MAD INSIGHTS
Governed AI: from a useful idea to a testable implementation
How to develop an AI-assisted visitor experience around approved content, visitor choices, local or external services, human control and a testable implementation scope.
BEGIN WITH THE USEFUL EXPERIENCE
What should the intelligence help someone do?
An AI-assisted experience needs a purpose a venue can explain. It might help a visitor explore an artifact, find an appropriate next exhibit or ask a question in their chosen language. The institution remains responsible for the story and for the actions available in its space.
Begin with one experience and the people it should serve. Define the useful result, the information it needs and the actions it can request. That creates a practical basis for choosing models, content and connected systems.
CONTENT AND VISITOR CHOICE
Give the answer an approved source and a clear boundary.
For Alice®, the venue controls the Body of Knowledge: its approved sources, editorial permissions and use boundaries. Curated Flexibility allows the experience to respond to different questions and interests within that agreed content scope.
Decide which preferences belong to the current visit and whether any should persist for a returning member. Anonymous session choices and a continuing profile have different retention and consent requirements. The implementation should make those choices understandable to visitors and staff.
AN ILLUSTRATIVE WALKTHROUGH
A question the approved material cannot answer.
Consider a visitor asking an avatar about an exhibit in their chosen language. In this design scenario, the request reaches an application limited to the venue’s approved sources. The material does not support the answer.
The intended behavior could be to acknowledge that limit, offer a relevant supported topic and point the visitor toward staff when appropriate. The system would need a defined route for that response, a way for an authorized operator to pause it and records that connect the interaction with the source and policy versions used.
Each part becomes a testable requirement. A convincing sample answer alone would not establish source handling, refusal behavior, language quality, operator control or recovery. The model, content and connected interfaces must be evaluated together.
WHERE THE PROCESS RUNS
Make the information boundary part of the design.
Mad has demonstrated Alice with a local language model. The development direction is to keep important venue and visitor information within the venue or participating venues. A local model is one part of that objective.
The project also needs to identify speech services, recognition, translation, content storage, diagnostics and any other connected service. Record where each processes information, what it retains and what happens when it is unavailable. This is how an intended local boundary becomes an implementation the team can inspect.
TECHNICAL DETAILTurn the experience into evidence.
Open the engineering detailClose this sectionLook for the red circuit symbol for engineering detail.
| Work package | Evidence to request |
|---|---|
| Content preparation | Approved source set, publishing responsibilities and representative supported and unsupported questions. |
| Preference handling | The information used for language, interests or accessibility, with consent, expiry and deletion behavior. |
| Action authority | Permitted actions, refusal and referral paths, operator controls and relevant safety precedence. |
| Connected services | Documented interfaces, locality, dependencies and fallback during loss of a model, network or sensor. |
| Visitor output | Review of the complete response in the required language and delivery format, using representative content. |
| Trace and recovery | Records linking request, sources, policy, decision and observed response, plus a tested recovery procedure. |
WorldModel’s ten-layer architecture and eleven policies provide the reference for governed coordination. OSOL has hard priority; AAL records without gating execution. The project chooses and qualifies the required implementation against that architecture.
FROM EXISTING WORK TO THE NEXT CAPABILITY
Keep the evidence connected to its actual scope.
At Vista del Lago, Mad’s QuickSilver and Alice exhibit installation uses age and language preferences supplied through QR or NFC interactions. That is a documented example of visitor-selected content in an exhibit.
An AI-assisted conversation, an avatar’s physical actions or venue-wide coordination introduces further content, authority and engineering requirements. Define those additions explicitly, with working demonstrations, development work and acceptance tests identified before committing to the installation.
WORK WITH MAD
What should the intelligence be allowed to change?
Bring one visitor experience and the boundaries that matter to your institution. We can turn them into a focused content, demonstration and engineering brief.
Discuss the experience