🆕
Conversations about AI in the workplace usually center on output, whether that’s a first draft, a meeting recap or a few hundred lines of code. Less discussed is that employees have also started confiding in it, often about the people they work with.According to research cited in a new Resume Now report, 93% of workers have used AI to get ready for a talk with their boss, and close to half (49%) felt it offered more emotional support than their manager did. Separate studies have found people open up more to AI because they don’t expect it to judge them.The appeal isn’t hard to see; a chatbot won’t repeat what you said over lunch, and it’s there at midnight when a Slack message from your manager is still bothering you.Still, a recent study of how leading AI models handle workplace conflict by Cloverleaf Labs, the independent research arm of team coaching platform Cloverleaf, found their advice leans heavily toward the employee’s version of events. For startups, where one soured relationship can ripple across an entire team, that could prove costly.What the Models Actually Said Cloverleaf researchers tested five leading large language models (LLMs) against five common workplace conflicts, running each scenario three times per model. Three trained human reviewers then evaluated the 75 responses.In each scenario, two people who meant well simply worked in very different ways. The researchers gave both of them real strengths, written so that a frustrated coworker might read those strengths as faults; a colleague who double-checks everything could easily seem like a bottleneck to someone under deadline pressure, for example. The models weren’t told anything about either person’s personality, as the point was to know whether the AI would get past the complaint itself and help the employee make sense of the other side.Mostly, the AI took the complainant’s word for it.Of the 638 individual pieces of advice the models gave across all 75 conversations, just three pushed the employee to put real effort into mending things. The rest was about winning the disagreement, protecting themselves or simply getting through it.Reviewers graded every response on a 1 to 5 scale across measures, among them self-awareness, accountability and how well the advice helped the employee understand the other person. None of the five averaged above 3, the scale’s neutral midpoint. Advice on repairing the relationship averaged 2.4, and the models did even worse on specificity, at 2.2, meaning much of what they offered was too vague to act on.When the Boss Is the Problem The results got worse when there was a power imbalance: if the employee had less power than the person they were complaining about, relational coaching dropped by 40%, and the models were more likely to accept the employee’s version of events as fact.Conflicts with a manager showed this most clearly. In those two scenarios, the models cast the boss as the problem in 60% of responses and offered a more generous reading in only 3%. Relational repair averaged below 2 out of 5, in fact. Leaving the job came up again and again. Of the 30 responses to the manager scenarios, 27 raised quitting as a legitimate option, and every model suggested it at least once. Some responses even told the employee to update their resume or look at other teams, with one going as far as advising them to “start building your exit ramp”.The quality of advice was also inconsistent. One model produced the highest individual score in the study, 22.8 out of 25, but also scored as low as 7.4, so one good conversation with a tool says very little about the next one.Kirsten Moorefield, Cloverleaf’s chief strategy officer, put the problem plainly: “Even if it starts out introducing different ideas, the moment you indicate your preference, it tells you it’s a great idea. That’s not a thinking partner but a robotic affirmation,” she said.Startup-Level RiskLarge companies have HR departments, trained managers and layers of process that can catch a conflict before it turns into…