A controlled workplace simulation published in NordiCHI ’26 on October 2, 2026, found that participants allocated a higher mean development reward to male-presenting avatar Johan (CHF 2.96) than to female-presenting avatar Johanna (CHF 2.54; p = .026). The money was designated for assistant development, not paid to the AI as wages.
What participants did
The study analyzed 189 knowledge workers. Each person interacted with a text chatbot, a desk robot and one of two human-like avatars: Johan or Johanna. All four assistant representations used the same model, gpt-4-1106-preview.
In a roughly 30-minute virtual-reality session, participants completed workplace tasks framed around employee wellbeing, inclusion or collaboration. They rated human-likeness, trust and credit on a 1–7 scale, and could allocate up to CHF 4 to each assistant’s development. Any money they did not allocate went to them. The reward analysis covered 159 participants because some employers prohibited workers from accepting compensation.
How Johan and Johanna compared
| Measure | Johan | Johanna | Statistical result |
| Human-likeness rating (1–7) | 5.09 | 4.59 | p = .029 |
| Trust rating (1–7) | 5.58 | 5.25 | p = .065 |
| Credit rating (1–7) | 5.24 | 5.05 | p = .164 |
| Mean development reward | CHF 2.96 | CHF 2.54 | p = .026 |
The reward difference was statistically significant in the study, while the trust and credit differences were not. Johan also received a higher human-likeness rating. Those results do not establish that gender presentation alone produced the reward difference.
Across the study, the human-like assistants received higher mean trust, credit and development rewards than the desk robot: 5.41 versus 5.05 for trust, 5.14 versus 4.90 for credit, and CHF 2.76 versus CHF 1.98 for rewards. The paper also reports that human-like assistants received more credit and rewards than the text chatbot, while trust did not differ significantly between them (p = .12).
Why the result does not isolate gender
The study’s authors say they could not cleanly separate gender presentation from perceived human-likeness and specific avatar and voice characteristics. The result belongs to a controlled, short workplace simulation, not a general rule about how people treat every AI assistant.
A separate study published on March 5, 2024, examined 1,400 recommendation letters generated by ChatGPT-3.5 for names historically associated with men and women. It found gender-linked differences in language that varied by prompt. That work measured text generation, not reactions to avatars or development-fund allocations.
