The demo is real

If you follow faith-tech, you've seen the demos: one API call, and out comes a Bible passage in Swahili or Slovak or Low German, reasoned over "from multiple perspectives," with a textual basis in the original languages. I want to take these systems seriously, because they're the most interesting thing to happen to Bible translation tooling in decades, and because I've spent the last year building something that looks superficially identical and is, underneath, close to the opposite.

An old practice with a new engine

Translating from translations is not an AI invention, and it isn't a scandal. It's an established move in Bible translation with a respectable pedigree: for a thousand years most of Europe read translations of the Latin Vulgate, itself partly a translation of the Greek Septuagint. The Living Bible, one of the best-selling English Bibles of the twentieth century, was a paraphrase of the 1901 American Standard Version. And in missionary translation work it's standard method to this day — translation teams routinely work from a "front translation" or model text rather than from Hebrew and Greek, because that's how you get scripture to a language community in years instead of decades. Usually a human editing pass finishes the job.

What AI changed is the speed and the scale. Strip away the prompt engineering and the new mechanism is: synthesize the target verse from a set of existing translations, checked against the original-language text — in one API call, for a hundred languages. One prominent system advertised exactly this, synthesizing "up to 14 public-domain reference translations, never blindly copied." The set has to be public-domain, because everything modern is copyrighted and license negotiations don't fit in an API call. Public domain means the corpus is mostly pre-1930.

Two consequences follow, and neither is fixable with a better prompt.

First, dated philology. A century of manuscript discoveries, lexicography, and Second-Temple scholarship postdates that corpus. Where the old translations share a mistake, the synthesis inherits it with confidence, because agreement across sources is exactly what synthesis rewards.

Second, a borrowed accent. The public-domain corpus isn't theologically neutral; it has the accent of its era, which trends narrow and conservative-Protestant. "Never blindly copied" is true and beside the point. When fourteen inputs lean one way, the output leans.

This is why I call the result a meta-translation: a translation whose sources are translations. As a bootstrapping tool for languages that have nothing, it's genuinely valuable, and I mean that without condescension. But it's a generator. It has no opinion about any verse. Nobody decided anything.

The same license logic bites even when the AI works from the original language. The Augsburger GenerativBibel, a German experiment I genuinely respect for its honesty — it labels itself "not a scientifically vetted Bible edition" and points readers to human-vetted translations for serious study — translates directly from Greek. But from which Greek? Tischendorf's edition of 1872, because it's public domain. That single constraint quietly imports a century and a half of missing manuscript scholarship. Copyright doesn't just pick your reference translations; it picks your source text.

What deciding looks like

The project I run (Aperto) inverts the order of operations. Before any target-language word exists, there's an exegesis: what does the Hebrew or Greek do, what did it do to its first audience, where do the traditions disagree. The exegesis feeds a decision document; the decision document binds the draft. AI writes the prose. A panel of adversarial AI judges, running in separate calls with hard gates, attacks it: archaisms, calques, register drift, back-translation fidelity, naturalness against the best modern published translations. And a human exegete rules on every flagged decision, in writing, in public.

Our north star is one sentence: the modern reader should be hit by the text the way the original audience was. Same clarity, and same friction. Both failure modes are real. Wooden literalism keeps friction the original never had. Smoothing removes friction the original did have. A synthesis of inherited translations can't even see this axis, because its sources already chose, mostly toward smoothing, a century ago.

And the text is only half of what gets made. Every decision that shapes a verse also feeds an apparatus a generator has no reason to produce: footnotes that say where a reading is disputed and — a category I haven't seen elsewhere — where a passage has historically been misused; podcasts and songs grown from the same exegesis; a public review tool where any reader can object to any verse, on the record, and trigger a re-examination. A meta-translation ends at the text. For us the text is the load-bearing wall of a larger room, built for readers who aren't already convinced.

One verse, concretely

Psalm 23:4, "thy rod and thy staff, they comfort me." The Hebrew šēbeṭ is a club: a shepherd's predator weapon, the thing David says he used against lions and bears. The comfort in the verse is the comfort of walking a dark valley next to someone armed. The English translation tradition softened it into pastoral furniture, and every synthesis of that tradition will reproduce the furniture, fluently, in every language it touches.

Our English draft says the club keeps the wolves off. Our Polish team rejected "różdżka" because to a secular Polish reader it sounds like a magic wand. Our German dropped Luther's "Stecken," a word that barely exists outside this one verse anymore. Three languages, three different decisions, one documented reason. You may think we ruled wrongly; the point is that you can look up who ruled, and why, and argue back. There's a verse-level record and a public comment channel, and reader pushback has already triggered re-translations.

The honest tradeoff

Meta-translation scales to 100+ languages in an afternoon. We've spent eight months on three languages and are still not done arguing about one psalm. If the goal is "some scripture in every language now," the generator wins and should. If the goal is a translation someone can trust, cite, and contest, provenance is the product, and provenance is precisely what a generator cannot emit.

I'd summarize the whole thing in one line: meta-translation is a generator; a translation is a set of defended decisions. Both are useful. Only one of them can be wrong in an accountable way, and being wrong accountably is what translation has always been.

Everything is public: the methodology, the judge architecture, the decision records, the text itself under a CC license. Next post: the day my own judges passed a text three times — and they were wrong every time.