Lyra

Students often use AI to save time, especially when balancing deadlines, coursework, and extracurriculars. While it can help them finish assignments faster, relying on it too heavily can lead students to move past material without fully understanding or retaining it.

Learning often requires revisiting concepts, practicing them, and working through the reasoning more than once. When students use AI to move through that process faster, they may complete an assignment without allowing themselves the opportunity to work through the problem on their own. This can create a gap between how well they think they understand something and what they actually understand.

This affected educators as well. When submitted assignments only showed the final work, it becomes difficult to determine what students understood, where they struggled, and what they may have offloaded to AI. Without seeing that process, educators also had less information about where a student needed help. This could allow students who were already struggling to continue relying on AI-generated answers without addressing the gaps in their understanding.

However, this does not mean that AI should be completely abandoned. AI can be an extremely useful tool for learning by helping students work through difficult topics quickly. One of its biggest advantages is that students can get help quickly and whenever they need it, even when a professor or tutor may not be available. The challenge was finding a way to keep that convenience while encouraging students to still work through the material themselves.

We used three research methods to understand how students were using AI and how we could design Lyra to better support learning:

I was also in charge of the survey. The survey showed a gap between how well students thought they understood AI-generated answers and some of their reported behavior.

Although most students said they usually tried to understand the answers AI gave them, 16 acknowledged submitting AI-generated work they did not fully understand or could not explain. This was a significant amount since it showcased an alarming number of students who were utilizing it for their work.

Around two-thirds also said they did, or possibly did, rely too heavily on AI for schoolwork. At the same time, most respondents were open to a tool that asked questions and encouraged them to think before giving them more help.

Across the survey, interviews, and literature review, we found several recurring patterns:

My literature review helped us connect what we were hearing from students and educators with research on how people learn.

We looked at concepts like scaffolding, self-explanation, metacognition, and cognitive load and considered how they could apply to Lyra.

We mapped the complete flow before moving into high-fidelity design. One of the challenges was giving students enough structure to support their learning without making Lyra feel restrictive or unnecessarily slow.

One area I was responsible for was the quiz students could take when setting up their preferences.

Lyra allowed students to choose how they wanted information to be presented. If they were unsure what to select, I designed a short quiz to give them a starting point.

Lyra also offered four modes for approaching a task:

The final prototype used a conversational interface to divide complex problems into smaller steps.

Rather than immediately displaying a completed answer, Lyra was designed to ask students questions and encourage them to explain their reasoning as they worked through a problem.

This connected back to our research. Students valued AI because it was fast, so we did not want reflection to feel like a completely separate assignment. Instead, the experience incorporated it into the process of getting help.

The interface supported this through shorter prompts, opportunities to revisit earlier thinking, and optional tools students could use without leaving the main conversation.

Lyra also included mind maps that organized concepts from a learning session into a visual overview.

We tested the high-fidelity prototype with five participants to evaluate its clarity, navigation, and overall experience.

Four of the five participants completed the guided flow without assistance. Participants understood the purpose of the reflective questions, and breaking information into smaller prompts made the guidance easier to process.

Because of the six-week timeline, we were not able to make another round of changes after testing. With another iteration, I would have focused on testing with more students and designing more of the conversation states.

Over six weeks, our team took Lyra from research to a high-fidelity prototype that we tested with students. This project changed how I thought about designing with AI. I started thinking less about what AI could do for students and more about how the way it responds and presents information could affect what students actually do with it. With more time, I would test Lyra with a larger group and compare it with a more traditional answer-first AI experience.