The Fish-Voiced Priest

A scribe in tenth-century Constantinople copied out a dream. A priest stands on an altar shaped like a flask; a voice describes being torn apart with a knife and reassembled as spirit. It is one of the strangest passages in the Greek alchemical corpus — the visions of Zosimos of Panopolis. When we asked a modern AI to read the page, it rendered the scene fluently, in confident English, and introduced a character who is not there.

Two sentences, one folio

Below are two English translations of the same six lines of Greek, from the same page — folio 93 recto of the Codex Marcianus graecus Z. 299, the manuscript that carries the Corpus of the Greek Alchemists. On the left is the reading printed in the standard critical edition. On the right is what our own pipeline produced when it transcribed the manuscript directly and translated its own transcription.

Sources, verbatim: Berthelot & Ruelle, Collection des anciens alchimistes grecs (1887), p. 14 → · Marcianus gr. Z. 299, folio 93r →

One of these says the speaker is a priest named Ion. The other says he is a fish-voiced one who wants to be left alone. They cannot both be reading the same words, and only one of them is. The name of the speaker — the whole point of the sentence — is simply gone from the right-hand column.

The divergence turns on a single misread word. Where the edition reads ἰσχνοφώνως — ischnophōnōs, “in a thin voice,” describing how the priest answered — the OCR saw ἰχθυόφωνος, ichthyophōnos, “fish-voiced.” From that one slip the sentence unravels: ἐγώ εἰμι ὁ Ἰων (“I am Ion”) dissolves into ἄφες με ὁ ἐμὲ (“leave me be, O me”). A proper noun becomes a plea. The English on both sides reads perfectly; only one side reads the page.

The hand that interpolates

The page in question — folio 93r

Look at the hand. This is tenth- or eleventh-century Greek minuscule: a fast, ligatured, heavily abbreviated book script where letters flow into one another and a single stroke can carry a whole syllable. And look at what the model has to read it from. The best public scan of this codex — from Internet Culturale, the digitization we and everyone else rely on — is about 892 × 1143 pixels for the entire two-column folio. That is low: perhaps a third of the linear resolution you would want for a hand this dense, an order of magnitude fewer pixels per character. Between the script and the pixels, this is exactly the kind of surface on which a modern vision model stops reading and starts guessing — and the guesses come out fluent. A model trained on oceans of Greek will always produce something that looks like a plausible sentence, whether or not it matches the ink.

You can measure the guessing without any answer key. A faithful transcriber is deterministic: give it the same page twice, get the same letters twice. So we transcribed all 393 text pages of the Marcianus twice, with the same model on the same images, and compared the runs. They agreed with themselves only about 62% of the time at the level of individual Greek words — the model disagrees with its own reading of roughly a third of the page every time it looks. (That figure is the Greek text alone; the editorial tags are stripped out before the comparison, and they in fact agree rather more — the instability is in the transcription, not the markup.) That is not the profile of a machine reading letters; it is the profile of a machine interpolating text. We have written before about why self-consistency is a floor on error and not a measure of accuracy — on clean printed pages it runs above 98%; here it collapses.

One honest caveat: that number cannot tell the two causes apart. A hard script and an under-resolved image both push a model from reading toward guessing, and self-agreement measures the guessing without saying how much is the hand and how much is the pixels. It is very likely both. What matters for a reader is that the instability is real at the resolution we can actually obtain — and higher-resolution scans of this particular codex are not publicly available. That is not a limit a better model removes. It is one the edition removes.

A footnote to a fabrication

The unsettling part is not the error. It is the poise. Beside its invented “fish-voiced one,” the model added a small, scholarly-sounding note explaining the word it had just misread:

“ichthyophonos,” likely referring to a silent or aquatic-themed initiate.

from the pipeline’s own annotation of folio 93r →

There is no fish. The model misread one letter, produced a word that does not belong in the sentence, and then reasoned confidently about what its own mistake might mean — an “aquatic-themed initiate.” Fluent, footnoted, and wrong. This is the pattern we have named elsewhere the confident hallucinator: the danger is never that the output looks broken. It is that it looks finished. The same failure mode, one layer down, is what turns a garbled Greek line into a confidently mistranslated English one.

To be fair to the machine: this is a dense, allegorical, narrative folio, the hardest thing in the book. On display script, recipe headings, and diagram labels the same model is nearly flawless — it reads the famous alchemical motto ἑν τὸ πᾶν, “the One, the All,” off a ringed diagram without a stumble. The problem is that you cannot know, from the output alone, which kind of page you are on. Every page comes back fluent.

