What this note does not claim
- Not that I worked in Hong Kong.
- Not a product name, price, launch, user count, or metric.
- Not that people already trust digital humans.
- Not peer-reviewed research.
Honest read
Catalogues and local demos are evidence for the Luddy build, not proof of engagement or evidence that an AI concierge should make personal decisions for someone. A listing is not disclosure before the first answer. I did not work in Hong Kong. I do not claim users already trust the face. Honest read: a lot of what you see is still video, anchors, and service-hall Q&A with a face on top. Some of it may help with ordinary enquiries. Still short of the harder work on trust, disclosure, and coming back.
What you meet
Most of these faces are not a new friend. They sit on an enquiry desk: a multilingual kiosk, a customer-service screen, a talking layer over a knowledge base. You ask a question. You need a clear answer. You need to know you are not talking to a person. That is the consumer day. Product responsibility starts there, not in a sales pitch or a white paper. The concierge version adds another duty: explain what private context shaped the answer.
Hong Kong catalogues as evidence
The Smart Government Innovation LAB and Hong Kong Productivity Council list AI digital-human customer service beside civic tools, in Cantonese, Mandarin, and English. The Smart Government Innovation LAB (Cyberport) and the Hong Kong Productivity Council’s Digital DIY portal publish those listings. A few public links:
- Smart LAB · AI Digital human customer service (China Mobile Hong Kong): smartlab.gov.hk
- Smart LAB · Tencent Cloud Digital Human: smartlab.gov.hk
- HKPC Digital DIY · DYXnet Digital Human Platform: ddiy.hkpc.org
- DYXnet product page (Hong Kong): dyxnet.com
Working checklist for a concierge
Working checklist I use for the Luddy build (Jade’s framework only ... not a published law and not a measured result): a concierge needs disclosure before the first answer, user-controlled memory, inspectable reasons for a suggestion, a human handoff, answers checkable against an owned source, a reason to return, and a knowledge base staff can edit after the demo. New sources that change the argument: Vesala et al. (ACM 2026) place disclosure at the latest at first interaction; PCPD’s 2024 Model Framework asks for prominent AI disclosure and human-intervention channels; mainland Deep Synthesis rules require prominent labelling when dialogue or face services may confuse the public. A face on a knowledge base is not automatic transparency. Looking human is not the product.
Two surfaces observation
Two surfaces opened 29 September 2026 (Jade’s observation only): Smart LAB A-0073 scored disclosure / handoff / checkable source / maintainability all not shown; the DYXnet Digital Human product page scored handoff pass (Contact Us) and the other three not shown. When the face is only a listing or a video, ordinary-day transparency is not yet evidenced. Not a user study.