Landscape review · August 2026
Conversational AI for cultural heritage is already an active field. Museums, archives and researchers have explored curated sources, retrieval, first-person characters, voice and animated bodies. Those projects provide precedents to learn from and make the remaining questions clearer.
Selected precedents
A voice-controlled exhibition character based on diaries, letters, articles and other records. The gallery exposed the human choices behind its sources and described the result as one of many possible versions.
A conversational representation associated with a 1931 wedding ensemble, built from letters, newspaper articles and historical documents curated with the museum's historians and digital team.
A voiced 3D digital human grounded in a substantial corpus of original writings. Subsequent research makes it a useful warning that relevant source links and an authoritative manner do not necessarily establish claim-level support.
A retrieval-based historical character evaluated for accuracy, temporal boundaries and voice. Its reported embellishment and temporal slippage show why careful sources and RAG are necessary but not sufficient.
What remains difficult
Citation is not proof.A source may discuss the right topic without supporting the generated claim.
Silence invites invention.Gaps in an archive can become a confident personality or memory supplied by the base model.
Embodiment adds authority.A face, voice and first-person manner can make uncertain material feel documented.
One voice is one interpretation.A conversational character is constructed from selected evidence, instructions and software.
The contribution being tested
Intelligent Living Art has produced a first Sepolia-bound Mary reference bundle and is developing the wider portable, versioned evidence package rather than treating the material behind a character as an invisible component of one application. A compatible implementation should be able to inspect:
Sources and transformationsWhat was used, what was mechanically prepared for retrieval and what rights govern reuse.
Evidence and known gapsDirect statements, synthesis, interpretation, uncertainty and questions that should not be answered as fact.
Identity and revisionWhich version is recognised, what changed and which earlier records remain available.
AuthorityWho may revise the canonical character, choose its operator or change its runtime services.
Model qualificationHow a particular LLM or VLM performed on facts, false premises, temporal limits and character pressure.
Replaceable presentationHow local, hosted and independently built experiences can use the same disclosed identity without being the same application.
The blockchain boundary: the Mary testnet reference anchors ownership, a recognised identity revision, an account and separate creator and owner actions. That verifiable record cannot make a generated answer true, resolve historical uncertainty or turn interpretation into evidence.
Next research step
The next useful test is to ask the same reviewed questions of an ungrounded persona prompt, a conventional RAG system and the Intelligent Living Art evidence-led implementation. Factual accuracy, passage-level support, resistance to false premises, appropriate uncertainty, historical boundaries and conversational usefulness can then be compared instead of assumed.
Further reading