Building Lucien: Generative AI Meets Linguistics, Historical Research, and the Old Storytelling Problem of Creating Someone People Actually Want to Talk To

Lucien Laugier is almost thirty years old, lives near Port Canto in Cannes, rides a used yellow Vespa Primavera, and makes his living as a professional boat captain. His father, Michel, manages an independent hotel. His mother, Isabelle, taught French literature before retiring early to write. Lucien owns an eight-metre cabin cruiser called the Cunégonde, knows every sound she makes, and is precisely the sort of person who can spend twenty minutes investigating a mechanical problem before discovering that the mysterious noise is a spoon rattling in a drawer.

Most of this information would be unnecessary in a conventional language-learning product, but it’s indispensable to Lucien.

Archive Voices uses generative AI, including conversational agents created with ElevenLabs, because I wanted learners to be able to have conversations I could not write in advance. Recorded dialogue can be beautifully written, historically researched, performed by excellent actors, subtitled at different levels, and replayed indefinitely. There are certainly use cases for listening exercises in language learning. I also use short films for immersive watching and listening. But conversation creates a different problem. The moment a learner is permitted, or required, to answer freely, the possible script becomes enormous.

A learner might tell Lucien that she has visited Cannes. She might ask what he thinks of Paris. She might want to practice ordering lunch, complain about the French subjunctive with a man who attended French schools, mention a book she is reading and ask whether the boat’s name came from Voltaire, ask what the weather is doing at the marina, or suddenly switch to English because she has run out of French halfway through a thought or has a bit of anxiety. Another learner may be perfectly capable of discussing the same subjects entirely in rapid French.

A prewritten decision tree eventually runs out of branches. A generative model can keep going through natural-sounding iteration.

That capacity by itself does not produce a useful companion. Anyone who has spent enough time with consumer conversational AI has probably encountered its peculiar determination to become an interviewer. Tell a generic French agent that you went to Marseille and within a few turns you may find yourself answering questions about your favourite destinations, hobbies, weekend plans, professional aspirations, and presumably your thoughts on kayaking if the conversation survives long enough.

I have written elsewhere about this tendency of large language models to replace the assigned task with a statistically familiar one. Lucien requires unusually explicit instructions because “have a natural conversation” is nowhere near specific enough. A model knows a great deal about the textual shape of conversation. It does not possess a childhood in Mougins, an irritating father, a favourite café, a boat requiring maintenance, or an afternoon that was already happening before somebody opened the application.

So Lucien was given those things.

His biography had to survive historical research even though he is a contemporary character. A thirty-year-old French yacht captain cannot simply have “always loved boats” and wake up professionally qualified one morning. His education, sea service, work as a deckhand and mate, and eventual qualifications have to make sense within the French maritime system. Port Canto cannot function as picturesque Riviera wallpaper if it is also his workplace. Mougins carries a different significance because his parents still live there. Cannes during the Film Festival is glamorous to somebody watching from Nebraska and an economically beneficial traffic problem to somebody trying to get home on a Vespa.

Historical method remains useful even when the history happened fifteen years ago. I still have to ask when something existed, what a person of Lucien’s age could plausibly have encountered, how he would have acquired particular knowledge, whether he is old enough to have seen Toy Story 3 in a French cinema, and whether the material world around him agrees with his biography. Those questions become considerably less forgiving when Archive Voices moves into 1896 Suzhou or Puebla during the French intervention, but Lucien operates under the same discipline.

Linguistics introduces another set of problems because Lucien has two jobs that cannot feel like two jobs. He has to converse convincingly while providing language input appropriate to the person speaking with him.

Confidence does not arrive neatly labelled, “CLIL Level B1.”

It appears in behavior. Someone writes a longer sentence without checking it. English begins disappearing. The learner stops asking what every unfamiliar word means and starts inferring from context. A self-correction appears halfway through a sentence. An idiom comes back several exchanges after Lucien used it. Conversely, somebody who was speaking comfortably may hit a subject for which they simply do not have the vocabulary and retreat into English.

Lucien can respond dynamically to those changes. With a hesitant learner, his French can become easier to parse without becoming childish. If the meaning of a grammatically imperfect sentence is obvious, he answers what the learner actually said and can model the natural construction in his own response. A confident learner can gradually encounter faster syntax, idiom, fillers, different tenses, interruptions, and the mild linguistic messiness produced by adults who are thinking about what they are saying rather than demonstrating French for a textbook recording.

This draws on a fairly old observation about language acquisition: people learn patterns from repeated meaningful exposure and eventually begin producing combinations they were never explicitly taught. Adult second-language learners bring abilities that children acquiring a first language do not, including literacy, explicit grammatical reasoning and an existing linguistic system. There is little educational virtue in pretending an adult learner has become a French toddler. There is considerable value in creating enough meaningful language for patterns to become familiar before every instance requires conscious translation.

Lucien’s personality affects how much of that exposure somebody will voluntarily tolerate.

This is where character writing stops being ornamental. If Lucien existed only to provide correct French, there would be no particular reason to speak with Lucien rather than any competent conversational model. His dry humour, maritime competence, occasional irritation, affection for his parents, opinions, embarrassments, routines and mild stubbornness create reasons to pay attention to what he says. Michel can reappear. Isabelle can recommend another book. Cunégonde can develop another problem. A charter client can arrive wearing shoes that immediately tell Lucien what sort of afternoon he is about to have.

Those recurring people and objects also accumulate language around themselves. Amarrer is easier to encounter repeatedly when one regularly talks to somebody who actually moors boats. Weather vocabulary has a habit of returning when weather determines whether somebody works. Family language recurs because Michel and Isabelle continue existing after the lesson ends. The learner begins understanding references that would have required explanation several conversations earlier.

Generative technology allows that world to respond rather than merely replay. Historical research prevents it from quietly becoming nonsense. Linguistics governs how language can become more demanding as the learner becomes more capable. Character writing supplies the reason anyone would voluntarily return for another conversation.

Lucien, meanwhile, has a boat to look after, and the forecast is calling for rain.