Musing and Rambling
tales of fevered imagineering.
  • Home
  • About
  • Contact
RSS
a corner house and street lamps at night in La Boca, Buenos Aires, with a puddle in the foreground

writing translation, language, stylometry, ai, music

the foreignization device

darqlaird ⋅ November 5, 2026

I've loved an Argentine trap song for years without fully understanding it, a thing you are allowed to do with songs and not, apparently, with people. For most of that time I assumed my attachment was to melody, or her voice doing that thing it does. When I finally sat down and translated the Rioplatense line by line, literally, refusing to smooth it, I got sentences like "your liquor is not cold" and "loving you so hard was hurting me," and those broken versions were better than any clean English I could have written. Somewhere in my chest that more fluent renderings never reached. The line that got me translates to something like "scream the songs of yours my name through every street of Buenos Aires that you walk it." That last bit, that you walk it, is just wrong English. It keeps a little resumptive pronoun the grammar threw out somewhere around Chaucer. And it is the wrongness I love. A corrected version, the streets that you walk, is fine, is fluent, it's banal.

So I started wondering what my attachment was latching onto, and it turned out to be structural, also too. This song loads a familiar shape with tokens the frame does not usually carry. Comprehensible container, surprising contents. This is the rhetorical move linguists call foreignization, Venuti's word for keeping the source tongue bleeding through instead of sanding it down into target-language smooth. A correct idiomatic translation domesticates: it files the interference down until the sentence sounds like it was born in English, at which point it sounds like everything. Beckett did it to his own French. Achebe let Igbo bleed into his prose on purpose. Díaz code-switches into Spanish and does not stop to hold your hand. The technique is old and literary and almost nobody uses it in ordinary prose, which is a waste.

You can run it as a pipeline. Take an English sentence, push it through another language, drag it back literally, then edit only enough to keep it parseable. What comes out isn't random strangeness. It is pointed. Arabic aims heavy. VSO order, the construct-state possession, subject doubling, and you get lines like "returned I, but not in the way of the dream," which I find strange and beautiful. Consider "I am what remained of what was, after the going of him." That second one is a verb turned into a noun bent into an ache, and English simply does not offer that construction as a native option. You have to import it. I read it out loud to my kitchen and we felt something like tripping over the tip of an iceberg. It did what the Cazzu line does, and I don't get that from sentences I write. Singlish gives you the opposite register, the aside, the shrug, aspect where the target wants tense, already doing work no English word volunteers for. These aren't swappable. Picking your substrate is choosing your register.

I ran a batch of these through GPT-2 to see what surprisal did. Foreignizing raised it almost everywhere. Singlish up over a point and a half on average, Nigerian variants up two and a half, one line, "Last time sure can one," spiking four and a half. The one that dropped was a Philippine reduplication, "handle it slowly, gently," because English already owns that trick as style, so the substrate stops registering foreign at all. That failure maybe the most instructive result in the set. When the interference reads yours, it costs nothing. Foreignization only works where English lacks the thing.

The uncomfortable part is what this does to the idea of fluency, and it stopped being abstract the moment I fed my own writing into the same machinery and watched it come back marked.

Fluency is exactly the thing I was ever praised for. Every English teacher ever wanted vocabulary, the rare and precise word over the plain one. A published detector from the top of the leaderboard scores exactly that word choice (wc) as machine-written, rare words flagging as AI in four hundred of four hundred-thirteen documents, and it does so causally, swap the words and the verdict moves, or at least that's how it looks from outside the inference engine. Another detector I rebuilt myself flagged all thirteen of my pre-AI era posts as synthetic, maybe on that count alone? The Navajo code talkers won a war int the Pacific around the opposite bet, the rare signal as proof you were on your own side. Now a too rare word does the opposite.

So maybe the resumptive pronoun is the tell that goes the other way. That you walk it, no machine trained on smooth English would leave that stranded pronoun in. It's an error a fluent writer deletes, a typo if it is not recognized as found art. Which is exactly why I'll keep it. Selective pressure.

Home » the screensaver

↑