When an employee turns in a polished marketing post or report drafted with ChatGPT, a manager faces an unfamiliar judgment call. The work may arrive faster and read smoothly, yet knowing a machine produced it can leave a boss wondering how much effort went into the assignment. New research published in The Leadership Quarterly investigated how managers respond when they learn a subordinate used a large language model, or LLM, to complete a task.
Across six studies involving corporate managers and university professors, the researchers found that using an LLM led supervisors to trust employees substantially less than when employees completed the work on their own. That drop in trust persisted even when supervisors used AI tools themselves and regardless of how important the task was. Only one factor eliminated the statistically significant penalty: an explicit organizational policy encouraging AI use.
Why managers face conflicting signals over AI
Most previous research on workplace AI has asked whether people trust the technology itself. Samuel Meyers of the Hebrew University of Jerusalem and his colleagues shifted the focus to a different question: how using AI affects a supervisor’s trust in the human employee.
Existing leadership theories do not offer a clear prediction for this situation. Social exchange theory, which views workplace trust as a system of mutual reciprocity, could cut either way. A boss might see AI-assisted output as a helpful contribution that boosts team productivity, or as a sign that the worker is shirking effort.
Similarly, managers hold mental models of what a good employee looks like, valuing traits such as competence, initiative, and hard work. While some supervisors might view prompting an AI tool as a modern technical skill, psychology research suggests people often see AI-generated work as inauthentic and extend that skepticism to the person who submitted it. To see which reaction wins out, Meyers and his colleagues designed a series of pre-registered experiments and an academic field study.
Testing trust with real money on the line
In their first four experiments, the researchers used an economic decision game called the “rely-or-verify” game to measure behavioral trust. Rather than asking managers hypothetical questions, the game placed real bonus money at stake.
Before running the main experiments, the team recruited pilot participants to act as subordinates and write a real social media post promoting a university research platform. The workers chose how to produce the post and then reported how many minutes they spent on it. The researchers ran those posts through Originality.ai, a commercial detection tool, to identify submissions written entirely by the worker, generated by an LLM, or plagiarized from another person.
Next, the researchers matched those real subordinates with temporary supervisors recruited online from the United States, the United Kingdom, and Canada. The supervisors never saw the actual social media post, removing any influence of writing style or tone. They were told only how the detection software classified the post, and they then had to decide whether to rely on the subordinate’s reported work time or pay to verify it.
The financial incentives mirrored a common management dilemma. If the employee reported their time honestly and the supervisor chose to rely on them, both earned a larger payout, with $1.50 going to the supervisor and $0.75 to the worker. Checking the report carried a fee that lowered the supervisor’s bonus to $1.00, but relying on an employee who exaggerated their time left the supervisor with $0 while the employee pocketed up to $1.50. Choosing to rely rather than verify meant accepting real financial vulnerability.
In Study 1, 392 full-time managers with an average of 10.5 years of supervisory experience took part in the game. When managers learned a subordinate wrote the post themselves, they had a 68% predicted probability of relying on the worker’s time report. When they learned an LLM generated the post, that probability fell to 24%.
Even so, managers did not treat AI use as harshly as copying from another person. Subordinates whose posts failed the plagiarism check had just an 11% predicted probability of being relied upon. The drop in trust for AI users also held up on standard survey measures of trust and remained even after accounting for ratings of the employee’s competence and professionalism.
When workplace context changes the reaction
Having established a baseline trust penalty, the researchers tested three workplace factors that might alter a manager’s response: company AI policies, task stakes, and the supervisor’s own experience with AI.
In Study 2, 601 managers played the same game under one of three conditions. Before evaluating the subordinate, managers were told that the organization either banned LLM use for the task, encouraged LLM use, or had no policy in place.
When no policy existed, managers again trusted AI users far less, with reliance dropping from 69% for self-written posts to 33% for LLM-generated ones. When LLM use was banned, the gap grew wider, falling from 71% to 19%. When company policy explicitly encouraged LLM use, however, the trust penalty shrank to a difference that was not statistically significant, with managers relying on 47% of AI-using subordinates compared to 58% of those who wrote the post alone.
