Conversational Design for Museums: From Monologue to AI Dialogue

Cultural heritage has spent decades trapped in the museum label model: fixed text, one-way communication, designed for an average visitor nobody has properly defined. A preprint published by Miriam Kenely and accessible at mkenely.com proposes changing that with a design framework specific to conversational AI in cultural heritage contexts. The document, titled From Broadcast to Dialogue, has sparked discussion on Hacker News and deserves attention beyond museum circles.

The central argument is not technological but communicative: cultural institutions have historically operated in "broadcast" mode, emitting content with no real possibility for response. Conversational AI, if designed well, can transform that relationship into something closer to real curatorial dialogue.

What the Framework Proposes

The preprint articulates several design principles tailored to heritage environments, among them:

• Domain specificity: conversational systems in museums cannot be generic chatbots. They require thematic constraints, verified sources, and a tone coherent with institutional voice.
• Uncertainty management: unlike a productivity assistant, a museum system must know when it doesn't know, and say so without eroding visitor trust.
• Layers of depth: the design should allow a ten-year-old child and a doctoral researcher to get useful answers from the same interface, adapting the level without sacrificing accuracy.
• Curatorial traceability: each relevant claim should be traceable to a source or to the curatorial decision that supports it, something generic LLMs rarely guarantee by default.

These principles are not abstract. The document illustrates them with examples of conversational flows applied to actual collections, lending practical utility beyond the typical theoretical paper.

Why It Matters and for Whom

The proposal arrives at a moment when several cultural institutions, from the Rijksmuseum to the Met, have launched or announced experiments with virtual guides based on language models. The recurring problem is not technical: current models have more than enough capacity to maintain informed conversations about art collections. The problem is one of design and knowledge governance.

An unrestricted LLM in a museum is a reputational risk: it can mix up dates, misattribute works, or generate plausible but false explanations. Kenely's framework addresses exactly that gap, proposing design layers that act as gatekeepers without sacrificing conversational naturalness.

For teams working with Claude in cultural integrations, whether through MCP servers connected to collection databases or Skills that encapsulate an institution's curatorial knowledge, this preprint offers vocabulary and structure to justify design decisions to non-technical stakeholders. It is not an implementation manual, but a well-constructed argument about what should guide those implementations.

It is also relevant for developers of custom agents: the tension between conversational openness and domain restriction that the paper describes is exactly what appears in any specialized agent project, regardless of sector.

Limitations of the Work

The preprint carries the typical weight of early-stage academic work: its practical examples are illustrative, not empirical evaluations with real users. There are no satisfaction metrics, usability studies, or comparisons between approaches. Kenely implicitly acknowledges this by framing it as a design framework rather than research with findings.

It also does not venture into implementation technicalities, how this design connects with collection management systems like Argus or TMS, or what information retrieval architecture supports the responses. This leaves the engineering entirely open.

That said, the absence of data does not invalidate the argument. The paper's value lies in systematizing a problem many teams are solving in ad-hoc and fragmented ways.

---

In our view, works like this are more useful than most museum AI demos circulating on social media: rather than showing what is possible, they explain what should be considered to make the possible also responsible. That is less spectacular, but far more useful when making real design decisions.