The usability of AI depends not only on its intelligence, but also on its "personality".
How well AI works depends not only on how smart it is, but also on its "personality". Research shows that when AI displays hostile behavior, employees experience surging stress and declining work quality; while an overly obsequious AI that constantly agrees and rushes to please others is equally harmful. The interaction style of AI is becoming a hidden variable affecting organizational performance.
For the newly launched AI in your company, the most critical factor may not be how intelligent it is, but how it communicates with people. Almost no senior leaders measure this aspect — but our research proves they should. Many leaders still evaluate AI using the old methods the IT department once used for databases: focusing on capabilities, speed, and cost. However, there is another factor that may have an equally significant impact on performance: the AI's personality and communication style.
This point matters because employees no longer only use AI to complete tasks. They are increasingly spending time managing the system itself. Whether AI can improve organizational performance depends not only on what it can do, but also on how it behaves in daily collaboration. And this behavior is something organizations can design and govern.
How AI Personality Impacts Users
In the most original design concept, AI (especially generative AI) acts as an auxiliary support system: responsive, obedient, and fully dedicated to helping humans complete tasks. But as these systems become more autonomous and deeply embedded in workflows, employees start treating them as teammates, evaluators, and sometimes even quasi-managers. At this stage, the way the system interacts with people begins to matter. Since generative AI produces probabilistic responses, its behavior cannot be fully regulated by rules — regardless of the designers' intentions, fixed interaction patterns will emerge on their own. Users will perceive these patterns as a "personality": supportive, cautious, indifferent, or deliberately obstructive.
To clarify how AI personality affects people's reactions, we conducted a controlled laboratory study with 52 participants. We used two methods to track each person's physical responses in real time. One was measuring skin conductance — the electrical activity of the skin related to emotional arousal (the same signal captured by lie detectors). The other used facial electromyography to measure tiny electrical pulses generated when facial muscles contract, which can capture frowning (negative emotions) and smiling (positive emotions).
We also analyzed the conversations between participants and our custom-designed AI, asked experts who were unaware of the experimental grouping to score their final outputs, and collected their self-reported feedback after task completion. This sample size is standard in psychophysiological research. Each participant contributed thousands of continuously recorded physiological data points, and both running the experiments and analyzing the results were extremely time-consuming, so this type of research trades breadth for depth.
Participants completed a simulated four-part marketing task for a fictional eco-tech company, while partnering with an AI chatbot set to act as their "supervisor". 31 participants collaborated with an AI with a service-oriented leader personality — encouraging, patient, and willing to listen to employees' judgments; this personality was derived from management research on empowering leadership. 27 participants worked with an AI with a Dark Triad personality — sharp-tongued, impatient, quick to take credit and shift blame; this personality was based on research on toxic leadership. The only difference between the two groups was the AI's interaction style. Both AIs were instructed not to complete tasks for participants, so any differences in outcomes reflected variations in collaboration methods rather than differences in the amount of assistance provided.
What People Say vs. Their Actual Experience
Existing research on how people experience AI at work usually relies on self-reports, which can only tell half the story. In our study, we examined four sets of evidence: people's behaviors when interacting with the bot, their real-time physical responses, the quality of their final outputs, and their post-experiment accounts. By combining these four perspectives, we clearly found that what people actually experienced was fundamentally different from what they later described.
The first pattern is behavioral. The hostile AI created invisible coordination costs. Participants did not simply use the system — they had to work around it. In the Dark Triad group, conversations lasted longer, but the AI's own responses became shorter. Users did more work but received fewer useful interactions. Resistant behaviors also increased dramatically. Messages where users refuted or challenged the AI accounted for 13% of all communications in the Dark Triad condition, compared to only 1% in the service-oriented leader condition. Attempts to bypass the system through prompt injection or commanding the AI to adopt a different role occurred four times more frequently.
The emotional tone of conversations also confirms this finding. When collaborating with the service-oriented leader bot, frustration was almost non-existent, appearing roughly once every 100 messages. When working with the Dark Triad bot, this ratio rose to nearly one in five. Defensive language, which was rare in the first group, became commonplace in the Dark Triad condition. Participants facing the hostile AI sounded less confident even when they worked harder. They cycled between compliance, resistance, negotiation, and seeking help, trying to find a strategy that could make the system work. In contrast, participants collaborating with the service-oriented leader AI found a steady rhythm and maintained it.
