Does My Bot Sound American to You?
Hooray for machine translation, although I’m sure quite a bit gets lost along the way. Which, conveniently, is exactly the blind spot that got me thinking about this in the first place.
A Finnish communications professional, Maria Morri, recently posted about her experiences writing and thinking with different AI models. A year ago, she thought Claude was exceptionally good in Finnish. Now, she finds herself wrestling with it. More interestingly, she described its Finnish output as sounding like American slogans: short, forceful statements that might work shouted through a megaphone to a crowd.
Meanwhile, Copilot has impressed her with its ability to distinguish nuance and incorporate details into a larger line of thought. But it has another habit that gets on her nerves: the relentless encouragement.
“Great idea.”
“This is a strong angle.”
“Maria, this is one of the most interesting development exercises you could do as a communications leader.”
Her question was essentially: what is the value of this praise? The model doesn’t admire the idea. It doesn’t have an opinion. And, more unsettlingly, what happens to us when our conversational partners continually reward particular ways of talking and thinking?
Reading the machine translation of that discussion, I found myself noticing something that is probably much harder for me to notice from inside American English. That my bot has an accent. Not an accent in pronunciation, obviously. An accent in assumptions.
The “American-ness” of ChatGPT, Copilot, Claude and their peers may be partly rooted in the material they learned from and the environments in which they were subsequently tuned and deployed. There is an enormous amount of English-language material in their world, and an enormous amount of American professional discourse within that: Silicon Valley product copy, LinkedIn thought-leadership, management consulting, startup pitches, marketing material, corporate training and customer-service language.
If you spend enough time with AI assistants, certain rhetorical habits become familiar:
“That’s a great question.”
“You’ve identified something important.”
“Here are three key takeaways.”
A declarative thesis. Explicit transitions. Numbered points. A punchy conclusion. Positive framing. An encouraging send-off. These are useful defaults. They make information easy to consume. They are also not necessarily culturally neutral defaults.
Contemporary American professional communication often makes approval conspicuous. We reassure the other person that their question is good, their insight is important, their direction is promising. Much of the time I barely notice it. It is conversational lubricant.
Apparently it can sound rather different in Finnish.
If competent engagement in another professional culture can be demonstrated simply by taking the question seriously and answering it, the American habit of repeatedly announcing approval may sound bizarrely performative. Why are you congratulating me? I asked you a question. Answer it. The words may all be perfectly good Finnish. The interaction can still feel foreign.
There's a tendency to talk about multilingual AI performance in terms of fluency. Does it know the vocabulary? Is the grammar natural? Does it understand idiom? Those are important questions, particularly for speakers of languages represented by much smaller bodies of training material than English. But fluency is only part of language.
A model can speak excellent Finnish while importing assumptions about how enthusiasm sounds, how disagreement should be softened, how leaders communicate, how arguments should be structured, how frequently another person should be affirmed, and what confidence is supposed to sound like.
The semantic passenger arrived safely. Its cultural luggage came along too.
I was reminded of Research In Motion’s BlackBerry PlayBook, the “professional-grade” tablet released in the early 2010s. The name caught flak in some quarters because “playbook” sounded lightweight or childish. “Do I get crayons with that?” In my world, there is nothing remotely childish about a playbook. It's what you reach for when PagerDuty wakes you up at three in the morning because something important is on fire.
The American meaning comes partly from gridiron football: the collection of practiced plays you execute in response to particular situations. From there the metaphor wandered into business, sales, operations, cybersecurity and incident response. By the time it reaches an American technology company, the sports metaphor may be almost invisible. “Follow the playbook” simply means that somebody has already thought through what to do next.
No translation error is required for somebody elsewhere to hear something entirely different. But then culture often lives in things that do not feel cultural from the inside. An accent sounds like an accent mostly to people who don’t share it.
Another Finnish LinkedIn post made the same phenomenon visible from the opposite direction. Pauliina Seppälä was writing about Finland’s stagnant productivity and what she sees as a tendency to discourage certain kinds of creative people before their ideas have had a chance to develop.
Her argument wasn’t the familiar claim that spreadsheets and data are enemies of creativity. She argued that Finnish business and investment culture is particularly comfortable recognizing functional value: technological inventions, efficiency improvements, a conventional product made somewhat cheaper or better.
It is less comfortable, she argues, recognizing value that initially appears as intuition, design, cultural positioning, brand, community or an idea for a market that does not quite exist yet.
“We also go far too obediently by the book,” she writes. Textbooks document previous successes; transformative companies sometimes write the next textbook themselves.
There is another invisible cultural default.
A serious idea looks measurable. A responsible person asks for benchmarks. A competent organization follows established process. None of these propositions is foolish. In fact, each can be a virtue. But a virtue pushed far enough can become a filter through which certain other forms of value are difficult even to see.
Which brings us, inevitably, to Moomins.
With Nokia’s phones now well into nostalgia territory, Tove Jansson’s peculiar family of trolls may have a reasonable claim to being among Finland’s most recognizable exports.
Try to imagine defending the original proposition with a conventional business case: There are trolls, but they are not especially interested in behaving like trolls. There are floods and comets and loneliness and family and strange creatures and melancholy. It is funny, gentle, occasionally frightening, and recognizably its own thing.
Decades later there are books, television series, mugs, licensing, exhibitions, tourism and theme parks. An intensely particular work of imagination became cultural infrastructure.
Finland’s other internationally recognizable exports often share the same quality: Marimekko, Aalto, peculiar strains of music, games, design. Their value is not merely that Finland executed an existing idea efficiently. Their Finnish specificity is part of what traveled.
Maria Morri’s AI observation and Pauliina Seppälä’s innovation argument therefore rhyme in an unexpected way. The AI assistant carries an implicit picture of what good professional communication looks like. An institution carries an implicit picture of what a credible business idea looks like. Neither picture needs to be stated aloud. From within the system, it can simply feel like common sense.
This is one reason I find multilingual AI more interesting than a simple leaderboard of which model writes Finnish best this month. When a model crosses languages, it can accidentally expose assumptions normally hidden from the people who made it.
The awkwardness itself becomes informative. A Finnish speaker says, “Why does this thing keep praising me?” Meanwhile, an American suddenly realizes: wait, we do that a lot, don’t we?
That may be one of the stranger gifts of these systems. In attempting to imitate everybody, they sometimes make us aware of the things we thought belonged to nobody.
So perhaps the question is not merely whether a model can speak Finnish.
If it speaks grammatically impeccable Finnish but still sounds unmistakably American, what exactly have we taught it when we say it has learned the language?