Archive · Past Project
05Conversational UI Research · UF · 2017

Designing Conversational AI
for Education

A Facebook Messenger chatbot to improve asynchronous communication between professors and students at UF. Research revealed the limits of AI in high-context human relationships — before those limits were widely understood.

Duration 2 months (Oct–Dec 2017) Role User research, conversation script design, prototyping, testing Team University of Florida · Human Experience Research Lab

The problem that seemed obvious — until the research

Professors were drowning in emails before every assignment deadline. Students were missing announcements scattered across email, learning management systems, class websites, and Facebook groups. The communication overhead was real, documented, and growing. A chatbot seemed like an obvious fix.

The research question was whether that obvious fix was actually right. We interviewed 17 students and 3 professors across undergraduate and graduate programs, conducted 2 focus groups, and spent time understanding not just what the communication problems were, but why existing systems were failing to solve them despite years of use.

"The most important finding was not about what the chatbot could do — it was about what kinds of communication a chatbot fundamentally cannot replace."

What the research actually found

01
Information overload, not information scarcity
Students weren't missing information because it wasn't available — they were overwhelmed by the number of platforms where information might live. The chatbot needed to consolidate, not add another channel.
02
The deadline email storm was solvable
Professors got 80% of their emails in the 24 hours before a deadline. Proactive assignment reminders from a bot could meaningfully reduce this — this was a real use case the chatbot was well-suited for.
03
Face-to-face was irreplaceable for nuanced questions
Students with questions that needed explanation, reassurance, or context preferred to wait for office hours rather than email. Asynchronous communication, including a chatbot, couldn't provide what they actually needed.
04
Personality mattered more than expected
A "friend-like" chatbot personality made students more willing to interact without formal tone anxiety. The conversational design was as important as the functional design.

The design: what worked, and what the research showed wouldn't

The chatbot was built on Facebook Messenger — where students already communicated — and designed around three core scenarios: proactive assignment reminders, professor meeting scheduling, and FAQ response for course logistics. These were the high-frequency, low-context interactions where asynchronous AI communication added genuine value.

Conversation script design was more complex than anticipated. Each scenario required handling fallbacks, intent switching, ambiguous inputs, and abandon states — all while maintaining a personality that felt natural rather than robotic. We mapped every branching path before writing a single response.

The most important finding
I would not recommend using a chatbot to enhance professor-student communication at scale. The research revealed that the communication problems that mattered most — nuanced questions, sensitive topics, learning difficulties — were exactly the ones that required human judgment. The chatbot solved the logistics. The logistics weren't the real problem.

This was an early lesson in the limits of conversational AI that still shapes how I approach AI research today: the question isn't "can AI do this?" — it's "should the human moment in this interaction be preserved?"

Why this project still matters to me

This was among the first projects where I had to recommend against the solution I was asked to design. The client brief was to design a chatbot. The research showed that the chatbot would solve the wrong problem. Delivering that finding clearly, with evidence, and having it accepted — that was an early formative experience in what research is actually for.

It also introduced me to conversational AI design at a time when most people in HCI hadn't started thinking seriously about it. The principles I developed here — designing for fallback, preserving human moments, calibrating personality to context — became foundation concepts I returned to repeatedly in later work on AI-native products at Meta.

← All work Thinking →