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

The Illusion of Personality: Why LLMs Don't Have Psychological Profiles

A new preprint from arXiv (2606.20205) delivers a sobering methodological critique: the psychological profiles we think we see in large language models are largely measurement artifacts. The researchers argue that applying human psychological instruments to LLMs produces stable-seeming results that are actually a function of the test design, prompt phrasing, and model architecture—not evidence of genuine personality traits.

What Happened

The study systematically examines how LLMs respond to established psychological questionnaires (e.g., Big Five, Myers-Briggs). The core finding is that LLM responses are highly sensitive to minor variations in question wording, ordering, and context—far more so than human respondents. When researchers controlled for these artifacts, the apparent "personality" of models like GPT-4 and Claude collapsed into near-random variation. The "stable profiles" reported in prior work appear to be an artifact of averaging across many responses, not evidence of a coherent internal disposition.

Why It Matters

This has significant implications for three critical use cases:

• Safety assessment: If we cannot reliably measure an LLM's "personality," we cannot trust evaluations that claim a model is "more agreeable" or "less neurotic." Such labels might mislead safety teams into believing they understand model behavior when they do not.

• Human proxy research: Using LLMs as stand-ins for human participants in psychology studies is increasingly popular. This paper suggests such proxies are fundamentally unreliable—LLMs do not possess the stable, context-independent traits that make human personality measures meaningful.

• Model alignment and control: The assumption that an LLM has a "default" personality that can be shaped through fine-tuning is undermined. If profiles are artifacts, then alignment efforts targeting "personality" may be chasing a phantom.

Implications for AI Practitioners

For those building or deploying LLM-based systems, the takeaway is practical: stop treating model outputs as expressions of a stable character. Instead, treat every interaction as a fresh inference conditioned on the exact prompt, context, and system instructions. This means:

• Evaluation protocols must control for prompt sensitivity explicitly—run multiple paraphrases and orderings before concluding anything about model tendencies.
• User-facing applications should not assume consistency of tone or behavior across sessions without explicit conditioning (e.g., system prompts that define a role).
• Research claims about LLM personality should be treated with extreme skepticism unless they demonstrate robustness to the artifacts identified in this paper.

The field has been seduced by the anthropomorphic ease of saying "this model is conscientious." This research reminds us that LLMs are not people—they are statistical machines that can simulate personality traits when prompted to, but possess no underlying self to measure.

Key Takeaways

• Psychological profiles assigned to LLMs are largely measurement artifacts, not stable traits
• Prompt sensitivity makes LLM personality assessments unreliable for safety or research use
• Practitioners should treat each LLM interaction as context-dependent, not as an expression of a consistent character
• Claims about LLM personality in published work require re-evaluation with proper controls for artifact effects