Museums Test AI Chatbots, Weigh Trust and Hallucination Risk

Museums are running a version of the generative-AI experiment that has already played out in newsrooms and law firms, and the tradeoffs are becoming clearer. The Financial Times is reporting on institutions using chatbots to let visitors ask natural-language questions about collections, with staff inside those institutions worried that hallucinated answers or inherited bias could erode the trust that makes museums useful in the first place.

The most publicly documented case in the sector is the London Museum's Clio 1.0, a conversational search agent for its collections. The museum has been unusually candid about the plumbing. OpenAI's text-embedding-3-small model creates embeddings of the museum's collection data, and Anthropic's Claude Haiku 3.5 generates conversational responses from what the embeddings return. The design bet is a strict one. Clio is trained only on the museum's own trusted data, so if the answer to a visitor's question exists elsewhere online but not inside the collection records, the bot will not surface it. That is the retrieval-augmented-generation trade in its purest form, fewer confabulations at the cost of a narrower set of questions the bot can meaningfully answer.

The bias problem is admitted in the same breath. The London Museum flags that offensive language and outdated terminology sit inside its older records, and that Clio may occasionally surface terms the institution no longer finds acceptable, even while it is being trained on inclusive standards. It also notes that dates, titles and materials can get simplified or rounded when a language model summarises a record, and that the service is currently English only. These are not marketing caveats. They are the failure modes staff have to weigh against the appeal of a friendly conversational interface.

The honest caveat is that most of the reporting still covers pilots and single institutions rather than a settled best practice. Retrieval-augmented setups reduce hallucination without eliminating it, and a bot that will only answer from inside a single collection is a much smaller product than the fluent general assistant visitors are used to from ChatGPT. What the reporting does not give you is a hard read on visitor uptake, or on whether trust in the institution measurably moves once a bot is in the mix. That is the number the sector actually needs.

The upside, if this generation of museum bots holds up, is broader. Institutions with deep, well-catalogued collections and thin staff numbers get a way to answer more questions from more visitors without having to grow the education team. Getting the guardrails right is what earns the right to that upside.