Can Artificial Intelligence Preserve a Culture Without Destroying It?
Imagine that the last fluent speaker of a language is dying.
Before she does, we record everything we can: her vocabulary and pronunciation, stories and songs, names for plants and places, jokes that do not quite translate, grammatical distinctions that disappeared from neighboring languages generations ago. We record conversations too, because a language is more than a dictionary. We want the interruptions, the hesitations, the way one expression suggests another.
Then we train an artificial intelligence on all of it.
Long after she is gone, the system can still speak. It can answer questions in her language and tell the old stories. Perhaps it can compose new ones. A child born fifty years later might have a conversation in a language no living human being remembers learning from a parent.
Have we saved the language?
Or have we built an extraordinarily convincing monument to it?
Human culture has always been an argument with disappearance. Languages die. Buildings collapse. Recipes change. Songs are forgotten. Crafts vanish when the last person who knows some apparently insignificant step dies without teaching it to anyone.
Most of our preservation technologies have helped by keeping traces. Writing preserves words, photography appearances, recording voices, film movement. Digital storage lets us accumulate these traces on an unprecedented scale.
AI introduces a peculiar complication because the archive no longer has to remain silent.
Give a model enough traditional songs and it may produce another song in the same idiom. Give it the records of an endangered language and it may generate sentences nobody ever recorded. The distinction is important. A recording retrieves something that happened; a generative model infers patterns from what happened and uses them to construct something that did not.
Of course, this immediately creates an awkward comparison. Isn't that what people do too?
A folk musician learns hundreds of songs before writing one. A potter imitates a teacher before developing a style. Children acquire language by absorbing patterns and then produce sentences nobody has ever spoken to them. Culture has always generated the new from the inherited.
So if a machine can do something formally similar, why shouldn't it count as the next participant in the tradition?
I think the answer has less to do with the object produced than with the relationship that produces it.
Consider a ceramic glaze made in the same village for four hundred years. A master potter knows when the kiln is ready partly by measurement and partly because she has fired it hundreds of times. The clay comes from a particular hillside. Apprentices learn by watching her, failing, being corrected and eventually developing intuitions they may themselves struggle to explain.
Now document everything. Analyze the clay and glaze chemically. Place sensors throughout the kiln. Record every movement of the potter's hands. Train systems to correlate thousands of variables with successful firings.
Eventually an automated kiln might produce the pottery perfectly. Perhaps it will produce it more consistently than the potters ever did.
Something has unquestionably been preserved. The objects still exist. Much of the technique exists. An enormous amount of knowledge that might otherwise have disappeared is available to us.
But has the craft survived?
The apprentice did not merely acquire information from the potter. She was taught by her. She argued with her. She misunderstood instructions, developed preferences, copied some habits and rejected others. Years later she may teach someone else, badly remembering something her teacher once told her and adding something of her own.
The machine can reconstruct patterns generated by that history. It does not thereby acquire a history with the people who generated them.
This is the difference I find difficult to get around. There is a difference between something continuing and something being reproducible.
We can see it in rituals as easily as crafts. Imagine documenting a festival completely: every song, costume, prayer, recipe and dance, every ceremony filmed from multiple angles.
The festival is still more than the resulting archive. It includes the fact that people come back next year. Children once carried through the procession eventually carry children themselves. Someone complains that the food isn't as good as his grandmother's. A family changes a recipe. A song gains a verse. Two people argue about who is supposed to stand where.
Someone gets it wrong.
That matters more than it might seem. Living cultures are full of errors, disagreements and mutations. Perfect reproduction is not their natural state. A tradition that cannot change may already have become something closer to an exhibit.
None of this is an argument against using AI for cultural preservation. In fact, refusing these tools in the name of authenticity could help destroy exactly what we hope to protect.
Human beings are very good at losing things. Languages have disappeared because nobody recorded their speakers. Traditional knowledge has vanished because it existed in the memories of a handful of elderly practitioners. Wars, migration, economic change and simple indifference have broken chains that might have survived if better records had existed.
