A language model does not translate the way a dictionary does — word in, word out. It reads the whole Greek clause, forms an internal representation of what the clause means, and then writes English out of that representation, one word at a time, each word chosen in light of the entire Greek sentence and everything already written. Researchers who study these systems from the inside find that meaning is stored in a form largely shared across languages: “bread” and artos light up much of the same internal machinery. Translation runs through that shared layer. The unit of fidelity is the sentence, not the word.
Each English word the model writes is a choice among candidates, and two forces pull on that choice. One is the evidence in front of it — the Greek text and its instructions. The other is its training — the patterns absorbed from everything it was built on. For most text the two agree. For scripture they pull apart. That training included the King James, its successors, and centuries of quotation, so wherever a famous verse appears, the traditional English wording is the path of least resistance: the strongest pattern in the data, and so the easiest to output. “Consider the lilies” is less translated than reproduced. It is the mechanical form of the problem a human translator has when the old version is so deeply learned it cannot be unheard — except that in a machine the pull is a measurable statistical bias, and can be partly corrected.
Two corrections were applied here. First, the instructions: the model was told to distrust the famous phrasings already in its training and to render each verse as if it had never been put into English before. Second, the choice of model: several were tried, and this draft was made by the one that leaned on the remembered wording least — that reproduced the traditional English phrase least often and stayed closest to the Greek in front of it. Why it did so is not something we can see from the outside; what we can report is what showed up on the page. The result is a fresher take.
The corrections partway succeeded, and the shape of that partial success is the most useful thing we can tell you. Most of the time the model did render the Greek in front of it rather than reach for the remembered English — that is why the draft reads the way it does. But the finished text was afterwards audited against the Greek word by word, and the pull had left its fingerprints in places. The tell is inconsistency. A model’s memory of the traditional wording is strongest where a verse is famous, and it varies from chapter to chapter — so the same Greek word comes out one way here and another way there. The errors arrive not as a policy but as a patchwork, and the patchwork is the signature. Some of what the audit found, all of it now corrected:
klaiō, which means to cry aloud — to wail — was rendered “weeping,” a word both older and gentler than the Greek. At John 11:35 the famous three words “Jesus wept” had been kept, though the Greek there switches to a different, quieter verb: he sheds tears, while the mourners around him are wailing. Doulos, a slave — someone owned — came out as “servant” sixty-eight times and “slave” ten. One Greek sentence, your faith has saved you, came out “healed you” in Mark and Matthew and “saved you” in Luke, though Mark and Luke are telling the same scene in the same words. Matthew has Jesus take up John the Baptist’s sentence exactly, word for word; our English gave the two men different words and hid it. Grammateus came out “scribes” in sixty places and “teachers of the law” in five — and those five sat together in a single chapter.
One more thing is worth saying, because it cuts both ways. The model that produced this draft was not built in the English-speaking world, and it was the one, of those tried, that leaned on the remembered English least — those two facts may well be related. But it did not escape the pull, and it is worth being clear about why not. The pull does not live in a model’s birthplace; it lives in the English it was trained on, and any model that has read the internet has read centuries of English Bibles. Nor is the King James the only groove worn into the language: some of what the audit removed was not old at all but the settled habit of modern English translation — “the Jewish leaders,” say, for a Greek phrase that reads simply “the Jews.” The machine reaches for whatever phrasing the language has already worn smooth, whichever century wore it.
The draft was then checked by other models and edited under human oversight. The whole chain stands on the shoulders of the scribes who copied the text and the scholars who established it; the machine is one more lens — with its own distortions, named here so you can look through it with open eyes. Read it as a draft to question.