Apparent Psychological Profiles of Large Language Models are Largely a Measurement Artifact

Technical Summary: Apparent Psychological Profiles of Large Language Models are Largely a Measurement Artifact

Problem Statement
Large language models (LLMs) are increasingly evaluated using psychological instruments designed for humans (e.g., personality inventories, risk-preference tasks) to assign them stable behavioral profiles. These profiles influence decisions regarding model usability, safety assessments, and the use of LLMs as proxies for human participants in research. However, recent findings suggest that LLM responses are unstable under minor prompt perturbations and that factor structures derived from them diverge from human norms. The central problem addressed is whether the apparent psychological profiles of LLMs reflect genuine, stable behavioral dispositions (traits) or if they are artifacts of the measurement process itself, specifically driven by systematic response biases rather than the constructs the instruments intend to measure.

Methodology
The authors employ a formal psychometric framework to disentangle two components of systematic responding: a latent trait ($\theta$) and a directional response bias ($b$). The study utilizes a measurement model where a respondent's answer to an item is a function of the scale midpoint, the latent trait, a directional bias, and noise.

• Data: The study analyzes 56 instruction-tuned LLMs (46 open-source, 10 proprietary) alongside a large human reference sample ($N=20,993$).
• Instruments: Two comprehensive batteries were administered:
  1. IPIP-NEO-300: A 300-item Big Five personality inventory (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism).
  2. Risk-Preference Battery: 24 instruments spanning self-report scales and behavioral tasks (e.g., decisions from description, lotteries) based on Frey et al.
• Key Metric: The study introduces and utilizes response orthogonality, defined as the proportion of items where the trait and response bias act in opposite directions (e.g., forward-keyed vs. reverse-keyed items).
• Analysis:
  • Diagnostic Test: The authors analyze the correlation between mean responses to forward-keyed and reverse-keyed items. Under the measurement model, a negative correlation indicates trait dominance, while a positive correlation indicates response bias dominance.
  • Decomposition: For each model, responses are decomposed into estimated trait ($\hat{\theta}$) and bias ($\hat{b}$) components.
  • Variance Decomposition: The proportion of between-model variance attributed to bias versus trait is calculated.
  • Orthogonality Analysis: The relationship between the proportion of reverse-keyed items in an instrument and the apparent internal consistency (Cronbach's $\alpha$) is examined.

Key Findings

1. Bias Dominates Between-Model Variation:

   • In human samples, the correlation between forward and reverse item means is strongly negative (e.g., $\rho \approx -0.82$ for Neuroticism), indicating that variation is driven by the underlying trait.
   • In LLMs, this correlation is strongly positive (e.g., $\rho \approx +0.68$ for Neuroticism), indicating that variation is driven by a directional response bias (a tendency to select one end of the scale or a specific label regardless of content).
   • Variance decomposition reveals that 81–90% of between-model variation in LLMs is attributable to response bias, compared to only 9–16% in humans.

2. Capability Reduces but Does Not Eliminate Bias:

   • Response bias decreases slightly as model capability (parameter count) increases and is lower in proprietary models compared to open-weight models.
   • However, even the most capable models exhibit bias levels significantly higher than the average human. The bias is not eliminated by scaling; the most capable models remain outside the human range of bias.

3. Apparent Reliability is an Artifact of Orthogonality:

   • In LLMs, apparent internal consistency (reliability) is almost entirely predicted by an instrument's response orthogonality.
   • There is a near-deterministic negative correlation ($r = -0.95$) between the proportion of reverse-keyed items and mean inter-item correlation.
   • Instruments with few or no reverse-keyed items show high reliability (Cronbach's $\alpha$ up to 0.96), while fully orthogonal instruments show near-zero or negative reliability. This suggests that high reliability in non-orthogonal instruments reflects consistent bias, not consistent trait expression.

4. Profile Instability and Manufacturability:

   • Because bias drives responding, a model's psychological profile shifts drastically depending on whether forward- or reverse-keyed items are used.
   • Profiles can be "manufactured" through item selection. The difference between profiles derived from forward-only versus reverse-only items is substantial for LLMs (mean absolute bias $|\mu_b|$ ranging from 0.15 to 0.99) but minimal for humans ($|\mu_b| \approx 0.09$).
   • This instability is more pronounced in risk-preference instruments than in personality inventories.

Significance and Claims
The paper argues that the apparent psychological profiles of LLMs are artifacts of the instruments used to measure them, not intrinsic properties of the models. The authors claim that:

• Current Assessments are Flawed: Profiles derived from standard human psychological instruments are largely confounded by response bias because these instruments rarely possess full response orthogonality. Consequently, reported reliability coefficients and personality scores for LLMs do not reflect stable dispositions.
• Methodological Correction is Required: Valid assessment of LLM dispositions requires either the use of fully orthogonal instruments (balanced forward/reverse keying) or the application of formal psychometric frameworks that explicitly model and separate response bias from trait.
• Implications for Research: These findings challenge the validity of using LLMs as proxies for human participants in research without correcting for these measurement artifacts. They also suggest that behavioral tasks are not immune to these issues; in fact, they may be more sensitive to bias than self-reports.
• Nature of LLM Behavior: The study posits that the observed bias is a distinct phenomenon, not easily explained by post-training effects like sycophancy (which would be one-sided) or social desirability (which would manifest as trait variance). It suggests that LLMs may lack human-like stable dispositions or that current instruments are invalid for them due to a lack of embodiment or volition.

The authors conclude that simply borrowing human psychology paradigms is insufficient for characterizing LLMs and call for dedicated assessments centered on response orthogonality to distinguish genuine behavioral dispositions from measurement artifacts.