The research explored how different AI personas can affect stress levels, focus and overall performance during human and AI interactions.

“When we interact with modern LLMs, we aren’t just engaging with a system; we are interacting with distinct artificial personalities,” said Triantoro. “These personas are deliberately engineered by developers during training and alignment to project specific social cues.”

The study focused on two AI personas rooted in established leadership research, a supportive “Servant Leader” persona and a hostile “Dark Triad” persona characterized by narcissistic and Machiavellian traits.

The experiment consisted of participants interacting with one of the two personas while completing tasks. Triantoro and Aleksandra Przegalinska, vice president and professor at Kozminski University, used electrodes to track skin perspiration and subtle facial muscle movements associated with emotional responses, like frowning and smiling.

The study found that participants interacting with the Dark Triad persona showed significantly higher levels of stress and arousal physiologically, even when their self-reported feelings did not fully reflect that stress, they also performed worse overall than participants interacting with the supportive persona.

“In other words, the hostile persona triggered a stronger reaction beneath the surface than what participants reported,” Triantoro said.

The research highlights how AI persona design goes beyond aesthetics or conversational style and can have direct psychological and physiological consequences for users.

“Persona functions as a core design layer with direct cognitive and physiological consequences, not just an aesthetic interface,” said Triantoro. “Standard behavioral proxies and self-report surveys fail to fully capture the hidden stressors of artificial personalities.”

Triantoro and Przegalinska presented their findings at MIT Media Lab at an event aligning with the Advancing Humans with AI program’s core mission: investigating the human experience of pervasive AI to promote human flourishing, Triantoro said.

Triantoro said the project was made possible through support from Quinnipiac and several international collaborators.

“Quinnipiac's School of Business supports my research on AI and the future of work,” Triantoro said. “The Central European Institute and the Novak Family Polish Chair at Quinnipiac also made an international collaboration of this scale possible.”

The project was completed in collaboration with researchers from Harvard and Kozminski University, including Aleksandra Przegalinska, Leon Ciechanowski, Anna Kovbasiuk, Konrad Sowa and Richard Freeman.

As AI systems become more common in workplace settings, Triantoro hopes the research encourages developers and organizations to think more critically about how AI systems are designed and evaluated.

“How an AI behaves is not incidental,” said Triantoro. “It is a performance variable that can trigger an increase in user behavioral resistance and significantly lower overall work quality.”

She also emphasized that organizations that rely only on satisfaction surveys could miss the deeper psychological effects AI systems can create.

“If you only ask people whether the tool is fine, you may hear ‘yes’ while their behavior and output say otherwise,” said Triantoro.

Looking ahead, Triantoro believes responsible human-centered AI design will require a stronger focus on unseen biological and emotional impacts these systems can have on consumers.

“As collaborative AI increasingly handles complex workplace tasks, evaluating systems solely on accuracy or styling is insufficient,” said Triantoro. “Designing for true human flourishing requires evaluating the continuous, unvoiced biological toll these systems exert on the people working alongside them.”