The (non-)alienness-dilemma in the use of simulated research subjects in psychological research

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Technological advances in machine learning, natural language processing, and generative artificial intelligence have created novel opportunities for integrating artificial intelligence into research. The resemblance between large language model (LLM) outputs and human cognition and behavior has sparked discussion about their potential use in simulated psychological trials. In this article, I explore the implications of simulating research subjects in psychological research, demonstrating fundamental challenges arising from both the assumption of alienness and non-alienness. Epistemic challenges are grounded in a lack of construct validity and measurement invariance of psychological measurements in LLM populations, constraining researchers’ ability to draw inferences about human psychological constructs from LLM outputs. This explanatory gap arises from LLMs being disembodied and lacking phenomenological consciousness. Increasing human-likeness in LLMs may reduce epistemic limitations but this may raise moral concerns – if LLMs acquire capacities for subjective experience or consciousness, they must fall under guidelines for ethical research. Ultimately, a structural dilemma emerges: the benefits of simulating research subjects rely on differences in moral status between humans/animals and LLMs, yet these differences simultaneously limit the insights obtainable from such simulations. Careful interpretation of results and consideration of ethical guidelines are therefore essential for research involving increasingly human-like LLMs.