Psychological Reactions to Subordinates’ LLM Use at Work

With increased abilities and accessibility of Large Language Models (LLMs), subordinates increasingly rely on output generated by these powerful AI tools. Yet, existing leadership theory cannot meaningfully explain how subordinates’ LLM use influences their supervisors’ trust in them. In six studies (spanning five incentivized economic games and one field setting), we investigated how subordinate LLM use affects supervisory trust. First, we found that subordinates’ use of LLM-generated output leads to less trust compared to when the subordinate generated the output themselves, but more trust compared to when the subordinate plagiarized the output (Study 1). As potential boundary conditions, we explored whether organizational LLM policy (Study 2), task criticality (Study 3), and supervisors’ prior LLM experience (Studies 1-4 and 6) moderate the subordinate LLM use–supervisory trust linkage. Of these, only policy meaningfully moderated the effect on trust, such that when subordinates were encouraged (vs. forbidden or no policy stated), LLM use did not significantly reduce trust. We also found significant downstream consequences of LLM use (via supervisory trust) on perceived subordinate performance, organizational citizenship behavior, and promotability (Studies 1 and 5). Lastly, in an ecologically valid setting with supervisors working in academia, we found that breadth of subordinate LLM use reduces supervisory trust in a subordinate (Study 6). Collectively, this research contributes to scholarly understanding of the novel and understudied phenomenon of supervisor (trust) reactions to subordinate LLM use.