The edition carries the text

The fix is not a better OCR pass. It is 140 years old. In 1887–88 Marcellin Berthelot and Charles-Émile Ruelle published the Collection des anciens alchimistes grecs, the critical edition of the Greek alchemical corpus — transcribed, letter by letter, from this very manuscript. Where our model guessed “fish-voiced,” Berthelot’s editors read Ἰων, Ion, and recorded in their apparatus that the surviving witnesses themselves disagree — one copy reads οἶων, another ὁ ὦν — before adjudicating between them. That is work OCR cannot do. It is not cleaner character recognition; it is scholarly judgment exercised across a thousand years of copies.

So we invert the usual assumption. For a hard hand, the manuscript is not the source of the text. It is the source of the facsimile. The edition carries the words; the manuscript carries everything that is only true of the physical object — the ink, the layout, the marginalia, the diagrams, the fact of the thing. This is not a demotion of the manuscript. It is a division of labour that respects what each one can actually be trusted to say.

The join

Berthelot’s edition and the digitized manuscript turn out to share a coordinate system. Because Berthelot transcribed from this codex, his printed pages are studded with its folio marks — “f. 92 v.,” an inline “(f. 171 r.),” a headnote “transcribed from M, f. 92 v.” — and the digitization’s page labels carry the same foliation. They lock together directly. We built that concordance and verified it against independent anchors; 195 folios now sit paired, page for page, with the running sequence staying monotonic — the signature of a correct alignment rather than a lucky one.

On the reader, Berthelot’s Greek and English now sit above the manuscript’s own AI transcription as the reading text of record. The machine reading is not deleted — it is demoted to a clearly flagged aid, labelled for exactly what it is. Nothing in the manuscript’s own data is overwritten; the pairing is an additive layer, so every folio now offers a three-way comparison: the facsimile, the AI’s attempt, and the critical edition, side by side. You can see it on folio 93r, where Ion is restored to his own sentence.

We are deliberately careful about how confident this looks. We do not publish an accuracy percentage for the AI transcription, because we have no ground truth to measure it against — both the transcription and any second reading are machine-made. What we publish instead is honest about its own limits: a stability signal drawn from how much the two OCR passes agree, and a badge marking whether a critical edition is available for that folio. Green on the clean pages, a caution flag on the dense narrative ones. A number we could stand behind, not a number that flatters the machine.

An old rule, newly kept

None of this is new. We already refuse to OCR the hardest scripts directly. When the library holds Akkadian or Egyptian material, the text is carried by the printed scholarly edition — King’s Seven Tablets of Creation, Budge’s Book of the Dead — whose transliteration and translation read cleanly, with the wedges and glyphs shown as plates. We have asked whether AI can read cuneiform and whether it can read hieroglyphs, and the honest answer, for now, is that it reads the printed edition far better than it reads the artifact. Greek minuscule quietly joins that shelf.

The rule generalizes cleanly: where a public-domain critical edition of a work exists — especially one whose base manuscript you are digitizing — prefer the edition’s text to raw OCR of the ancient hand. OCR the clean print; keep the manuscript for the facsimile, the images, and everything object-native. It is the same instinct that runs through our other work: trust the catalogued, adjudicated record over a model’s unaided fluency, and be loud about uncertainty rather than papering over it.

The flashlight and the witness

The manuscript still matters more than the edition. It is the only witness to what a scribe’s hand actually did in the tenth century; the edition is a reading of it, and a good reading can still be revised. But a witness is not a text, and an AI is not a scholar. The model is a flashlight you carry into the archive — useful, fast, and blind to its own mistakes. On folio 93 the flashlight found a fish. Berthelot found Ion. The point of pairing them is not to decide who wins. It is to put the facsimile, the machine, and the edition on the same page, and let the reader see the difference.

Sources & further reading

• Marcianus graecus Z. 299 (=584), Corpus of the Greek Alchemists — the tenth–eleventh-century manuscript, folio 93r quoted above.
• Marcellin Berthelot & Charles-Émile Ruelle, Collection des anciens alchimistes grecs (Paris, 1887–88) — the critical edition whose base manuscript is the Marcianus.
• The Confident Hallucinator — why manuscript OCR fails fluently, and why consistency is not accuracy.
• Measuring OCR Consistency — the self-agreement method, and the baseline for printed text.
• Can AI Read Cuneiform? and Can AI Read Hieroglyphs? — the ancient-script series this post extends.