TAIHU
Designing trustworthy conversational AI for humanities research
TAIHU is a conversational AI platform that helps humanities scholars explore Taiwan-focused primary-source databases and verify AI-generated answers with cited sources. The project received a 2025 Future Tech Award.
I conducted UX research and owned the end-to-end redesign, turning insights from professors and PhD students into product features and working closely with PMs and engineers on a team spanning 7 departments and 3 universities.
Role
UX Researcher
Product Designer
Team
50+ people across UX, PM, Eng. & Research
Timeline
July 2024 – Aug 2025
Scope
Research, interaction design, design system
TL;DR
I rebuilt TAIHU’s product experience from the ground up, redesigning its core flows, interactions, and design system, and turned continuously evolving requirements and researcher needs into production-ready designs with PMs and engineers.
Future Tech Award
Team award, 2025 Taiwan Innotech Expo
150+
Attendees at the 2024 international symposium
1 of ~200
Research projects in the Future Tech Pavilion
CONTEXT
TAIHU brought together leading AI and humanities researchers across Taiwan, in partnership with major cultural institutions and historical databases, to build a conversational knowledge platform that scholars could trust for serious academic research.
50+ members, not a single designer. So I became one.
I was hired as a research assistant on the UX research team, working with scholars to understand their needs. From the inside, I saw the problem: the team had strong research, advanced AI capabilities, and authoritative data sources, but no one was designing the product. The insights we gathered had no consistent path into what was being built.
I brought this concern to the project lead with an argument: if we wanted scholars to trust the product, we had to take its design seriously. Then I asked to take on that work myself.
Trust had to be designed, not assumed.
We were asking a lot of our users. Scholars who had spent decades relying on traditional research methods now had to trust AI-generated information, and eventually cite it in their own academic work. Strong research and good data wouldn’t earn that trust on their own; the product itself had to feel credible, intentional, and transparent.
As the team’s sole Product Designer, I translated research insights and team requirements into design, set the product’s direction, and redesigned the experience from the ground up.
The original interface, built without a designer.
Sources appeared as a list of links after the answer, with nothing showing which claim each one supported. The only guidance on accuracy was a footnote asking users to check answers themselves.
Research
10
researchers
20 hrs
of interviews
645
coded observations
Our UX research team conducted semi-structured interviews with 7 professors and 3 PhD students across 6 disciplines. Participants walked us through their research workflows and tested ChatGPT, Perplexity, and the early TAIHU system, showing us not just what they said they needed, but how they actually evaluated AI research tools.
We synthesized the data through thematic analysis into personas and journey maps, which revealed the real challenge.
“
I need to know where the answer came from before I can trust it.
— humanities researcher
”
The challenge wasn’t getting AI to answer, it was helping scholars verify what they could trust.
01
AI output needs evidence for scholars to believe in
Scholars felt existing AI tools weren’t yet reliable enough for academic research. Before using an AI-generated claim, they needed evidence to judge whether it was actually supported.
They also wanted the complete source, so they could check AI outputs against the original material themselves.
02
Verification can’t break the workflow
Scholars needed a quick way to inspect evidence, see where it came from, and trace it to the original material without leaving what they were reading.
what i learned from research
Citations alone weren’t enough.
Building trust in AI required visible evidence and verification that fit into the scholar’s workflow.
DESIGN APPROACH
Make verification part of the reading experience
What research showed
How I designed for it
01
AI output needs evidence for scholars to believe in
→
Make the connection between claims and evidence visible
I placed citations and source excerpts right next to the claims they relate to, giving scholars a direct connection between an answer and its underlying material so they can judge each claim for themselves.
Show which claims the evidence supports
Fact-checking began as a proposal from the engineering team. Working with the engineers, I designed its interface. Rather than judging an answer as a whole, scholars can see which claims hold up and which need a closer look.
Green marks supported claims; pink marks insufficient evidence.
02
Verification can’t break the workflow
→
Let scholars choose how deeply to verify
I organized source information in layers, from quick previews to publication details and original text, keeping the answer readable while offering a clear path to deeper verification.
Quick
Detailed
Citations at the claim
Numbered citations sit right under the paragraph they support, not in a list at the end.
A quick preview
Hover a citation to preview its source and an excerpt, enough to decide whether to dig deeper.
The full source
Publication details and the original text open over the conversation, so scholars never lose their place.
Final design
sources & citations
Follow the Evidence
Numbered citations connect AI-generated answers and historical images to their sources. Scholars can hover over a citation for a quick preview, browse the source list, and open publication details and full source text without leaving the conversation.
fact-checking
See What the Evidence Supports
Fact-checking highlights each claim by how well the evidence supports it, with labels in the upper right to help first-time users read the results.
Expanding a highlighted claim reveals a source excerpt right inside the answer, with the relevant passages in bold, so scholars can check claims against the source without leaving the reading flow.
image viewer
Cite What You See
Each historical image is displayed with a source citation beneath it, formatted to academic conventions. In one viewer, scholars can enlarge images, copy a citation in one click, download an image, or open its source details.
landing page
Animated Walkthrough
The landing page animation walks visitors through a typical interaction, from asking a question to getting an AI-generated answer with related historical images, so first-time users, including those new to conversational AI, see how it works before logging in.
chat interface
Branded Loading Animation
While the system generates a response, Taihu, the brand’s bird character, takes flight, reinforcing the brand and making the wait feel shorter.
impact
Recognized with the 2025 Future Tech Award and presented to an international audience
Showcased at the 2025 Taiwan Innotech Expo
TAIHU received the 2025 Future Tech Award and was one of nearly 200 research achievements in the Future Tech Pavilion, a national showcase organized by the National Science and Technology Council, Academia Sinica, and two ministries.
Verification at the center of the live demo
Over three days of hands-on demos, visitors from academia, industry, and government saw TAIHU search Taiwan’s humanities databases in real time and fact-check its answers. That traceability is what the redesign set out to make visible.
150+
Participants at the 2024 International Symposium
Presented to humanities scholars from around the world
TAIHU opened the 2024 International Symposium on AI and Humanities Research, which welcomed international scholars from universities including Cornell, Stony Brook, and HKUST. A project lead also gave a keynote at DADH 2024, Taiwan’s leading international digital humanities conference.
UMI CHEN
Product Designer
© 2026 UMI CHEN