Not yet human-reviewed
This source record was written automatically by the research pipeline and has not yet been reviewed by the maintainer or integrated into a topic page. Treat it as a provisional summary and verify claims against the original source.
From Tasks to Traits: Can LLMs Infer Personality From Brief Assessment Center Behavior?
Citation. Härtel, T. M. (2026). From Tasks to Traits: Can LLMs Infer Personality From Brief Assessment Center Behavior? International Journal of Selection and Assessment, 34(3). https://doi.org/10.1111/ijsa.70074
Summary. This peer-reviewed study evaluates whether an LLM can infer personality traits from transcripts of brief simulated online assessment-center group discussions. Across 363 participants, four Gemini 2.0 Flash runs per transcript produced reliable aggregate trait scores but only modest convergence with self-report trait tests. The authors position LLM ratings as a low-cost supplement to multi-source assessment rather than as a standalone selection measure because of weak discriminant validity, gender-related adverse impact, and input dependencies.
Key claims
- Aggregating four independent LLM runs yielded ICC(3,4) reliability of 0.75–0.92 across the evaluated traits, whereas single-run reliability ranged from 0.43 to 0.74.
- LLM trait inferences correlated modestly with trait tests, from r = 0.07 to 0.20, with conscientiousness not above chance and openness and intelligence at r = 0.20.
- The LLM’s trait-validity estimates were descriptively higher than trained human-rater benchmarks for five of seven traits, but the study does not establish suitability for standalone personnel-selection decisions.
Evidence & limitations
- The study reanalyzed recordings from 363 predominantly student, German-speaking participants in virtual assessment-center group discussions; the LLM pipeline used Whisper transcription, Pyannote speaker diarization, and Gemini 2.0 Flash trait inference.
- Validity was assessed against self-report trait measures and a human benchmark rated a 100-transcript subsample, so results are limited by brief transcript input, the simulated assessment-center context, modest absolute validity, discriminant-validity concerns, and reported gender-related adverse impact.