Master’s thesis · Interaction design

Designing the line between helpful and human.

Two interactive prototypes exploring how tone, modality and explanatory cues shape trust, safety and a sense of control in everyday chatbot conversations.

The brief

Why interaction design matters for AI

People increasingly turn to generative AI for everyday reflection and support, not just quick answers, yet most chatbots are still plain text boxes with no sense of role or boundaries. For our master’s thesis in Human Computer Interaction, we explored how tone, interaction design, and modality shape trust, safety, and a sense of control when talking to a GenAI assistant.

We designed and tested two Framer prototypes, then studied how everyday users experienced them through interviews and think aloud testing.

Research

Talking to real, everyday users

We ran exploratory interviews first, then scenario based prototype testing with a think aloud protocol, asking participants to complete tasks and reflect on the assistant’s responses as they went.

Synthesis

Research insights → Design decisions

Concept, prototype one

Three roles, three ways to respond

Instead of one generic assistant, users chose between three conversational roles, giving the system a clearer, more legible identity before the conversation even started.

Designing an AI that feels present, not human

Generative AI is often experienced as text appearing on a screen. We wanted the assistant to feel present without pretending to be a person, so we gave it a visual companion rather than a chat window alone.

We explored the balance between approachability and clear AI boundaries. Early on, the concept was inspired by the metaphor of a spirit in a lamp, exploring how an AI could appear when needed while remaining clearly identifiable as an AI rather than a person. This evolved into an abstract visual companion that felt expressive and socially present without implying genuine emotion or consciousness.

Design decision 1: a distinct visual language per mode

Each mode had its own visual identity, giving conversations a distinct tone from the start: Personal Reflection used a broader range of expressions to support reflective conversation; Analytical Assistant had its own look, including glasses and a clipboard, and expressions that reinforced structured, analytical reasoning; Source Critical Guidance had its own identity and expressions reinforcing evaluation, questioning and critical assessment.

A few examples of Personal Reflection’s expressions.

Design decision 2: signalling interaction states, regardless of mode

While each mode’s visual language stayed consistent throughout a conversation, the companion also shifted expression when input was received, while a response was generated, and when it appeared, making system activity visible without relying on additional interface elements.

Together, these two decisions gave the companion a way to communicate both conversational tone and system activity at the right moments. Its abstract form kept it clearly identifiable as AI rather than a simulated person.

Structuring the interaction

Low fidelity wireframes in Figma helped us structure the interface and the user flow before investing in visual design. For prototype one, they concretised the three mode concept: a start page where users pick a mode, and the conversation structure within each one. These wireframes became the foundation for the next, more refined iteration.

Before building the interactive prototypes, we mapped every conversational path to understand user decisions, system behaviour and interaction states.

User flow diagrams mapping each mode's conversation path.

Prototype 1: Three user flows exploring how Analytical, Personal and Source-Critical conversations guide users through different interaction styles.

Visual & interaction design

Calm enough to feel safe, clear enough to trust

A minimal layout, a single purple palette used consistently across background, buttons and the AI figure, and simple, readable typography, all to reduce memory load and keep users oriented in a conversation rather than a control panel.

Mode cards (icon, title, short description) made the assistant’s options visible before the conversation began. Chat bubbles, a “Thinking…” progress indicator, and a pulsing microphone animation for voice input gave the system a visible, predictable state at every step.

Transparency & boundaries

Telling users what the system can’t do

Explicit disclaimers and source references worked as boundary markers throughout, small signals that told people when to trust the assistant and when to look elsewhere:

Testing & iteration

Testing showed that interaction style mattered more than simply offering multiple AI personalities. Rather than refining all three roles further, we focused the second prototype on comparing text and voice interaction.

From three roles to text vs. voice

Testing prototype one told us tone and structure mattered more than the number of roles on offer, so prototype two narrowed the focus to compare modality directly:

User flow diagrams for the refined text and voice based chat.

Prototype 2: Updated user flows comparing structured text interaction with voice interaction after insights from user testing.

The result

Two working prototypes, real insight

Prototype 1 walkthrough

Prototype 2 walkthrough

Reflection

Trust is designed, not assumed

Want to see more of my UX work?

I’d love to talk about research, trust, and designing for AI driven products.