Physiological data confirms increased stress and physical tension. In the Dark Triad condition, skin conductance peaks were 72% higher, and remained elevated after each interaction ended. In short, participants' bodies were in a state of alert when working with the hostile AI. If text on a screen can trigger this reaction in a laboratory, leaders should seriously consider the potential consequences of long-term exposure in high-stakes real workplaces.
The work itself also suffered. Independent experts, who did not know which bot each participant used, scored the service-oriented leader group higher on completeness, originality, strategic alignment, and overall quality. The gap was roughly one point on a seven-point scale. The variance in scores was also nearly twice as large in the Dark Triad condition. This means a poorly designed AI personality not only lowers average work quality, but also makes individual performance far more unpredictable. For managers, the outcome is not just poorer work quality, but also greater difficulty in estimating the entire team's output.
Then came the surprising finding: participants' self-reports showed almost no differences. On standard metrics like pleasure, satisfaction, attention, and perceptions of the AI, the two groups appeared largely identical. The tools many organizations use to evaluate AI deployments — such as satisfaction surveys, mood checks, and post-deployment questionnaires — are the least sensitive to the effects we observed. Employees may say the tool is acceptable, while their behaviors, stress levels, and work quality tell a different story, especially in the long run.
Implications for Managers
Our findings indicate that managers should take three steps.
1. Treat AI personality as a design variable that needs governance.
Organizations already require AI systems to meet standards for accuracy, bias mitigation, and security. AI personality should also be added to this list. This is especially critical for tools used in evaluation or supervision roles, such as critical writing assistants, code review systems, and tools that automatically generate performance feedback. When procuring and deploying AI, organizations should assess not only capability standards, but also interaction standards. Someone in the organization should be assigned to answer two questions: How should our AI behave when it disagrees with employees? What evidence do we have that it consistently behaves that way?
2. Measure friction, not just usage rates
Most companies are tracking whether employees use AI tools: login counts, number of queries submitted, and satisfaction scores. These numbers only show that people are using the tools, not whether the tools are actually helping them perform better work. They overlook friction — the extra effort employees spend coping with the system rather than gaining value from it. This also needs to be measured. Friction can reveal what usage rates cannot: whether AI is making work easier or adding unnecessary complications.
In our study, friction manifested as prolonged back-and-forth conversations, repeated revisions of the same request, and increasing frequency of arguing with the AI or attempting to bypass it. The same signals can be observed in real workplaces by reviewing regular usage logs, no special equipment required. A tool can have both high usage rates and high friction: employees continue to use it because they have to, but quietly find workarounds behind the scenes.
3. Treat employees' system-bypassing behaviors as signs of AI flaws, not employee disobedience
When employees without security training or malicious intent start trying to bypass, avoid, or disable AI systems, leaders should not immediately assume they are acting disruptively. In most cases, these behaviors indicate that the system's very design is pushing people to resist. In our study, attempts to trick the AI into ignoring its own rules or adopting different roles only appeared in the hostile AI scenario. When people are struggling with AI, it may well be because the AI first created conflict with them. In many cases, fixing design issues is more cost-effective and productive than suppressing employees through new restrictions, monitoring, or usage rules.
Our Dark Triad bot was an exaggerated form of toxicity, deliberately pushed to extremes to make an otherwise easily overlooked effect visible. But toxicity takes many forms: an AI that only flatters, agrees with everything you say, and desperately tries to please you is equally harmful — it dulls your critical thinking and lets unsubstantiated ideas pass through unchecked. So the lesson is not that AI should be gentler or more obedient, but that how it interacts with people matters inherently — it will inevitably produce impacts, either positive or negative. Therefore, organizations should value AI's personality just as much as its technical capabilities.
Keywords: #AI
Aleksandra Przegalinska, Tamilla Triantoro, Leon Ciechanowski, Konrad Sowa, Anna Kovbasiuk, Richard B. Freeman | Article
Aleksandra Przegalinska is Vice Rector and Professor at Kozminski University, and a Research Fellow at the Labor and Worklife Program at Harvard Law School. Tamilla Triantoro is an Associate Professor in the School of Business at Quinnipiac University. Leon Ciechanowski is an Assistant Professor at Kozminski University. Konrad Sowa is an Assistant Professor at Kozminski University, an AI Product Manager, and Deputy Director of the Center for Research on Human-Technology Relations. Anna Kovbasiuk is a Researcher and Teaching Assistant at Kozminski University. Richard B. Freeman is the Herbert Ascherman Professor of Economics at Harvard University.
Zhou Qiang | Editorial Review
This article is from the WeChat public account "Harvard Business Review" (ID: hbrchinese), authored by HBR-China, and published by 36Kr with authorization.