UX Research · The Regency Town House × University of Brighton · 2025
Coding and synthesising 40+ visitor interviews from the public evaluation of REX — a retrieval-augmented virtual expert deployed in a real Regency drawing room.
[ 01 · The brief ]
A heritage AI that sounded certain even when it was wrong.
REX was deployed as a virtual expert in a working heritage house, answering visitor questions about the building using retrieval over a curated archive. The team needed to know whether visitors actually trusted it — and what happened to that trust when the model got something wrong.
My scope was the evaluation: designing the interview protocol, running sessions on the floor with real visitors, and coding the transcripts into a set of findings the build team could act on.
[ 02 · Process ]
What I did
- Designed a semi-structured protocol covering expectation, verification behaviour and repair
- Ran sessions with 40+ visitors in situ, immediately after their interaction with REX
- Coded transcripts for trust signals, hesitation moments and abandonment triggers
- Mapped findings against the retrieval architecture to locate where confidence was being manufactured
“Confident hallucination, not limited capability, was the real threat to trust.”
Headline finding, REX evaluation
[ 03 · Findings ]
What the evidence said
Visitors forgave a model that said it did not know. They did not forgive a model that answered fluently and turned out to be wrong — and once that happened, they stopped verifying anything else it said.
The interface was doing nothing to signal the difference between a grounded answer and an extrapolated one. Confidence was uniform, so trust collapsed uniformly.
[ 04 · Outcome ]
What changed
The findings fed directly into how the team surfaced provenance and uncertainty in the interface. The recommendation was not to make the model better — it was to make its confidence legible.
The lesson I carry into every AI product now: users do not need a model that is always right. They need one whose certainty they can calibrate against.
Reflection