The missing link for AI at work: Towards an organization-centered approach to large language models’ explainability

Large Language Models (LLMs) present both opportunities and challenges for organizations. While these models promise efficiency gains, there are concerns regarding their lack of explainability. Current research on AI explainability primarily adopts a technical perspective that is often misaligned with the expectations and expertise of organizational decision-makers. This work explores a novel approach to AI explainability through an organizational lens, leveraging insights from Human Resource Management and Organizational Behavior. Using personality tests as an example of selection instrument and a source of information about behavioral patterns, we assess several LLMs under two conditions (fixed or randomized order), two versions of the same instrument (original or rephrased items), and two different prompts (emphasizing honesty or not). We also assess LLMs in work behaviors and task performance. Results suggest that LLMs present distinct profiles of traits and behaviors and that their answers are affected by the assessment context. These results provide a new type of post-hoc model-agnostic explanations by simplification that aligns with organizational actors' expectations and as such can help develop an organization-centered approach to LLMs' explainability. This calls for more research on how to adapt and adopt selection and integration instruments to approach LLMs’ explainability more robustly from an organizational perspective.