A director of a memorial museum recently asked us, with care and some reluctance, whether her site should offer an audio guide at all. The framing was honest: it feels commodifying. We don't want this to be a "tour" in the consumer sense. The visitors who come here are paying respects.
We took the question seriously, because it deserves to be. The short answer is yes, at most sites. But the reasons are not the reasons audio guides are usually sold. None of them are about engagement, retention, or visitor satisfaction scores. They are about respect, accuracy, accessibility, and whose voice carries.
Why the case at a memorial is not the engagement case
Most arguments for audio guides are operational. Visitors stay longer in front of objects. They remember more. They report higher satisfaction. The guide pays for itself in upgraded tickets or premium add-ons. None of that translates to a memorial site, and pretending it does is part of why the question feels uncomfortable in the first place.
A visitor at the United States Holocaust Memorial Museum is not a customer being served an experience. A visitor at the Apartheid Museum, at Tuol Sleng, at the Whitney Plantation, at the Memorial to the Murdered Jews of Europe, is doing something closer to bearing witness. The museum's job is to make that possible without distortion. Whether an audio guide helps depends on whether it serves that act or interferes with it.
The honest case for audio at these sites rests on four things, and only these four:
- It can carry survivor and descendant voices in a way panels and films cannot.
- It can give international visitors access to the site in their own language, which matters more here than almost anywhere else.
- It can hold visitors to a tone the institution chooses, instead of leaving them to set their own.
- It can remove physical and sensory barriers that would otherwise exclude survivors and elderly families.
If a proposed guide does not deliver on at least two of those, it probably should not exist. If it delivers on all four, the question shifts from whether to how.
What audio uniquely does at these sites
Testimony. That is the single strongest argument, and it is enough to carry the rest.
Memorial museums hold archives of recorded testimony, often built over decades, often by people who are no longer alive. The Shoah Foundation has more than fifty thousand interviews. The District Six Museum holds oral histories from displaced communities. The Sixth Floor Museum in Dallas, the Ulster Museum's Troubles archive, the National Civil Rights Museum in Memphis — every serious memorial site has voices in its collection that printed text cannot reproduce. The pause before someone names their sister. The grain of an elderly voice. The accent that places someone in a specific village in a specific year. None of that survives transcription.
A wall label can tell you that Maria was twelve when she arrived. A recording lets her tell you. Those are not the same act, and the visitor's relationship to the history changes accordingly.
Audio is the medium for this material. A film locks you into a single screen and a single seat. A panel paraphrases. A docent retells. A well-designed audio guide hands the testimony to the visitor unmediated, in the place where it belongs, at the moment they encounter the object or the room it refers to. That is not a service feature. It is the closest a contemporary visitor can get to the voice of someone who was there.
The second thing audio does well is language. Memorial sites attract international visitors more than most museums, and often for personal reasons — descendants tracing what happened to their families, students from countries with adjacent histories, communities in diaspora returning. A site that speaks only the local language asks those visitors to encounter their own family history through a translator app or not at all. A guide in twenty or forty languages, properly localized rather than machine-translated, removes that barrier. We have written separately about why a multilingual museum audio guide matters, and the argument is sharper at memorial sites than anywhere else. International remembrance is not optional at a site of genocide. It is the point.
The third function is harder to name but real. An audio guide, if the script and pacing are right, sets the tone for visitors who arrive without one. People who have never been to a memorial before do not always know how to behave at one. Some take selfies on the railroad tracks at Birkenau. Some speak loudly in rooms that should be silent. Some treat the visit the way they would treat any other museum stop on an itinerary. A guide that begins quietly, that pauses, that asks for stillness in specific spaces, does work that signage cannot. It does not police visitors. It models a register. Most people will follow it.
What audio cannot fix, and where it is wrong
It is worth being honest about the cases where the answer is no.
Some sites are small enough that the building and a few panels do the interpretation cleanly. A single room with twenty objects and a careful curator's text does not need narration layered over it. Audio adds friction and a sense of management to a space that was working without it. The Anne Frank House works partly because most of it is bare. The interpretation happens in the rooms themselves and in the brief written context. A constant voice in your ear would diminish that.
Some sites are designed around silence as the interpretive medium. The Stolpersteine in European cities are not accompanied by recordings, and they shouldn't be. Maya Lin's Vietnam Veterans Memorial is meant to be encountered without narration. The empty library at Bebelplatz. The reflecting pools at the 9/11 Memorial. These are spaces where a designer made a deliberate choice to withhold language, and an audio guide layered over them works against that choice. A site can offer audio elsewhere — in the museum building, in the documentation center — and leave the memorial space itself uncommented.
Some sites operate under stakeholder constraints that rule out a guide entirely, or rule out the format an outside vendor would propose. Indigenous sacred sites, sites where descendant communities have asked for a particular mode of engagement, sites where survivor groups have specifically said no to recorded interpretation — these are not cases for a technology pitch. The institution's job is to listen first and decide whether audio fits at all. Sometimes the answer is no, and that answer is correct.
