Read our new white paper: Scripture Quotation in Generative AI

A Short History of Getting Scripture Wrong (and Fixing It)

← Back to Posts

You’ve heard the concern that generative AI is an unprecedented threat to how faithfully Scripture gets passed down. It’s a reasonable worry, but history says this exact anxiety, in different clothes, has shown up at nearly every major transition in how God’s word has been copied and distributed.

The short version: every new medium for transmitting Scripture has introduced its own specific way to get the wording wrong, and every one has eventually gotten a fix. AI is the newest entry, not an exception.

Ancient Scribes: Word Spacing and Abbreviations

Even before the printing press, two habits created real risk. Early Hebrew scribes wrote in scriptio continua, no spaces between words, marking boundaries with dots or strokes instead; systematic spacing didn’t arrive until the Persian period. Isaiah 2:20 and Amos 6:12 remain the textbook cases of word divisions apparently lost or misread. Separately, nomina sacra, abbreviating divine names out of reverence, created near-identical collisions: in 1 Timothy 3:16, the Greek words for “who” and “God” differ by a single stroke once abbreviated, a variant that later became grounds for theological accusations, though some scholars argue it was a deliberate clarification, not a slip.

The Codex: a New Format, a New Failure Mode

Binding pages into a codex solved a real problem and created a new one. A scroll’s ending was wound inward and protected; a codex’s final leaf was physically exposed and could simply be lost. That’s one leading, still-debated explanation for why Sinaiticus and Vaticanus both end Mark at 16:8, while roughly 99.8% of surviving manuscripts include verses 9 through 20. Dictation added a second risk in this same era: one reader dictating to many scribes converted visual copying into auditory copying, and merged vowel sounds in Koine Greek made some words sound identical though spelled differently, likely the source of the “we have peace” / “let us have peace” split in Romans 5:1.

The Printing Press: Speed as a Liability

Gutenberg’s press made Scripture affordable within fifty years, and the same speed nearly cost Erasmus a book of the Bible. Racing to publish the first printed Greek New Testament before a rival edition in 1516, Erasmus found his only manuscript of Revelation missing its final leaf. Rather than wait, he translated the missing verses back into Greek from the Latin Vulgate, reconstructing readings absent from the Greek tradition he actually had. Some of those readings entered the Textus Receptus and then the King James Version, shaping English Bible reading for centuries. The press didn’t corrupt the text. Haste and a limited source base did, and the press just made the error easy to reproduce at scale.

The Digital Age: Not a New Problem, Less Friction

Misinformation isn’t unique to digital media; what changed was the infrastructure, not the appetite. Content got cheap to produce, editorial gatekeeping disappeared, and claims started arriving with a friend’s implicit trust rather than an anonymous broadcast. That infrastructure didn’t invent bad information. It removed the friction that used to limit its reach.

Where AI Actually Fits

AI is the newest entry on a list that already includes lost dividers, an exposed final leaf, and a rushed print deadline. Its version of the risk comes from probabilistic generation, a model reconstructing a quotation from training rather than retrieving one, with every generated token a fresh chance to drift.

What’s different this time: the failure mode has actually been measured, and a fix demonstrated with a deterministic guarantee. A recent 33,000-observation study found one method, output transform, that hit 100% exact reproduction across every model, translation, and passage tested. Unlike Erasmus’ notes, where the doubt was documented but the fix still took centuries, digital Scripture can be validated and fixed instantly.

FAQ

What is scriptio continua? Writing without spaces between words. Early Hebrew scribes used dots or strokes instead, and a handful of textual cruxes trace back to word divisions lost or misread.

Why does Mark’s Gospel end at 16:8 in some manuscripts? Two of the oldest manuscripts end there, while most others include verses 9-20. One debated explanation: the codex format left the final leaf exposed to loss in a way scrolls weren’t.

Did Erasmus really invent Bible verses? Not intentionally. Missing the last page of his only Revelation manuscript, he translated the missing verses back into Greek from Latin, and some of those readings shaped the King James Version.

Is AI actually worse than earlier copying technologies? Different, not simply worse. It can generate wrong text at a scale no scribe could match, but this specific failure has already been measured and a deterministic fix demonstrated.

What’s the fix? Output transform: the model only names the reference, and a deterministic process supplies the real, licensed text, so the model never generates Scripture itself.

Conclusion

Every era of Scripture transmission has introduced a new way to get the wording wrong, and every era has closed the gap eventually. This one comes with something earlier ones didn’t have: a measured, already open-source fix. Read the full study to see how it was tested, or explore the open-sourced test harness to see it built out.