Task importance did not change the pattern. In Study 3, 398 managers evaluated a subordinate on a social media post described as either high-stakes and central to attracting subscribers, or low-stakes and minor. Managers trusted AI users less in both cases, with reliance falling from 77% to 33% on lower-stakes tasks and from 61% to 30% on higher-stakes ones.
Nor did a manager’s personal familiarity with AI soften the penalty. Pooling survey data from 1,249 managers across the first three studies showed that supervisors who regularly used LLMs penalized AI-using subordinates at roughly the same rate as supervisors with little AI experience.
To test whether hands-on exposure directly changes attitudes, Study 4 recruited 398 adults who had never used an LLM before, including 233 with supervisory experience. Half watched a brief tutorial and spent five minutes using ChatGPT, while the rest learned and played the computer game Minesweeper. Brief experience with ChatGPT made no meaningful difference: reliance dropped from 77% to 28% among participants with no AI experience, and from 71% to 28% among those who had just tried the chatbot.
The ripple effects on careers and academic work
A loss of trust does not stop at double-checking timesheets. In Study 1, lower trust in AI-using employees was associated with lower ratings of their job performance, voluntary helpfulness beyond their formal duties (known in management research as organizational citizenship behavior), and readiness for promotion.
Because those ratings were collected alongside trust in Study 1, the researchers ran Study 5 with 204 managers to test whether lower trust directly causes harsher career evaluations. Managers read a real reference note written by an earlier supervisor who had either relied on or verified a subordinate, without any mention of AI. To give the evaluations real weight, managers were told that higher ratings increased their likelihood of being paired with that subordinate in a future paid collaboration. Managers who read that a subordinate was trusted rated the worker substantially higher on expected job performance, voluntary workplace contributions, and promotability than those who read the subordinate was not trusted.
Finally, Study 6 examined whether the pattern holds outside simulated business games and whether the amount of AI used matters. The researchers recruited 321 professors who currently supervise research assistants, graduate students, or postdoctoral fellows. Each professor read the real introduction section of one of 14 papers written by graduate students for the 2024 Academy of Management Conference.
Half of the papers registered 0% AI-generated text on Turnitin, a detection tool widely used in universities, while seven matched papers of similar length, topic, and research method registered Turnitin scores between 38% and 100%. Professors saw the paper’s introduction along with its Turnitin score and rated both its writing quality and how much they trusted the author. Even after accounting for perceived writing quality and topic familiarity, higher detected AI usage was linked to lower trust in the author. Unlike the corporate managers in earlier studies, professors with more personal LLM experience penalized AI use slightly less heavily than less experienced peers, though their trust still declined significantly.
What the findings mean for workplace policies and their limits
For organizations adopting generative AI, the results highlight a hidden friction in unwritten norms. In Study 2, having no formal AI policy still produced a substantial trust drop, though a smaller one than under an outright ban. According to the findings, employers that expect workers to take advantage of AI tools without suffering informal penalties from their managers may need clear policies that explicitly encourage AI use.
At the same time, several features of the research limit how broadly the results apply. In the first four studies, supervisors never saw the subordinate’s actual writing, leaving open whether noticeably higher-quality AI output could offset skepticism in day-to-day business settings. The rely-or-verify game also measures willingness to accept financial risk, which can blend trust with individual risk preferences, though survey trust measures yielded the same patterns.
Other caveats relate to the samples and tools used. Because few managers on the recruitment platform had zero prior AI experience, Study 4 included non-managers and gave participants only five minutes of ChatGPT practice, which may not replicate months of workplace adoption. In Study 6, professors evaluated conference authors rather than their own direct students, and as the authors noted in their participant debrief, AI detectors like Turnitin can produce false positives and false negatives, meaning the study captured reactions to perceived AI use rather than verified authorship.


