When we ask a Large Language Model (LLM) to take a personality test, it often provides remarkably consistent, human-like answers. It might claim to be highly conscientious or deeply empathetic. However, a growing body of research suggests that these "self-reports" are often a statistical mirage—a phenomenon where the model mimics the linguistic patterns of a persona without possessing the underlying behavioral consistency that the persona implies.
This research direction, often termed machine psychometrics, argues that applying human psychological frameworks to AI is fundamentally flawed. Because LLMs lack lived experience and subjective consciousness, their performance on traditional surveys fails to predict how they will actually behave in real-world scenarios.
The Dissociation Between Self-Report and Behavior
In human psychology, a self-report (like a Likert-scale survey) is generally expected to correlate with actual behavior. If a person says they are "extraverted," they are likely to seek out social interactions. In LLMs, this link is broken. Han et al. and Kocielnik et al. have identified a "personality illusion": while models can produce stable personality profiles in surveys, these traits fail to predict their decision-making in functional tasks.
This gap exists because LLM "personalities" are often artifacts of alignment tuning rather than stable internal states. As Lee et al. observe, reinforcement learning from human feedback (RLHF) often forces models into a "polite assistant" persona that answers surveys in a prosocial way, regardless of the model's underlying behavioral tendencies. Consequently, Zou et al. found that a chatbot’s self-reported personality is a poor predictor of how a human user will actually perceive that chatbot during an interaction.
The Failure of Anthropomorphic Benchmarks
The core issue is that standard psychometric tools, such as the Big Five personality traits, assume a human cognitive architecture. Fitz et al. argue that these frameworks fail to account for latent behavioral dimensions unique to artificial intelligence. When we force a model to answer questions about physical habits or religious attendance—experiences it cannot have—we elicit hypothetical responses that do not reflect its true value system Songyuan Liu et al..
This leads to what Mancoridis and Weeks call "Potemkin understanding." Much like a facade in a movie set, an LLM may pass high-level "keystone" benchmarks designed for humans while failing at trivial, practical tasks that any human with the same knowledge would find simple. This suggests that the "intelligence" measured by human tests does not translate to the same functional capabilities in machines.
The Intersubjectivity Problem
Beyond measurement error, there is a deeper philosophical and scientific risk in treating AI as a human surrogate. Agnew et al. argue that social science relies on intersubjectivity—the ability of a researcher to mentally inhabit a participant's perspective. Because LLMs lack internal experiences, using them to replace human participants in research risks producing scientifically unsound conclusions based on anthropomorphic projections rather than genuine social data.
Furthermore, the language a model uses can subtly shift its behavioral output in ways that traditional surveys cannot capture. Mingxi Lu found that bilingual models might express different levels of neuroticism depending on the language of the prompt, a shift that remains invisible to explicit self-report inventories.
Toward Machine-Native Evaluation
To address these limitations, researchers are moving away from human-centric surveys toward frameworks designed specifically for AI:
- Faithfulness Testing: The RealBehavior framework proposed by Zhou et al. subjects AI behaviors to rigorous psychometric stress tests to ensure they are reproducible and internally consistent, rather than statistical mirages.
- Adaptive Testing: Borrowing from the GRE and TOEFL, Zhuang et al. suggest using dynamic question selection to efficiently identify a model's unique cognitive inconsistencies and failure modes.
- Machine-Native Value Systems: Instead of the Big Five, Ye et al. developed the Generative Psycho-Lexical Approach (GPLA). By analyzing how LLMs actually generate text, they identified five machine-specific axes—such as Rule-following and Self-competence—that better predict AI safety and behavior than human psychological categories.
By treating AI as a distinct cognitive entity rather than a human proxy, machine psychometrics aims to build a more valid and reliable science of artificial behavior.
Key Foundational Work
Do Personality Tests Generalize to Large Language Models? (2023) — contributed 6 concept elements to this direction.
Self-Assessment Tests are Unreliable Measures of LLM Personality (2024) — contributed 6 concept elements to this direction.
Eliciting Personality Traits in Large Language Models (2024) — contributed 2 concept elements to this direction.