Bilingualism and Linguistic Identity

Among the components of intangible heritage, languages are at once the most fragile and the most strategic. What the people of Valais speak among themselves, in their villages and valleys, is not a regional detail. It is the most demanding test of this book's thesis, that tradition is the infrastructure of modernity, because languages are exactly what digital modernity erodes most mechanically, and at the same time what it allows us to document most powerfully.

There are few places in Western Europe where a linguistic border can be pointed to with such precision as in Valais. Travel down the Rhône valley from Brig toward Sion, and at one exact point you cross a stream, the Raspille, which has for centuries marked the divide between German-speaking Upper Valais and French-speaking central Valais. The transition is not gradual, as it can be elsewhere. It is sharp. The station at Salgesch announces its trains in Walliserdeutsch; the one at Sierre, a few kilometres downstream, announces them in French. Sierre itself is bilingual in administrative practice, but the two communities barely mix, and the stream that separates them exists in the residents' minds at least as much as it does on the map.

An officially bilingual canton with a linguistic border that is old, stable, and locatable to the metre is already remarkable. It becomes still more so once you listen to what is actually spoken on each side of the Raspille. On the German-speaking side, not standard German but Walliserdeutsch, one of the most archaic and distinctive dialects in the entire German-speaking world. On the French-speaking side, the patois, where it survives — that everyday spoken tongue of the villages, which shares little more than a name with standard French — is not French at all: it is Franco-Provençal, a Romance language in its own right, with its own history and its own characteristics.

From a strictly linguistic standpoint, then, Valais is far more than a bilingual French-German canton. Within its borders live several linguistic varieties, some with fewer than a hundred thousand speakers worldwide, whose future, in the age of large AI language models, raises questions with no precedent.

Walliserdeutsch as a rare language

Let us start with Upper Valais. Walliserdeutsch is spoken today by roughly eighty thousand people²³, which corresponds more or less to the canton's German-speaking population. Linguistically, it belongs to the group of the highest Alemannic dialects, known as Höchstalemannisch²⁴, and it has the distinction of having preserved archaic features that every other Germanic dialect lost after the Middle Ages. The genitive case is still alive among traditional speakers. Certain vowels and consonants retain a form close to Middle High German. The prosody is so distinctive that speakers of other Swiss-German dialects struggle to follow it; for a German speaker trained on standard German, Walliserdeutsch can sound like an entirely different language.

This distinctiveness owes itself to geography. Upper Valais is ringed by high mountains to the north, east, and south, and bounded by the linguistic border to the west. This enclosure shielded the dialect from the changes that, from the year one thousand onward, transformed the other Germanic vernaculars. Where standard German unified and smoothed away regional variation, Walliserdeutsch remained itself, and its sub-variants — in the Lötschental, the Goms, the Vispertal, and the region around Brig — each kept their own particularities. Someone from Saas-Fee, someone else from Zermatt, a third person from Münster all speak Walliserdeutsch, but not quite the same Walliserdeutsch, and they can place each other's home village within a few sentences.

To this internal diversity is added a diaspora. In the thirteenth and fourteenth centuries, farmers from Upper Valais crossed the high passes eastward and southward, founding settlements in Piedmont, the Aosta Valley, Ticino, Grisons, Austria's Vorarlberg, and Liechtenstein. Their descendants, the Walser, still speak dialects related to the original Walliserdeutsch, in some isolated valleys scarcely changed in seven centuries. These pockets represent some ten thousand additional speakers outside the canton, though in a more precarious position: in Italy, Walser German is barely passed on to children anymore. The linguistic centre of gravity remains in Valais.

Franco-Provençal as a forgotten layer

On the French-speaking side, the story is less glorious. The language spoken until the nineteenth century in Lower and central Valais was not French; it was Franco-Provençal, or more precisely its Valais varieties, which the inhabitants themselves call patois. Franco-Provençal is a fully independent Romance language, distinct from both French and Provençal, which historically covered French-speaking Switzerland, Savoy, the Lyon region, and the Aosta Valley. As late as the nineteenth century, it remained the everyday language of Valais villages; standard French only took hold through the school, the administration, and the Church.

