In the Blink of an AI: Large Language Models Can Infer Traits From LinkedIn

Abstract

Large language models (LLMs) are increasingly used in HR tasks, including leveraging LinkedIn profiles to infer personality during hiring decisions. Building on the lens model, we explore LLMs' capability to take over the role of human recruiters when drawing LinkedIn-based trait inferences. For this purpose, we fed 406 LinkedIn profiles twice to Microsoft Copilot (powered by GPT-4) and single-shot prompted it to evaluate users’ traits in terms of personality (Big Five, narcissism) and intelligence. The LLM-driven inferences demonstrated good test-retest reliability for selected traits (.31 < r < .81) and moderate correlations with ground-truth trait test scores, with intelligence (r = .24), openness (r = .20), and extraversion (r = .20) most accurately assessed. Based on a set of 32 coded cues of basic LinkedIn information, we explain this accuracy by Copilot using LinkedIn information consistently across users and showing reasonable sensitivity for valid information. However, we also identified general and cue-specific positivity bias, range restriction, and adverse impact leading to less favorable and less accurate inferences for gender and age groups. This study extends the lens model to LLMs as perceivers, providing insights into their potential and limitations for assessment in recruitment and broader HR contexts.

Ähnliche Arbeiten