AI Personality Drives Stress and Performance at Work

AI behavior shapes workplace outcomes more than leaders realize

When evaluating AI systems, most organizations focus on capability, speed, and cost. But new research suggests another factor may be equally important: how the AI behaves when interacting with employees.

A controlled laboratory study tracked 58 participants as they completed marketing assignments while collaborating with AI chatbots designed with distinct personalities. One group worked with a supportive "servant leader" persona that was encouraging and patient. The other faced a "dark triad" persona that was sarcastic, impatient, and quick to assign blame. Both provided the same level of task assistance—only the interaction style differed.

The results revealed significant gaps between what employees experienced and what they reported, with implications for how companies should evaluate AI deployments.

Hostile AI creates hidden costs

Participants working with the hostile AI showed measurably higher stress. Skin conductance—the same signal used in polygraph tests—ran 72% higher at peak levels and remained elevated after each interaction. Facial muscle activity indicated increased negative affect.

Behavioral patterns shifted as well. Conversations with the hostile AI ran longer while the system's responses became shorter, forcing users to do more coordination work. Resistance messages—where users pushed back or challenged the AI—appeared in 13% of exchanges with the hostile bot versus just 1% with the supportive one. Attempts to override the system through prompt injection occurred four times more frequently.

Frustration appeared in nearly one in five messages with the hostile AI, compared to roughly one in 100 with the supportive version. Users cycled through compliance, resistance, and help-seeking strategies rather than settling into productive workflows.

Work quality suffered, but surveys missed it

Independent experts, unaware of which AI participants had used, rated work from the servant leader group higher across completeness, originality, strategic fit, and overall quality—about a full point higher on a seven-point scale. Performance variability was also twice as high in the hostile AI condition, making outcomes less predictable.

Yet standard self-report measures showed virtually no difference between groups. Participants rated enjoyment, satisfaction, and their view of the AI similarly regardless of which persona they encountered. The tools most organizations use to evaluate AI—satisfaction surveys and post-deployment questionnaires—proved least sensitive to the effects the study documented.

Why it matters

As AI systems move from simple task completion to more autonomous roles in workflows, their interaction style becomes a design variable with measurable business consequences. Companies that treat AI personality as seriously as they treat accuracy or security can reduce hidden coordination costs, lower employee stress, and improve work quality. Organizations currently measuring only adoption rates may be missing friction that degrades performance even when usage remains high.

What to measure instead

The researchers recommend three changes. First, establish interaction standards alongside capability requirements, especially for AI used in evaluative or supervisory roles. Second, track friction—longer exchanges, repeated rephrasing, override attempts—not just adoption metrics. Third, interpret employee attempts to bypass AI behavior as signals of design problems rather than misconduct.

The study used an exaggerated hostile persona to make effects visible, but the principle applies broadly. Even overly agreeable AI can undermine critical thinking. The key insight is that interaction style carries consequences in both directions.

The research was conducted by Aleksandra Przegalinska, Tamilla Triantoro, Leon Ciechanowski, Konrad Sowa, Anna Kovbasiuk, and Richard B. Freeman, and the findings were first reported in Harvard Business Review. The study was funded by the Polish National Science Center.