After Babel, The Machine?
An experiment in literary translation
Three AI translators rendered the same chapter of Gabriele Tergit's novel Effingers into English, alongside the established human translation. Across the three machines a single variable changed — how each formed its translating persona: inherited from a human translator's writing, built from its own reading, or none at all. Everything else was held constant.
The four columns
A · inherited
From a human's writing
A persona built from a human translator's critical and biographical writing about Tergit.
B · self-built
From its own reading
A persona the machine assembled itself, from a wide corpus — Tergit's books, scholarship, her contemporaries. Kept apart from the human translator's writing.
C · control
No persona
No persona and no research. It read the novel and translated.
H · human
The anchor
The published English translation by an established human translator of Tergit.
Findings summary
Set against the German, and against one another:
The three machine translations converge; the human stands apart.
The models stay close to the German; the human moves toward fluent English — fusing the recurring refrain into prose, voicing the dialect, naturalizing the names.
The arm built from the human translator's writing did not converge on her.
It took on her reading of the author, not her hand on the page — and sat no closer to her than the other two, in places further.
The control, with no persona, converged with the other two.
Reaching many of the same choices from the bare text — so the likeness among the models tracks a shared model and method more than the personas built for them.
Where the personas surfaced was in the working, not the result.
Less in the finished text than in the notes each arm kept and the choices it stopped to justify: the self-built arm reasoned at length and caught its own errors; the inherited arm worked by a declared "lightest hand"; the control argued its way to many of the same calls from the bare text.
The full analysis, with quoted evidence, is in the Findings.
What's on the following pages
- Method & integrity — how the translators were built and the constraints they worked under.
- The personas — the two AI personas, and how one of them changed after reading the novel.
- During the run — what happened as the translations were drafted and revised.
- Findings — where the translations converge and diverge.
- The translations — all four, aligned beside the German, to compare directly.
- Appendix — the notes, personas, and revision logs the machines produced.