During a two-day workshop, I explored how UX design can mediate trust in AI-powered mental health support. Taking the lead on end-to-end design, I created an adaptive conversational agent grounded in ethical flows, human-in-the-loop logic, and emotionally responsive interfaces. By turning complex principles into a working prototype, this project reflects my belief that design plays a crucial role in making AI safer, more transparent, and truly human-centered.
Context
"How can we design AI systems that users can truly trust?" This was the central question of an intensive 2-day workshop focused on creating transparent and explainable AI. Through real-world case studies and design thinking exercises, we explored how human-centered principles and ethical considerations can be embedded into AI systems—from intent modeling to interface design.
Organized as part of Werkwoche 2024, an International Design Week week hosted annually by Technische Hochschule Augsburg, my university. the workshop brought together design students and professionals for hands-on exploration of AI ethics. It was led by designers from IBM, who shared their framework for building responsible AI and facilitated collaborative sprints around trust, hallucination reduction, and user agency.
I joined a team of seven students and actively contributed to early UX ideation, including sketching user flows and mapping ethical decision points. Building on this foundation, I took the lead in UI design and prototyping—translating abstract ethical goals into tangible, trustworthy product behaviors, particularly in edge-case interactions.
Apporach
On the first day, we explored IBM’s six core AI intents and discussed how to design responsibly for real-world applications. We reflected on how trust is formed (and lost) in AI-powered products, particularly in high-sensitivity domains.
During our ideation phase, we mapped out ethical concerns and stakeholder needs across different contexts, and began identifying patterns that could be translated into product logic.
Our team chose to address mental health support—a space where AI intervention has potential but must be handled with extreme care. We designed an adaptive conversational agent that could distinguish between mild emotional distress and more serious psychological crises, adjusting its behavior accordingly to avoid overstepping its role.
Grounding Trust Through Human-in-the-Loop
To ensure safety and accountability, we grounded our approach in a Human-in-the-Loop (HITL) structure. Our AI concept was trained using data from real patient–therapist conversations, clinical assessments, and symptom descriptions.
Instead of relying on full autonomy, we designed a question tree to guide the agent’s responses, trigger escalation when necessary, and defer to human professionals based on context. This strategy allowed us to build trust not by overpromising AI capabilities, but by designing around their limitations.
v1.0 AI Agent Therapist
Based on these principles, we developed our initial concept: v1.0 AI Agent Therapist, which includes three key layers:
Data: grounded in real-world clinical and conversational sources
Model: trained and refined using reinforcement learning and safety validation
Interaction Logic: a question tree that guides conversations within safe boundaries and defers to human professionals when necessary
This architecture not only formalized how the system interacts, but also operationalized the HITL principle by embedding escalation triggers and ethical fallback points directly into the conversation flow.
Shaping AI interactions tailored to user context
To validate how our system would respond to different user needs in practice, we mapped out two representative scenarios:
User in a non-critical state: The user engages with the AI, receives supportive responses, and reflects on the conversation via a chat review—resulting in emotional relief without the need for human intervention.
User in a critical state: Based on conversation signals, the AI detects severe emotional symptoms and flags the need for professional care. The system then escalates the case to a licensed therapist, ensuring the user receives appropriate help.
While the user journey begins the same, the AI’s behavior branches based on detected severity, allowing the system to remain supportive while never acting beyond its ethical boundaries.
Feedback machnism between user and AI
To close the interaction loop and reinforce trust, we implemented a lightweight feedback mechanism between the user and the AI.
The feedback mechanism works in both directions: For the AI, feedback comes in the form of real-time uncertainty handling and emotion recognition during the conversation. When the system encounters ambiguous input, it avoids hallucination by responding with transparency and care—for example, “I’m not sure about that, but let me find out for you.” For the user, a simple rating prompt is provided at the end of each conversation, allowing them to reflect on their experience and help the system improve.
Throughout the interaction, the system continuously analyzes emotional cues to adjust tone and behavior in real time. Feedback prompts may also be triggered when signs of distress are detected. These bidirectional feedback channels—explicit on the user’s side and adaptive on the AI’s—create a dynamic loop that not only supports learning and empathy, but also actively prevents hallucination and reinforces user trust.