A weak audio guide is worse than none. A recording from 2008 with a cheerful voice describing atrocity will damage a visitor's trust in the institution faster than any other failure. If the budget cannot support a properly produced guide with current scholarship, current testimony permissions, and proper localization, the right move is to wait. Memorial sites cannot afford the version of an audio guide that other museums can sometimes get away with.
The AI question, taken seriously
Most discussions of AI at museums treat hallucination as a quality risk. At a memorial site it is a category of harm.
A general-purpose language model trained on the open internet has read Holocaust denial. It has read pseudohistorical revisionism, false equivalence essays, and the comment sections of every news article about slavery, partition, the Troubles, the Khmer Rouge, the Rwandan genocide. The model knows what those arguments sound like, because it was shown them. Asked an open-ended question by a visitor, it can produce a fluent and confident sentence that is also false, minimizing, or sympathetic to a position the institution has spent decades repudiating. The audio guide would have introduced into the memorial the exact thing the memorial exists to refuse.
This is not theoretical. Large models have produced answers minimizing the death toll at specific camps, suggesting moral complexity where the historical record is unambiguous, and offering "both sides" framing on questions that do not have two sides. Memorial sites should assume that any open-web AI system will eventually do this, and design accordingly.
What curatorial control looks like in practice is a closed knowledge base. The system can only draw from sources the institution has vetted: the curatorial text, verified scholarship, oral history transcripts, documentary evidence. It cannot reach outside that corpus. Free-form questions get answered from approved material or not answered at all. Outputs are logged, reviewable, and reversible. A persona designed for the site holds a fixed tone across every interaction. None of this is a guarantee, but it changes the failure mode from "the system says something invented and harmful" to "the system declines to answer a question outside its scope," which a memorial can live with.
The conservative deployment for sensitive sites is curated-only. The AI delivers the institution's narration, in the institution's languages, with no open Q&A surface. The benefits are operational: many languages, fast updates as scholarship moves, no hardware to maintain, accessibility features built in. The interpretive content stays human-authored. We have written about this design pattern in more detail in audio guides for sensitive heritage and memorial sites, and the short version is: start there, and only widen the system's freedom once you trust it under your specific conditions.
If you are evaluating vendors, the question to ask is not whether the system is "safe" or "responsible." Every vendor will claim that. The question is whether the knowledge base is closed, whether you can review every generated output, whether you can switch the system back to fully curated mode in a single afternoon, and whether the platform is built to defer to your scholarship rather than its own model's training data. If those answers are weak, the answer to "should we deploy AI here" is no. The choice between platforms is covered separately in our piece on the best audio guide options for memorial museums.
How to assess fit for your site
A short set of questions, in order.
First, does the site hold testimony or primary source audio that would benefit from being heard rather than read? If yes, audio is a serious candidate. If the collection is primarily objects and architecture, the case is weaker.
Second, does the site receive significant international visitation, or does it serve communities whose first language is not the local one? If yes, the multilingual argument alone may carry the decision.
Third, is there a stakeholder community — survivors, descendants, indigenous custodians — whose consent shapes the interpretation? Have they been consulted on whether an audio guide fits, what voices it should carry, and what should remain unmediated? If that conversation has not happened, it should happen before any procurement.
Fourth, is the budget sufficient to do this well? Sufficient means licensed testimony, proper localization (not machine translation), human script review at every stage, ongoing maintenance, and the ability to update content as scholarship advances. A one-off project that ships and then sits is a liability at a memorial site.
Fifth, is the curatorial team prepared to own the system after launch? An audio guide at a memorial is not a deliverable. It is a permanent piece of the institution's voice. Someone has to read every output, respond to community feedback, and revise the content as understanding shifts.
If those five answers line up, the case for an audio guide is strong, and the question becomes which platform and which design. If two or three of them are weak, the right move is to wait until they are not.
A note on what a soft pitch looks like here
We build Musa, an AI audio guide platform used at sites where the interpretive stakes are high. We are not the right vendor for every memorial. Some sites should not have an AI guide at all. Some should have a closed, curated, human-narrated guide and nothing more. Some should have nothing, and the silence is the design.
The one economic point worth making here, and only here: the platforms now available run on usage-based or revenue-share pricing rather than capex. For memorial sites that have historically gone without any interpretive audio layer because the old production model priced them out, this removes the financial barrier. It does not add any pressure to monetise. A memorial site can run a carefully curated, quiet, multilingual guide at no upfront cost, and use it to serve visitors who could not otherwise access the content. That is the change worth paying attention to, not anything to do with revenue.
If you are working through this question for your institution and want a careful conversation about whether the technology fits — including the cases where we would tell you it doesn't — that conversation is one we are willing to have. There is no rush on our side. Getting this right at a memorial site matters more than any deployment timeline.