Term project at
Carnegie Mellon University,
Emma Zelenko, Saloni Sabnis
Form exploration, user journey mapping, UI design, high fidelity mocks and prototypes
“For users, conversational AI offers for the first time a means to interact with technology using their own words. For technology to understand them, not the other way around.”
– Andy Peart, Chief Marketing & Strategy Officer at Artificial Solutions
Through this project, we sought to design a screen-based, interactive application that responds to content and emotions expressed in conversations between counselors and clients. We explored how we could leverage the information the system has on the client through their AI-led counseling sessions to provide a customized doctor-facing application interface.
We started with a transcript of the therapy session of a person dealing with a traumatic relationship and breakup. The brief involved two parts: the first to create an audio-visual experience for counseling using AI, and later, to extend this design to help a client access a real/human counselor.
We explored existing conventions for visualizing an AI counselor by using applications like Woebot and Replica. As a regular user of Replika, I was now looking at it from a different perspective. The egg-shaped representation of the AI was very abstract yet to me it didn't seem to form a cognitive connection with an AI counselor. Woebot, on the other hand, was very robot-like and was successful at setting the right expectation. The robot-like form didn't seem to afford information to be mapped on to it. In terms of the kind of support the apps provided, both had systems that would prompt the client to seek human assistance if specific ‘red flags’ were hit.
Interviews with COUNSELORs
We had conversations with counselors to understand the appropriate posture towards a client, the information that helps them further investigate the state of the client or form assessments, and the type of interactions that occur between counselors and their clients. Following are insights from the interviews we conducted:
Best mood to convey through the form of the AI is empathy
Mirror back the client’s emotion to show they have been understood
The visual should be responsive to what the client is saying
All notes taken during the session should be the ones that the client would be okay seeing
It is okay to share general trends of moods and habits and is helpful to get a better
understanding of the client’s case (eg: showing depressive symptoms, sleeping a lot,
To guide our design principles, we created a persona of a client based on the transcript we chose. Additionally, we created the persona of another client who could benefit from an AI-based counseling service.
We utilized literature and research in emotion mapping, affective computing, and VUI, to create a culturally universal visual language for our system. We explored forms that would be most appropriate to use for AI-led counseling sessions through which the client is expressing their deepest feelings. Here are some of the considerations that informed our design decisions:
Creating a form that embodied empathy to establish a stronger connect with the client
Moving away from human-like forms to avoid creating false expectations or the
Relaxing and soothing visuals during therapy
Create visuals that are intentionally ambiguous and open to interpretation
Swarms inspired our form because of their soothing yet intelligent movements. We explored how nuances of various moods and emotions could be mapped on to the swarm. We used parameters such as density to emulate body language (tense versus relaxed), color to represent the kind of emotion (anger, disgust, fear, happiness, sadness, surprise) and movement to communicate the energy level (calm versus agitated) and govern its behavior during therapy sessions.
SafeSpace aims to provide accessibility to human counselors when the AI system recognizes a need for additional help and leverages the data gathered from the AI based counseling sessions and presents it in a meaningful way to the counselor.
Due to the intimate nature of therapy, we made efforts at all points to ensure a human-centered experience, whether building in consent requests into the VUI flows, or providing the user agency for sharing only as much data as they feel comfortable doing.
Designing for AI’s Limitations
The system would detect symptoms and stages of mental health issues from specific question templates based on clinical methodologies. If it hits specified ‘red flags’ the system pushes customized suggestions to book an in-person counselor to ensure timely treatment.
Data versus Information
We mapped the system of particles to emotion data, creating different visualizations customized for the client and doctor apps. While the chief purpose of the client visualization was to reflect the mood of the client gently, the doctor visualization provided navigable insights into the patient’s emotional states over time.
the Concept Video
what i takeaway
Diverse backgrounds, diverse perspectives: I had a fantastic experience working with my teammates who were from backgrounds that were very different from mine. They had different approaches to design and perspectives that were very different from mine, which made the outcome multi-faceted.
Visualizing emotion: It was fascinating to learn about how color communicates emotions so powerfully and when paired with form and movement, the visualization can even convey the nuances of various emotions and cross lingual and cultural barriers effortlessly. When shown the swarm visualizations, I was so excited to see people point out with such consistency the emotion the visualization was trying to reflect.
Efficient prototyping: Initially, we started off prototyping the swarm on a 3D modeling software, which generated beautiful results but was very power hungry. This slowed down our iteration dramatically. We then switched to After Effect, in which we could create similar visuals in almost half the time. This allowed us to tweak multiple parameters and generate many iterations in a short period.
More data does not equal more information!
Setting clear expectations: While working with material such as AI, it is crucial to set the right expectations from it. The design should communicate clearly what the technology can and cannot do.