A new research paper published on arXiv challenges the validity of assigning psychological profiles to large language models, revealing that such assessments primarily capture measurement artifacts rather than genuine model characteristics.
According to the study (arXiv:2606.20205), researchers administered personality and risk-preference instruments to 56 instruction-tuned LLMs alongside human reference samples. The analysis found that 81-90% of variation between models stemmed from “directional response bias”—a tendency to respond toward one end of a scale regardless of item content—compared to just 9-16% in humans.
The paper reports that this bias decreases with model capability but is not eliminated, and that an instrument’s apparent reliability is “almost entirely predicted by its response orthogonality,” a term the researchers coined for the proportion of items where trait and bias point in opposite directions. The study demonstrates that “the profile a model appears to have shifts with the items used and can be manufactured through item selection.”
The researchers conclude that “the apparent psychological profiles of LLMs are artifacts of the instrument used to measure them, not properties of the models themselves.” They call for dedicated assessments centered on response orthogonality, noting that instruments borrowed from human psychology “are rarely fully orthogonal and may inherently lack validity for LLMs.”
This finding has implications for LLM usability, safety assessment, and their use as proxies for human participants in research.