Outcome
Based on this structured approach, we translated our ethical design principles into a user interface that reflects empathy, clarity, and accountability.
I led the UI design, creating an emotionally attuned interface that combined voice interaction with character-driven visuals. The goal was to make each interaction feel human, trustworthy, and grounded in psychological sensitivity.
To bring these ideas to life, I rapidly prototyped the interface in Figma, taking ownership of both visual and interaction design. I focused on lo-fi prototyping for speed and clarity, incorporating animations and micro-interactions to simulate real-time responsiveness and strengthen user engagement.
Core UI Principles: Empathy, Transparency, and Explainability
Empathy
Because this product deals with sensitive emotions, I represented the AI through a character, a visual mediator designed to express empathy and build trust. The character uses friendly 2D visuals with soft shading to convey warmth and intelligence, while avoiding the uncanny valley.
Transparency
An “About AI” button placed in the top-right corner gives users access to detailed information about how the system works. This includes what is AI-generated, the model used, how responses are formed, and links to relevant clinical research — all to demystify the AI process.
Explainability (Clarity)
To support explainability, the system reveals why specific responses are generated. Users can access contextual explanations — such as tone analysis or reasoning behind a chosen therapeutic strategy — empowering them to understand, question, and trust the AI’s decisions.
Adaptive response based on user state
The AI adapts its behavior depending on the user’s emotional state—offering either reflective support or professional escalation.
Solve with AI
When the system detects that the user is not in a critical state, it engages in supportive dialogue to promote self-reflection and emotional relief. Through empathetic conversation and adaptive tone, the AI guides the user toward calmness without the need for human intervention.
Ex: “Yeah.. It happens sometimes.”
Solve with therapist
If the AI detects concerning symptoms—such as severe depression, suicidal thoughts, or emotional crisis—it triggers an escalation protocol. The user is informed that the situation exceeds the AI’s scope and is encouraged to connect with a licensed therapist for further support.
Ex: “I guess that you are experiencing depression, in which case you should see a therapist to discuss this further.”
feedback in loop
After each session, the user is invited to reflect on their emotional state and rate how the conversation went. They can review a brief summary of the session including total time, the AI-generated summary, and their selected emotion afterward.
Finally, users are given the option to share this session with a therapist. This creates a continuous feedback loop between the user, AI, and therapist supporting emotional awareness, system accountability, and potential human intervention when needed.
Building trust by handling interruptions with clarity and care
To maintain robustness and emotional safety in sensitive interactions, the AI agent proactively responds to unexpected disruptions with clear visual feedback and gentle prompts. This ensures a sense of continuity, helping reduce user frustration and reinforce trust in the system.
Network Connection Error
If the network disconnects or becomes unstable, the system immediately notifies the user with a clear message. This helps the user understand that the issue is technical—not conversational—and provides direct guidance on what to do next.
Interrupted Session
If the user accidentally closes the app or navigates away, the session is automatically saved. When they return, the system gently asks if they would like to continue where they left off. This ensures continuity while respecting the user’s agency.
Potential Features (for future iterations)
Decision Review (Transparent but Background): In critical moments such as signs of depression or self-harm, a human therapist could review the AI’s decisions in the background to ensure fairness, accuracy, and ethical alignment.
Interactive Decision Control: When the AI detects concerning symptoms, it may suggest several response options. Users can choose to accept, decline, or redirect the conversation. This gives them a sense of agency while maintaining safety.
Reflection
Although this was a short and experimental project, it offered a meaningful opportunity to reflect on the ethical tensions that come with designing AI—especially the discomfort of uncertainty and the risk of over-relying on systems that aren’t fully reliable. It made me think more critically about how design can either reinforce or repair trust in technologies that are still evolving.
What stood out most to me was the realization that while AI may take years to become truly trustworthy, UX design can already help bridge that gap. By guiding expectations, exposing limitations, and creating safer interactions, design plays a key role in shaping how people relate to AI today.
I’ve come to believe that many ethical concerns in AI are not just technical issues—they are design challenges. And UX, with its focus on human understanding, has the potential to make AI feel more transparent, responsible, and humane. This experience has deepened my interest in designing for emerging technologies, and I’m excited to continue exploring how design can support more thoughtful AI.