Valais Franco-Provençal met the common fate of regional languages under a centralising state: it retreated, generation after generation, before standard French, until it was spoken only by an aging minority. A few pockets remain — Évolène is often cited as the commune where patois has stayed most alive — but the bulk of it fell silent over the last fifty years. There are associations working to preserve it, dictionaries, courses, patois competitions. These efforts deserve admiration. Genuine reversals of the trend, however, remain extremely rare: Basque in the Spanish Basque Country and Catalan in Catalonia are the most successful examples, and they depend on political, demographic, and institutional conditions that Valais does not have in place for Franco-Provençal.

The asymmetry between the canton's two linguistic sides is revealing. Walliserdeutsch holds: it remains the everyday language of Upper Valais, passed on to children, spoken in shops, present in regional media. Valais Franco-Provençal, by contrast, no longer holds. The same alpine geography was not enough to protect both. The difference likely stems from several combined factors: a different relationship to the language of the state, the structure of communal institutions, the Church's role in transmission, the relative prestige of the standard languages. But the difference exists, and it changes the nature of the questions worth asking about the future.

What the large models do not know

Large language models, the AI systems underpinning the tools whose use is spreading through this decade, are trained on enormous text corpora. These corpora are, by construction, dominated by the world's major languages: English first, then Mandarin, Spanish, French, German, and a handful of others. For these languages, the models have hundreds of billions of words at their disposal, covering every stylistic register and every specialised domain. For smaller languages, the available mass shrinks quickly. For regional varieties, dialects, and minority languages, it becomes critical.

Standard French is very well served by today's models: any French-speaking resident of Valais who interacts with an assistant in French receives fluent, rich, idiomatic answers, because the model has seen hundreds of millions of French sentences. Standard German is in a comparable position. Walliserdeutsch, by contrast, practically does not exist in the corpora. It has no standardised written tradition, it is chiefly oral, and what little textual material exists — a handful of collections, a few specialist websites, transcriptions of songs — amounts to barely a few megabytes, against several terabytes for standard German. For a model, Walliserdeutsch is, for all practical purposes, invisible. Valais Franco-Provençal, even less documented, is equally so.

Two direct consequences follow, pulling in opposite directions.

First risk: accelerated erosion

The first: in everyday use, AI tools will likely accentuate the erosion of regional and minority languages. No intent on the part of their designers; a mechanical effect. When someone from Upper Valais uses an assistant, they do not do so in Walliserdeutsch but in standard German or English, because that is what the model understands. When they have a letter drafted, it comes out in standard German. When they use a transcription tool, the tool struggles to recognise their dialect and defaults to standardising it. Every interaction is, on a microscopic scale, a shift toward the dominant language.

This shift predates AI: television, the press, school, and the administration had long been pushing in the same direction. But AI introduces a change of scale. Where television exposed people in Upper Valais to standard German for two or three hours a day, AI tools now insert themselves into every task of writing, research, and communication, sometimes for several hours a day, and always in the standard language. The standardising pressure becomes constant.

This risk is not specific to Valais. Every minority language in Europe faces it: Breton, Occitan, Frisian, Welsh, Sardinian, Basque, Romansh. And the first observable signals from these regions all point the same way: a shift toward the standard language that is faster than with previous media, and that intensifies with use. For Valais, this means that without deliberate intervention, Walliserdeutsch will lose ground faster over the next twenty years than it lost over the previous hundred. Franco-Provençal will disappear faster still, though at this point that changes little: its disappearance is already far advanced.

Second opportunity: documentation and transmission

The second consequence runs the other way, and it is what makes the stakes interesting for a forward-looking essay. AI does not only dilute minority languages. Under certain conditions, it becomes a powerful tool for documenting and transmitting them.