AI could make those records dramatically more useful. Recordings of an endangered language become more accessible if software helps descendants learn pronunciation. Thousands of handwritten documents become more valuable when they can be searched and compared. Scattered recordings of traditional songs can be catalogued. Hundreds of hours of footage documenting a craft can become a teaching resource rather than an inaccessible archive.
The mistake is not preserving too much.
It is confusing preservation of information with preservation of the thing that information describes.
And that leads to a problem I suspect will become more important as cultural AI systems become better: who decides what goes into them?
There is no neutral version of a culture waiting to be uploaded.
Which dialect of the endangered language should the model speak? Whose pronunciation? Which stories belong in its training material? Which version of a ritual is correct? Which recipes are traditional enough? A practice documented by an anthropologist in 1920 may differ from one remembered by a family in 2026. A diaspora community may preserve customs abandoned generations ago in the country its ancestors left.
These disagreements are normal. They are part of culture.
A dataset, however, requires choices.
Someone selects the recordings. Someone labels them. Someone decides what counts as representative and what counts as an outlier. Some material will exist in abundant, machine-readable form; other material will survive only in private memories, inaccessible archives or communities that do not want it recorded.
The model then gives those choices extraordinary reproductive power.
What began as a dataset becomes a canon, and the canon can talk.
That may be more consequential than simple loss. An AI does not have to erase a culture to alter our relationship with it. It might freeze one particular interpretation and reproduce it so fluently, so conveniently and at such scale that later generations mistake the preserved version for the whole.
Convenience creates another danger.
Return to our endangered language. Suppose its remaining speakers are elderly, scattered and difficult to reach. Learning from them takes patience. They forget words. They disagree. They have lives of their own and cannot teach thousands of students.
The AI tutor is available at three in the morning. It remembers every word in its training material. It never becomes tired of correcting pronunciation.
Soon it may speak the language better, by many measurable criteria, than most of the people trying to revive it.
Why not learn from the machine?
There is no foolish answer to that question. If the alternative is that nobody learns the language at all, the machine may be an extraordinary gift.
But there is an uncomfortable possibility hidden inside its success. We might build the machine because human transmission has become fragile, then make the machine so useful that maintaining the fragile human transmission no longer seems necessary.
The preservation tool becomes the replacement.
I arrived at some of these questions through fiction. In The Last Vines on Earth, people try to reproduce things whose value seems to reside partly in properties that can be measured and partly in accumulated time. A great wine can have its soil, climate, vines, methods and oak studied and reproduced, yet something remains stubbornly attached to generations of people making it in the same place. Elsewhere, an old fermentation culture survives only because people have continuously kept it alive.
The question that interested me while writing was not whether copies are inferior. Sometimes the copy is technically better. Sometimes transplantation is precisely how a tradition survives. In the novel, a French grape disappears in France but continues for generations in Chinese soil until the supposed transplant is the only living lineage left.
That makes the problem harder, not easier.
If continuity does not require purity, and authenticity does not require staying where something began, then perhaps the important thing is neither perfect reproduction nor an untouched original. Perhaps it is the existence of a continuing relationship between people and what they have inherited.
Which makes me wonder whether preserve is the wrong verb altogether.
We preserve objects partly by preventing them from changing. Cultures survive by doing almost the opposite.
They are mistranslated and misremembered, argued over and abandoned and rediscovered. Children disappoint their parents. Apprentices alter techniques. Migrants combine traditions that were never supposed to meet. Recipes acquire ingredients someone's great-grandmother would have regarded as an abomination.
The thread survives not because every link is identical. It survives because somebody makes the next link.
AI could be extraordinarily useful in helping us do that. It can recover, organize and teach what might otherwise disappear. It can give people access to parts of their inheritance that history has made difficult to reach.
But perhaps we should be wary when the technology moves from helping someone inherit a tradition to becoming the thing from which the tradition is inherited.
The distinction is subtle. In practice it will probably be messy, and different communities will draw the line in different places.
That may be exactly as it should be.
Because if a culture is still capable of arguing about what should happen to it next, then at least one important part of it is still alive.
Comments