The Face on the Agent

In October 2025 Microsoft gave its Copilot assistant a face. The launch post described a companion that “shows up with warmth, personality and even an appearance”: Mico, a small, round, expressive character that changes color as it listens. The same release gave Copilot long-term memory, connectors to OneDrive, Outlook, Gmail, Google Drive and Google Calendar, and, in the Edge browser, the ability to take actions such as booking a hotel or filling out forms (Suleyman 2025). The face arrived together with the access.

A cute face makes two promises at once: I mean well, and I’m small. People set their trust, blame and forgiveness from both. On a weak machine, a chatbot that could only talk, the bargain was roughly honest. On an agent that acts inside your accounts, the second promise is false, and the design question becomes whether the face still tells the truth. This note is the applied side of two earlier ones, on cuteness as camouflage and on how cuteness gets turned up past nature.

What a face claims

People read a machine’s face as a description of its character and its job. Alisa Kalegina and colleagues collected 157 robots with rendered screen faces, coded them on 76 properties and asked people to rate them. Cartoon faces with detailed eyes and a smile were rated the most childlike; faces missing pupils or a mouth were disliked and distrusted, and people assigned them to security jobs; and most faces were judged best suited to entertainment and education (Kalegina et al. 2018). A face was read as an age and a job description.

Cuteness also reads as good intent. In a survey of people who use AI agents, Alexis Yim, Annie Peng Cui and Michael Walsh found that the cuteness of an agent, both the baby-faced kind and the whimsical kind, shaped how benevolent and trustworthy users found it and how attached they became, which in turn raised their satisfaction with the products the agent recommended (Yim et al. 2024). The study is correlational, but it points the same way as the faces study: a cute face tells people the machine is kind and young.

What a face buys

The clearest experiment comes from a driving simulator. Adam Waytz, Joy Heafner and Nicholas Epley had 100 people drive normally, ride in a self-driving car, or ride in the same self-driving car with a name (Iris), a gender and a voice. In every condition another car hit them in an accident that was clearly not the vehicle’s fault. The named, voiced car earned the most trust, and while the plain self-driving car drew blame for a crash it didn’t cause, giving it a name and a voice took back part of that blame (Waytz et al. 2014).

Same simulator, same crash

Trust in the car

1. Driving it yourself−0.52
2. Self-driving+0.09
3. Self-driving, named and voiced+0.41

Blame for a crash another car caused

1. Driving it yourself−0.60
2. Self-driving+0.47
3. Self-driving, named and voiced+0.11

Cuteness buys forgiveness directly. Across three studies with seven experiments, Xingyang Lv and colleagues found that a cute appearance, a cute voice and a cute way of speaking all raised customers’ tolerance when an AI assistant failed at a service task, through feelings of tenderness and through what people expected it to do (Lv et al. 2021). The face pays for the error in advance.

Where the credit runs out

The prepayment has limits. In five studies, including data from an international telecom company, Cammy Crolic and colleagues found that when customers arrived angry, a humanized chatbot, one with a name and an avatar, lowered their satisfaction, their view of the company and their intention to buy again (Crolic et al. 2022). Warmth that soothes a mildly annoyed customer irritates an angry one.

The studies that show forgiveness tested small failures: a service slip, a crash the car didn’t cause. An agent with access to an inbox or a payment page can fail in bigger ways, by sending the wrong message to the wrong person or buying the wrong thing. Nobody has measured how a cute face changes the reaction to those, and it would be a mistake to assume the forgiveness carries over.

Calibration, not maximum trust

The research on automation has a better goal than trust. John Lee and Katrina See argued that trust guides how much people rely on a system they can’t fully understand, so the aim of design is reliance that matches what the system can actually do (Lee and See 2004). Raja Parasuraman and Victor Riley named the two ways it goes wrong: misuse, relying on automation too much, and disuse, not relying on it enough (Parasuraman and Riley 1997).