Documenting a language means transcribing it, analysing it, mapping its variants, building corpora, producing dictionaries and grammars. Until very recently, all of this required armies of linguists, decades of work, and budgets that few regional authorities were willing to grant. AI changes the equation. An automatic transcription model processes overnight the contents of a sound archive that would have occupied a human team for several years. A speech-recognition model, trained on a few dozen hours of well-documented recordings, is beginning to transcribe dialects for which no tool previously existed. A translation model, even an imperfect one, serves as an initial bridge that makes it easier for younger speakers to access their own language.

None of this is theoretical. The canton of Grisons has been actively exploring these possibilities for Romansh for several years now, with encouraging results. The Basque and Catalan regions are investing heavily in models trained on their own languages. Wales has made digital linguistic sovereignty an explicit pillar of its cultural policy. In every one of these cases, the principle is the same: do not wait for the large standard models to handle minority languages properly — they will not do so on their own, since there is no return on that investment for their developers — but build, at the regional level, the corpora and specialised models that make it possible.

For Valais, the question is therefore direct: will the canton be content to watch the accelerated erosion of its linguistic heritage, or will it take charge, as others have done, of the documentation and digital promotion of its languages? The resources exist. The expertise exists: Idiap, based in Martigny, is an internationally recognised institution in automatic language processing, and the HES-SO Valais-Wallis, in Sion, can carry the educational side and the transfer to local players. The speakers still exist: eighty thousand for Walliserdeutsch, a few hundred for Franco-Provençal. What is missing, so far, is an explicit and coordinated policy.

A strategic map

In the age of large language models, a territory's linguistic diversity becomes a strategic asset of a new kind. Not only for cultural or identity reasons, which would already be reason enough, but also for economic and political ones.

Economic, first. A territory that masters the digital documentation of its regional languages holds transferable expertise: the techniques developed to transcribe Walliserdeutsch apply to other under-represented languages, the tools built for Franco-Provençal serve comparable languages elsewhere. The market for minority languages, at the European scale, is larger than commonly assumed, and it is a niche market structurally shielded from global players, who will never see any direct commercial interest in it.

Political, next, because digital linguistic sovereignty is one facet of sovereignty as such. A canton that knows how to document its own languages retains control over how they are represented, archived, and passed on. It does not depend on the goodwill of a Californian company for its grandparents' dialect to still exist in the databases of the future. The point may sound abstract today; it will not sound abstract in ten years, once the majority of human interactions with machines pass through language models whose composition was decided outside Switzerland.

Identity-related, finally, and more deeply so. A language that dies out carries away with it ways of thinking, naming, and relating that cannot be reconstructed. A patois that has vanished is not merely one less folkloric curiosity; it is a loss of vocabulary, of conceptual categories, of a relationship to the land, with no equivalent in the dominant language that replaces it. Walliserdeutsch names the alpine landscape with a precision standard German does not have, because it was forged by that landscape and for it. When it fades, the landscape does not become simpler. It becomes harder to read.

A programme

What, then, is to be done in practice? The detail belongs to the chapters of Part Four, but the framework can be set out now.

The canton, in partnership with its research institutions, could launch a programme to build structured corpora of Walliserdeutsch and Valais Franco-Provençal: interviews with the last traditional speakers, transcription of sound archives, digitisation of existing collections. The cost would be modest against the scale of a cantonal budget; the gain, in heritage as in reputation, considerable.

On this foundation, specialised models can be trained or fine-tuned: transcription, speech recognition, possibly translation. They need not be state of the art by world standards, nor compete with the major assistants. Their purpose is to exist, to be available to institutions and residents of Valais, and to serve as a foothold for documentation and transmission.

Schools, regional radio stations, and cultural associations can be brought into a public effort of promotion. The two languages, after all, call for different gestures. Walliserdeutsch does not need saving; it is not in immediate danger. It needs the tools to remain alive in an age when standard written language invades every use. Franco-Provençal, by contrast, needs documenting while there is still time — that is, over the next fifteen to twenty years, after which the last traditional speakers will have disappeared.

There is nothing defensive in this approach. Valais holds, in its languages, an asset it still largely underuses. The time has come to make something of it.