What Makes a Historical AI Character Believable? A Curator’s Checklist



Picture a visitor standing in front of a museum case, phone out, asking ChatGPT what she is looking at. Nobody at the museum suggested it. She simply did what she might now do in a restaurant, shop or unfamiliar city.

That is not an unusual possibility. Thirty-two per cent of museum visitors used a general AI assistant in the past 12 months to ask about something they saw at a museum, and 17% did so on their most recent visit (Musa Guide, State of AI in Museums 2026, surveyed 5 May 2026, n=395 across the UK, US, Germany and France).

A museum with no AI character on its floor does not have an AI-free floor. It has AI in the room already, just not the museum’s own.

Believability is not primarily a visual standard. It comes from evidence, boundaries, consistency and accountability.

The trust bar is already high

The presence of AI on visitors’ phones does not mean they want a museum-branded version of the same technology without controls. Once an AI speaks with the museum’s authority, visitors apply a much higher standard.

In a January 2026 survey of 2,045 US adults, the American Alliance of Museums found that only 9% of respondents were comfortable with museum staff using AI whenever they chose, including in exhibition text. Forty-five per cent wanted to know every time AI-generated content was used (AAM, AI in Museums and Community Trust, surveyed 20-23 January 2026, published 30 March 2026).

Put the two findings together and the decision facing a curator comes into focus. It was never simply “AI or no AI.” Visitors have already brought general AI into museums themselves. The decision is whether the AI voice carrying the museum’s name is bounded, sourced and disclosed, or whether it is allowed to speak with an authority it has not earned.

An off-the-shelf avatar does not answer those questions by itself. Trust depends on the sources, boundaries, disclosure and review process behind it.

For museums, the sharper question is not simply whether to use an AI character, but how tightly that character is governed. That is a governance problem before it is a technology problem. It sits alongside the questions explored on our Museums hub and in our examination of where museum visitor engagement tends to lose depth.

Here is the checklist we use with curators before a historical AI character goes anywhere near a gallery floor.

The curator’s checklist

1. Curator-approved knowledge boundaries. Build the character around source material the museum has approved, test its answers against that material and prevent it from answering confidently beyond the defined knowledge base. Where the evidence is incomplete, the character should say so rather than fill the gap with a plausible-sounding answer.
2. Appearance, language and behaviour grounded in evidence. Check clothing, equipment and physical details against reliable references. Where recordings exist, they can inform the voice. Where they do not, distinguish clearly between what is supported by evidence and what is informed interpretation. Written records may reveal vocabulary, beliefs and perspective, but they cannot tell us exactly how a person sounded.
3. A setting that fits the period and place. The environment surrounding the character, whether physical or digital, should be connected to the actual period and, where possible, the real location. A figure can be accurate in isolation and still feel false when placed in a generic or historically inconsistent setting.
4. A visible boundary for what it does not know. When a visitor asks something outside the approved material, the character should acknowledge the limit and, where appropriate, redirect the visitor. It should not invent an answer simply to keep the conversation moving.
5. Disclosure the visitor can actually see. If visitors are interacting with AI-generated content, say so before the interaction begins. Disclosure belongs beside the experience itself, not in a policy document or on a wall label three rooms away.
6. A consistent voice and point of view. A guide from another century who suddenly slips into modern marketing language breaks the credibility the exhibition has worked to create. Vocabulary, manner and perspective should remain consistent across open questions as well as prepared responses.
7. An honest language plan. Decide which languages the character can support reliably, test them and state them clearly. A character that quietly fails a non-English-speaking visitor is worse than one that never claimed to support that language.
8. Accessibility built in from the beginning. Consider captions, volume, pacing, physical positioning and an alternative route to the same content for visitors who cannot, or would prefer not to, use the AI interaction.
9. Clarity about what is captured. Curators and visitors should be able to understand what the system records, why it is recorded, how long it is retained and who can access it. The system should be documented with the same care as any other visitor-facing technology.
10. A review cadence with a named owner. Someone should be responsible for approving what the character can say and reviewing it on a defined schedule. New research, changing language or an exhibition refresh should not leave the character repeating material the museum would no longer put on a wall label.

None of this is exotic. It is the editorial discipline museums already apply to wall labels, interpretation and audio-guide scripts, carried into a format that can respond to questions nobody wrote in advance.

What this looks like when built

We applied this discipline when developing Private Forsythe, the AI Coldstream Guard created for Waterloo Uncovered. He was an English-language character, governed by curated material rather than left to draw freely on a general model.

Appearance. His uniform, equipment and bearing were developed using primary reference photography of professional Napoleonic reenactors. The kit was researched and reviewed at regimental level for the Coldstream Guards.

Language and perspective. The memoirs of Arthur Clay, who fought at Hougoumont Farm during the Battle of Waterloo, informed the character’s vocabulary, restraint and point of view. The written record did not tell us exactly how he sounded, but it provided a stronger foundation than inventing a generic historical voice. Where the source material was incomplete, responses could remain measured or incomplete rather than becoming more confident than the evidence allowed.

Behaviour and setting. The brief avoided ceremonial or mythologised imagery. Private Forsythe was designed as a private soldier shaped by routine, weather and lived experience, including his posture, restraint and stillness. A military veteran collaborator reviewed details of the uniform, daily routine and regiment. The character was positioned beside one of only two surviving beech trees from the Waterloo battlefield, connecting the setting, not just the figure, to the real place and event.

Designing an AI Coldstream Guard for Waterloo Uncovered explains the full development process.

The same discipline applies whether a museum is developing a permanent gallery experience or considering how to engage school groups at a heritage site without adding staff. Interactive museum technology that cannot pass its own curatorial checklist is not ready for either audience.

Frequently asked questions

What makes an AI character believable in a museum setting?

Evidence and consistency working together. Its facts, appearance, language, behaviour and setting should follow a coherent interpretation approved by the museum. It also needs a clear boundary on what it does not know. Photorealism may suit some exhibitions, but a character does not have to look real to feel authentic. Inconsistency and false confidence break believability faster than visual stylisation.

Should a historical AI character speak multiple languages?

Only when each language can be supported and tested reliably. The experience should state honestly which languages are available rather than allowing visitors to discover its limitations halfway through a conversation.

Do visitors need to be told when they are talking to an AI?

Yes. Disclosure should be visible before the interaction begins. In the AAM’s 2026 survey, 45% of respondents wanted to know every time AI-generated content was used. Transparency allows visitors to decide whether and how they want to engage.

Can a historical AI character answer any question?

It can accept open questions, but that does not mean it should attempt to answer everything. Its knowledge should be governed by approved sources and defined boundaries. When reliable material does not support an answer, it should acknowledge that limitation rather than guess.

How is this different from a chatbot on a museum website?

A website chatbot often handles practical information such as opening hours, ticket prices and directions. A historical AI character represents a particular figure or point of view within an exhibition. Because visitors may experience that voice as part of the museum’s interpretation, the standards for historical evidence, behaviour and oversight are higher.

Is your organisation ready to operate an AI character?

This checklist focuses on historical accuracy and curatorial trust. The AI Character Readiness Scorecard examines the wider operational question: whether your organisation is ready to take responsibility for the character’s content, behaviour, data and ongoing management.

It takes around ten minutes and produces a scored report. It is worth running before the vendor conversation begins, not after.