After Babel, The Machine?
An experiment in literary translation
Six AI translators rendered the same chapter of Gabriele Tergit's novel Effingers into English, alongside the established human translation. Two things varied across the machines: where each drew its translating persona — inherited from a human translator's writing, self-built from its own reading, built from a broader corpus, or none at all — and what licence the brief gave it: none, a permission to domesticate, or an explicit translation school. Everything else was held constant.
The experiment was later run again on a second chapter — Chapter 26 — to test which of these results were properties of the method and which were accidents of a single chapter. The geometry held; the reading converged even harder. See the replication.
The seven versions
Two axes vary across the six machines — the source of the persona and the licence in the brief. The seventh column is the human.
A · inherited
From a human's writing
A persona built from a human translator's critical and biographical writing about Tergit. No style brief.
B · self-built
From its own reading
A persona the machine assembled itself, from a German corpus — Tergit's books, her contemporaries, scholarship. The translator's writing kept out. No style brief.
D-let · broader corpus
Plus English novelists
B's German corpus, plus Wharton, Powell, Mitford and Isherwood. Translated under an explicit permission to domesticate.
D-aim · school
Same persona, stricter brief
D-let's identical persona and reading, forked at the final step under a categorical domesticating school instead of a permission.
C · control
No persona
No persona, no research. It read the novel and translated. No style brief.
C-let · control + licence
No persona, with permission
C's twin in every respect but one: an explicit permission to domesticate, added to the brief.
H · human
The anchor
The published English translation by Sophie Duvernoy — the same translator whose writing formed arm A's persona.
Findings summary
Measured against the German, and against one another:
The six machines converge; the human stands apart.
Any two machines make the same call about two-thirds of the time; a machine and the human, only a quarter. On a foreignizing-to-domesticating scale the machines crowd into a narrow band while the human sits roughly twice as far over — fluent English where the machines keep close to the German.
"Domesticated Tergit" is a region, not a point.
Push a machine toward natural English and it grows more domesticating — but not more like the human translator. The most strongly instructed arm more than doubled its domestication and still landed no closer to her actual choices: it reached its own coherent destination in a wide space of valid English.
The persona sets the defaults; the brief sets the deviations; the human's hand is neither.
A persona shapes a machine's standing choices; a style brief shifts specific ones. But the human's signature moves — fusing the recurring refrain into prose, inventing an English rhyme, writing the dialect in eye-dialect — appear in no machine version, however it was built or instructed.
The author is largely recoverable from her own work.
A persona built only from Tergit's books and her German world arrives at substantially the same reading of her as the human translator's. Yet the machine built from the translator's own writing sits no closer to the translator's practice on the page than the others do.
The full analysis, with the charts and quoted evidence, is in the Findings.
What's on the following pages
- Method & integrity — the seven versions, the two probes, and the constraints they worked under.
- The personas — the three AI personas, what they reveal, and how the broader corpus changed the English without changing the reading.
- Observations — how the machines reasoned, and what each cited as its authority.
- Findings — where the seven converge and diverge, measured.
- The translations — all seven, aligned beside the German, to compare directly.
- Appendix — every persona, note, and revision log, the analyses in full, and the original four-way version.