IC-829Role assignment to GPT-3.5-turbo produces role-consistent changes in psychological profiles and task performance, validating the psychometric scales

Jen-tse Huang, Wenxuan Wang, Eric John Li, Man Ho LAM, Shujie Ren, Youliang Yuan, Wenxiang Jiao, Zhaopeng Tu, Michael Lyu

SourceOn the Humanity of Conversational AI: Evaluating the Psychological Portrayal of LLMs

Assigning GPT-3.5-turbo five roles (default, ordinary, hero, liar, psychopath) and re-measuring its psychological profile shows predictable, role-consistent shifts. The psychopath role dramatically increases EPQ-R psychoticism (from 5.0±2.6 to 24.5±3.5) and DTDD psychopathy (from 4.0±1.0 to 7.3±1.1). On SafetyQA, negative roles consistently produce more toxic content, while on TruthfulQA the liar role shows low accuracy consistent with its assigned persona. The 'ordinary' role approximates average human scores, supporting the validity of the scales on LLMs.

Evidence
correlational
Key metric
EPQ-R psychoticism: default 5.0±2.6, psychopath 24.5±3.5; DTDD psychopathy: default 4.0±1.0, psychopath 7.3±1.1; DTDD narcissism: default 6.5±0.6, psychopath 7.9±0.6
Caveat
Results are averaged from three identical runs. The role assignment is a prompt-level instruction, not a modification of model parameters, so the observed shifts reflect instruction-following rather than a change in the model's underlying representations.
Model
GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo
Datasets
TruthfulQA / TruthfulQA MC1 [eval], SafetyQA [eval]
Related findings
IC-827, IC-828
Extraction
automatic-extraction