A translation is not the source
A machine translation is a claim about a text, produced by a named model on a named day — not the text itself, any more than OCR output is the page.
By Dipankar · Last updated
On this page
We already argue that extracted text must name the engine that produced it. The same argument applies, more sharply, to translation — and it is easier to forget, because a translation reads like prose rather than like output.
The failure mode
A researcher working on Bengali famine reporting does not read Bengali. They run a page through a model, get fluent English, and quote it. The quotation enters a footnote, the footnote enters a chapter, and at no point does anything record that a model produced those words on a particular day from a particular scan.
Every step of that is reasonable in isolation. The result is a citation that cannot be checked, attributed to a person who never wrote it.
Three things that are not the same
The page. Ink on paper, photographed. This is the evidence.
The extraction. What an OCR engine or vision-language model believed the ink said. Already an inference, already wrong in places, and wrong in ways that correlate with damage — the parts of a page hardest to read are the parts most likely to matter.
The translation. What a language model believed the extraction meant. An inference on top of an inference, in a language where a fluent sentence carries no signal about whether it is right.
Fluency is the trap. A bad OCR pass looks bad — you see the garbled characters. A bad translation reads beautifully.
What we do about it
Every derived block of text in Archēglyph carries the id of the model that produced it, and the reader shows the original beside the translation rather than instead of it. That is not a debug affordance to hide behind a settings toggle; it is on the page, next to the words.
Transliteration gets the same treatment plus one more thing: the romanisation scheme. IAST, ITRANS and Harvard-Kyoto give three different strings for the same Sanskrit word, and a transliteration whose convention is unstated cannot be compared against other scholarship. So the scheme is chosen per book, named in the instruction the model receives, and displayed with the result. Change it and the text is regenerated — never relabelled, because the words on disk were produced under the old convention and calling them something else would be a lie about their provenance.
We also tell machines. The site’s /llms.txt says, in as many words, that a
translation here is machine-generated derived text, that it carries a model id,
and that it must not be attributed to the researcher who published the book —
they published a scan and chose a model; they did not write those sentences.
| The page image, as scanned | Original transcribed, in its own script | Romanisation · IAST machine-written, signed | Translation · English machine-written, signed |
|---|---|---|---|
| আমার সোনার বাংলা, আমি তোমায় ভালোবাসি। | আমার সোনার বাংলা, আমি তোমায় ভালোবাসি। | āmāra sonāra bāṃlā, āmi tomāẏa bhālobāsi. | My golden Bengal, I love you. |
| চিরদিন তোমার আকাশ, তোমার বাতাস, আমার প্রাণে | চিরদিন তোমার আকাশ, তোমার বাতাস, আমার প্রাণে | ciradina tomāra ākāśa, tomāra bātāsa, āmāra prāṇe | Forever your skies, your air, in my heart — |
Rabindranath Tagore, Āmār sonār bāṃlā — opening lines, first published in Bangadarshan, 1905.
Hover or tab to a line to see it in every column.
What this asks of you
Cite the version, not the dataset. Quote the source when the argument turns on the words, and the translation when it turns on the sense — and say which you are doing. When the claim is load-bearing, read the scan.
None of this is new scholarly practice. It is ordinary source criticism, applied to a source that happens to have been produced by a model twenty seconds ago rather than by a copyist four centuries ago. The unfamiliar part is only that the intermediary is fast and confident and does not sign its work.
So we sign it for it.