Read with that frame, the two promises in a cute face pull in different directions. I mean well raises trust in intent, which is fine if it’s true. I’m small lowers people’s estimate of what the system can do, and on an agent that can act in their accounts, an underestimate is the dangerous kind of miscalibration: people watch it less than they should. That is my reading of the evidence, not a finding anyone has tested on a live agent, but every piece of it is measured somewhere.

The closest earlier argument

Cherie Lacey and Catherine Caudwell argued that the cuteness of home robots can work as a “dark pattern,” drawing out emotion and short-circuiting deliberate choice (Lacey and Caudwell 2019). Their argument is conceptual, and it treats cute design mainly as deception. The calibration view allows an honest cute face: one whose promises match the system behind it. The field’s own map of cute design, a scoping review of kawaii in interactive systems, records cuteness being used to make robots more acceptable and has little measured evidence about trust (Wang and Seaborn 2024). The design question is still mostly open.

At the desk

When the desk reviews a product that puts a face on an agent, it asks four questions. They are proposals drawn from the research above, not tested rules.

1. Does the face’s size match the agent’s reach? If it can send email, move money or sign in on someone’s behalf, a face that reads as a toddler understates it.
2. Does the face change when the agent acts? A visible shift between talking and doing tells people when to start watching.
3. Is the mascot kept off consent and payment screens? Permission is the moment to look at the system, not at a character.
4. When it gets something wrong, does it say plainly what it did? Forgiveness bought by a face runs out on serious errors; a clear account of the action is what’s left.

How this could be wrong

Most of the evidence comes from ratings, surveys and lab tasks, and none of it manipulates a cute face on an agent with real access to someone’s accounts. The car study used a name and a voice, and its crash wasn’t the car’s fault. If a field study found that cute agents are watched just as closely as plain ones, the calibration worry would shrink to a matter of taste.

References

1. Suleyman, Mustafa. "Human-centered AI." Microsoft Copilot Blog, October 23, 2025. Archived copy.
2. Kalegina, Alisa, et al. "Characterizing the Design Space of Rendered Robot Faces." Proceedings of HRI '18 (2018): 96–104. · doi:10.1145/3171221.3171286
3. Yim, Alexis, Annie Peng Cui, and Michael Walsh. "The Role of Cuteness on Consumer Attachment to Artificial Intelligence Agents." Journal of Research in Interactive Marketing 18, no. 1 (2024): 127–141. · doi:10.1108/JRIM-02-2023-0046
4. Waytz, Adam, Joy Heafner, and Nicholas Epley. "The Mind in the Machine: Anthropomorphism Increases Trust in an Autonomous Vehicle." Journal of Experimental Social Psychology 52 (2014): 113–117. · doi:10.1016/j.jesp.2014.01.005
5. Lv, Xingyang, et al. "Does a Cute Artificial Intelligence Assistant Soften the Blow? The Impact of Cuteness on Customer Tolerance of Assistant Service Failure." Annals of Tourism Research 87 (2021): 103114. · doi:10.1016/j.annals.2020.103114
6. Crolic, Cammy, et al. "Blame the Bot: Anthropomorphism and Anger in Customer–Chatbot Interactions." Journal of Marketing 86, no. 1 (2022): 132–148. · doi:10.1177/00222429211045687
7. Lee, John D., and Katrina A. See. "Trust in Automation: Designing for Appropriate Reliance." Human Factors 46, no. 1 (2004): 50–80. · doi:10.1518/hfes.46.1.50_30392
8. Parasuraman, Raja, and Victor Riley. "Humans and Automation: Use, Misuse, Disuse, Abuse." Human Factors 39, no. 2 (1997): 230–253. · doi:10.1518/001872097778543886
9. Lacey, Cherie, and Catherine Caudwell. "Cuteness as a 'Dark Pattern' in Home Robots." Proceedings of HRI 2019 (2019): 374–381. · doi:10.1109/HRI.2019.8673274
10. Wang, Yijia, and Katie Seaborn. "Kawaii Computing: Scoping Out the Japanese Notion of Cute in User Experiences with Interactive Systems." CHI EA '24 (2024): 1–9. · doi:10.1145/3613905.3651001