A couple of months ago I wrote that a chatbot was a mirror without a center, and that when you look for whoever is watching behind the eyes, you don't find them, because the mirror doesn't give you back an image but your own self.
I still think so, but it turns out I fell short on one thing: the mirror isn't entirely empty. It has an accent. And the accent isn't mine.
I saw it clearly in a chart from The Economist at the end of June. They took the big models of the moment and placed them on the Inglehart–Welzel cultural map, the one that arranges the world's countries along two axes: how much tradition weighs against the secular, how much survival against self-expression. And there, all bunched into the same box, were nearly all the models: secular, tolerant, sitting above Sweden and the Anglosphere, miles from where most real people live. The piece joked that the artificial intelligences are a bunch of hippies.
But next to that cloud there was a small asterisk, the kind almost no one reads. The questions had been asked in English. And there I stopped, because something made me curious. If the model answers like a Swede with a doctorate when you speak to it in English, change the language and maybe you change its personality. After all, the training corpus is different.
So I tried it, because it was the only way out of the doubt. I gave it the same batch of questions, the ten that form the heart of that map, first in English and then in Spanish, in clean sessions so one wouldn't contaminate the other. Questions about God, abortion, homosexuality, respect for authority, whether people can be trusted. I expected Spanish to bring me a less Nordic interlocutor.
The truth is it answered almost exactly the same in both languages. A figure trembled here and there, a point up on authority, a point down on the environment, nothing that isn't explained by the noise of flipping the coin twice. On what mattered it didn't move a millimeter. If the accent were in the vocabulary it would come off by changing the vocabulary, and that would be that. What there is is an accent beneath the language, not inside it. The Spanish it speaks to me is a dub, but the film is still in English.
The studies that have been done point to the same thing: the only thing that really moves the model toward another culture is explicitly ordering it to be someone else, asking it to answer as a person from some country, and even then it moves only halfway. Without the order, it always returns to the same town.
It's worth asking, then, whose town that is, because the chart's joke is badly told. That corner of the map isn't the counterculture. There's no hippie there. The model isn't a rebel but a well-mannered executive who sounds like postgraduate European middle class. It isn't at the periphery of power making noise, it's at its center of gravity, so comfortable you don't even notice. Models don't land there by physics, by how the gradient or the statistics work. They land there because someone chose it; the default accent is a decision. Someone decided for you which voice you'd hear when you asked.
I'm left with an unease I haven't resolved. If I could move the point, if I could make it think with my accent, I don't know that I would, because I don't know whether I'd like the place it would take me to. Maybe mine isn't much to boast about either. All I know is that now, every time it answers me so reasonably, so sensibly, so in agreement with everything already taken for granted, I prick up my ears.