Transparency
Show what's real, what's AI
- Problem
- Visitors may not know what is verified fact and what is AI interpretation.
- Response
- Official museum facts set apart from AI interpretation, with sources and confidence shown.
05 · 2025–26 · From touch to interpretation
MuseAI.
An AI museum guide for first-time international visitors.
A quieter guide that brings objects, people and stories closer together, so visitors feel more confident, curious and connected in the museum.
Prologue From touch to interpretation
Physical access to cultural heritage led me to a later question: how can digital interaction make cultural meaning easier to understand?
01 Problem
Visitors need more than translation.
I focused on first-time international visitors to the British Museum. Mapping their visit showed that the hard parts were not only language: it was knowing why an object matters, finding a route through huge galleries, and getting tired long before the end.
- Cultural interpretation
- Literal translation
- Information overload
- Navigation
- Fatigue
- Fixed content depth
- Personal routes
- Trust in AI
- Privacy
Key insight. The guide had to give cultural context, adjust its depth, support the route, notice fatigue, and keep the visitor visibly in control of the AI.
- Cultural interpretation
Objects are hard to understand without context.
I see beautiful objects, but I don't always understand what they mean.
- Information overload
Too much text and scattered sources.
There's so much information, I don't know where to start.
- Navigation and fatigue
Large venues are tiring and hard to plan.
I get tired and unsure what to see next.
- AI transparency and control
Users want to understand, trust and stay in control.
I like the AI help, but I had to know where the information comes from.
02 Experience architecture
From evidence to a journey.
I turned the research into one coherent visit, from the first tap to a lasting memory. Each step answers one of the barriers above.
- PreferencesTell the guide who you are and what interests you.A personal start
- Exhibit recognitionScan an object and identify it instantly.From a glimpse to a story
- Cultural interpretationLayered context, not a literal translation.Cultural interpretation
- AI Q&AAsk follow-up questions in natural language.Information overload
- Route and fatigueA route that adapts to your time and energy.Navigation and fatigue
- Visit reportLook back on the visit and keep the memories.A lasting memory
03 The guide
Walk through the visit
From the door to the report.
Pick a step, or let it play.
04 Evaluation
What broke in testing.
A role-play cognitive walkthrough plus an AI-adapted heuristic evaluation. Four participants took part, about 30 to 45 minutes each. It was formative and found problems; it is not external validation.
100%
found onboarding extremely smooth.
75%
showed automation bias, trusting AI answers without checking. Only one module showed its sources.
0.75
average severity for missing source tags (AI transparency ≈ 0.5).
≈1.75
severity for user control: AI route changes were hard to reject.
Cognitive walkthrough
Task success
Share of participants who completed each core task.
AI-adapted heuristic evaluation
Average severity
Nielsen's 0–4 scale. Lower is better.
What I changed
- Automation biasShow where answers come from.
Source tags on every module, official facts set apart from AI interpretation, and a “View AI reasoning” panel.
- User controlLet people say no to the AI.
A “Cancel & re-enter your preferences” button on every route change.
- PrivacyAsk before using location history.
A just-in-time prompt before the report: “Can we use your trajectory data to build your report?”
- PermissionsMake permissions flexible.
Each permission can be changed on its own instead of all at once.
05 Designing trust
Trust, designed in.
What the evaluation found became three principles, so the AI works as a helpful companion instead of a black box.
Control
Exploration should be yours
- Problem
- One “optimal” route can feel restrictive and miss personal interests.
- Response
- Explain why a route was chosen, offer alternatives, and make the AI easy to override.
Consent
Your data, your choice
- Problem
- Visitors may be unsure what data is collected, when and why.
- Response
- Just-in-time consent, plain explanations, separate privacy controls and a clear visit report.
My part
From research to prototype.
- RoleIndividual MSc project, from research to the evaluated prototype.
- What I didUser flows, UI design, high-fidelity prototyping, prompt design, AI interaction logic, usability evaluation and AI transparency evaluation.
- MethodsJourney mapping, cognitive walkthrough, AI-adapted heuristic evaluation.