Method

Six AI renderings of one chapter and the human translation. What was held the same, what varied between arms, and how the seven are compared.

The design

Seven English versions of Chapter 25 of Tergit's Effingers. Six come from the same underlying AI model, given the same source, the same novel-reading step, and the same revision budget; the seventh is the published human translation. The arms vary on two axes: the source of the translator's persona (or its absence), and the style brief at the translation step (none, permission, or instruction).

• A — inherited. A persona built from a human translator's critical and biographical writing about Tergit. No style brief.
• B — self-built. A persona the machine wrote for itself, after assembling and reading a wide corpus: Tergit's own books, scholarship about her, and her German contemporaries. The human translator's writing was excluded. No style brief.
• C — control. No persona, no research. No style brief.
• C-let — control + permission. Same as C, with an explicit permission added to the brief: leave to anglicize foreign vocabulary, render dialect in plain English, fuse paragraphs, naturalize syntax, find English rhyme for embedded lyric, or domesticate culturally specific references — where the choice serves the English text. A permission, not a requirement.
• D-let — self-built, broader corpus + permission. A persona built on B's German materials plus a set of anglophone reference writers (Wharton, Powell, Mitford, Isherwood) added to the corpus. Translated under the same permission paragraph C-let received.
• D-aim — the same as D-let, under a school instruction. Identical persona, identical novel-read; the two differ only in the instruction given at the moment of translation. D-aim was told to follow an explicit translation school — domesticating translation, in the Schleiermacher tradition — as its operating principle.
• H — the human translation. Sophie Duvernoy's published English version of Chapter 25 (Effingers, New York Review Books Classics, 2025). The same translator whose writing formed A's persona.

The experiment was extended over time. The original four arms (A, B, C, H) established the baseline. The control-plus-permission arm (C-let), the broader-corpus arm (D-let), and the school-instruction fork (D-aim) were added to probe specific dimensions opened by the original findings. The site as it stood at the original four-way version is preserved at /archive/four-way/.

The procedure

The AI arms followed a four-step protocol. Arms with a persona (A, B, and the shared D) did all four; the controls (C and C-let) joined at the third.

D-let and D-aim are the same machine through the first three steps. They share the persona and the reading notes, and part ways only at the last: the conversation was wound back to the point just before translation — a fork, a fresh copy of the machine's state at that moment — and one copy was given the domesticating-school instruction where the other carried only a permission. The two were run separately under identical conditions, so the single thing that differs between their translations is that final instruction.

The constraints

• No internet. Each arm worked only from what it was given.
• A full read, not a skim. Each arm read the entire novel — 151 chapters — before translating. An early trial run showed a machine will economise (sample rather than read) even when the whole text fits in front of it; the real run used fresh instances and an explicit instruction to read in full.
• A wall around the self-built arms. B and the shared D were built from corpus material that excluded the human translator's writing about Tergit — checked across corpus, notes, persona, and translation. Their sense of the author is independent.
• The same five-pass limit. All AI arms had the same revision budget: up to five passes, each logged. Each arm decided when to stop within the budget.
• The style brief is the experimental variable. A, B, and C received no style guidance; their interpretations are their own. C-let and D-let received an explicit permission to domesticate; D-aim received an explicit school instruction at the translation step. Because each of these two pairs holds everything else constant, the difference the brief makes can be read directly in the Findings.
• Pre-converted source materials. All source texts (the German Effingers, the primary corpus, scholarly works, anglophone reference writers) were converted once to plain text and shared across arms. No instance had to extract binary text from PDFs or EPUBs; no agent's context carried serialized source files.

Reading the results

The translations were read in stages. The original four (A, B, C, H) were compared first blind — twice, on the texts alone and then with the German source beside them — and only afterward with the labels visible, a guard against reading in what one expects to find. A separate analysis compared A's, B's, and the human translator's readings of Tergit, side by side. C-let was then added for a five-way analysis, and the full seven-way comparison — folding in D-let and D-aim — completed the picture.

What that reading found: the six machines cluster tightly, while the human stands apart; a stronger domesticating instruction moves a machine further toward English without moving it closer to the human's own choices; and the human's structural moves — fusing the refrain, inventing a rhyme, writing the dialect in eye-dialect — appear in no machine version. The measured picture is in the Findings; how the arms reasoned, in the Observations.

A full procedural history of the experiment — naming changes, corrections, methodological discoveries along the way — is at How this experiment evolved.