AI Is Rewriting the Museum Audio Guide — and Museums Are Testing the Limits
A new generation of museum interpretation is taking shape in Tokyo, Bangkok, and Miami. Artlas, launched in 2025 by former Google engineer Grace Yao, is being piloted at the Mori Art Museum, Dib Bangkok, and the Institute of Contemporary Art Miami, where it is offering visitors personalized audio guides and conversational tools built around approved institutional content.
According to company figures, the platform has generated more than 25,000 personalized audio guides since December 2025. It supports more than 20 languages and adjusts its responses to a visitor’s interests, available time, language, and level of knowledge. In practice, that means the same artwork can be framed in sharply different ways depending on who is standing in front of it.
Yao describes Artlas as “an AI layer for museum knowledge and visitor experience,” rather than a conventional audio guide. Using Georges Seurat’s “A Sunday Afternoon on the Island of La Grande Jatte” (1884-86) as an example, she says a child might be led through a looking game, while an adult could receive broader social and historical context. Specialists, meanwhile, might be offered discussion of color theory or pigment change over time.
The platform arrives at a moment when museums are weighing the promise of AI against longstanding concerns about authority and accuracy. Yao argues that Artlas is designed to supplement, not replace, human educators and guides. “A great curator, educator or human guide brings something very special: lived experience, personality, judgement, taste, emotion and sometimes a story that only that person can tell,” she says. But she also notes that live interpretation is constrained by time, staffing, language, and cost.
To reduce the risk of hallucinations, Artlas limits responses to approved museum content and verified sources. Partner museums can review, edit, and approve the material the system generates. Yao says that safeguard is especially important in a museum setting, where a mistake about identity, religion, colonial history, or artistic intention can carry more weight than a simple factual error.
Privacy is part of the same equation. Artlas is designed to collect as little visitor information as possible, while keeping interpretation standards and approval processes under institutional control. For museums, Yao suggests, trust now depends not only on what an AI says, but on how carefully it is governed. As these systems spread, the question is no longer whether visitors will encounter AI in the gallery, but how much authority institutions are willing to share with it.
