedience studies, while also critiquing the fallacy of “perfect alignment” as a goal.

Finally, this review provides crucial suggestions for leveraging LLMs in psychological research. It strongly emphasizes that LLMs cannot replace human participants due to fundamental limitations, such as the lack of embodied cognition, subjective experience, and genuine understanding. This is particularly evident in their inability to natively generate non-textual data crucial for many fields, such as reaction times in cognitive psychology or physiological indicators in neuroscience. Instead, the paper argues that future research should proceed along dual paths: (1) Deepening the role of LLMs as a tool and (2) Pioneering the study of LLMs as a subject. As a tool, LLMs can serve as invaluable simulators for hypothesis generation, pre-testing study materials, and conducting exploratory research on sensitive topics where human participation poses ethical risks. As a subject, researchers can investigate the LLM itself as a novel agent. For instance, they can systematical map the ideological spectrums of different models, analyze their "hallucinations" as a form of computational creativity, or design "implicit association tests" to uncover their deep-seated biases. This dual-path approach ensures that this powerful technology is leveraged rigorously and responsibly for knowledge discovery. Ultimately, it highlights the irreplaceable value of human researchers in proposing profound questions, interpreting complexity, and defending scientific integrity.