If you go to Rome, to the church of San Pietro in Vincoli, you’ll find one of Michelangelo’s most famous sculptures. It’s Moses, carved around 1515 for the tomb of Pope Julius II. He has an imposing beard, a stern gaze, the arms of someone who just finished carrying stone tablets. And, on top of his head, two horns.
By Jörg Bittner Unna – Own work, CC BY 3.0, Link
It’s not a whim of Michelangelo’s. He sculpted exactly what his era believed to be true. The error is much older, and it’s a translation error.
In the 4th century, St. Jerome translated the Hebrew Bible into Latin. That translation, the Vulgate, was the West’s reference text for over a thousand years. In the book of Exodus, when Moses comes down from Mount Sinai, the Hebrew describes his face with the word qaran. It means “radiant,” “shining.” But qaran shares a root with another word, qeren, which means “horn.” Without written vowels, the two sit very close to each other.
Jerome chose horns. He wrote that Moses’s face had become cornuta — horned.
And here’s the part that matters. Jerome wasn’t wrong out of ignorance. He knew the word was ambiguous. He knew almost every earlier Greek and Aramaic translation had read “radiant.” He chose the minority reading anyway. He made an interpretation, presented it as a translation, and for over a thousand years almost no one in the West could challenge it, because almost no one read Hebrew. One man’s choice became stone in Michelangelo’s hands, and the stone is still standing.
This is exactly it
I think about this story every time I work with AI.
An agent takes an ambiguous source — my request — makes dozens of interpretive choices I never see, and hands me back a finished result, confident, with every appearance of truth. And most of the time, I have no way to read the Hebrew.
Let me give you a concrete, recent example.
The ten-minute website
A few months ago I wanted my own site to publish my writing. The problem was never the will to do it, it was the time: designing, building, reviewing, publishing. The only raw material I had was the articles and a vague idea of what the site could look like.
I asked a model to handle it. And it did. It installed WordPress. Built a theme, with some direction from me along the way. Generated images. Fetched the articles, translated them, handled the basic SEO, put together the essential features. From start to finish, I spent no more than ten minutes giving instructions.
Producing an entire site — something that a few years ago was a weeks-long project — now costs ten minutes of my attention. It’s intoxicating. And this is where the Moses story starts to make sense.
What “done” hides
The site looked done. It wasn’t.
In one case, the agent built a contact form. It worked. The field was there, the button submitted, the success message appeared. Except nobody had told it where the messages should go, and it didn’t ask. It decided on its own. For a while, anyone who tried to contact me was writing into the void.
It wasn’t a mistake that jumps out at you. It was a silent decision, made underneath something that looked finished. Like choosing “horns” instead of “radiant” and moving on.
There was more. SEO, which is basic on any site, only got done after I explicitly asked for it. The images were whatever the model had on hand, and were never optimized — and a site full of heavy images is a slow site and a bad experience for whoever visits it. None of this was “wrong” in any obvious way. Everything was functional on the surface and flawed underneath.
What I realized was simple to say and hard to digest. My job had stopped being production. The model produces better and faster than I do. My job became being able to see what it had decided, and choosing what to trust.
You don’t read Hebrew
With a small site, I can still open everything and check it. With a large project, with dozens of decisions, scripts, and dependencies, I can’t anymore. At some point, the amount of things to check stops fitting inside one person’s head.
It’s the same problem the medieval reader had in front of the Vulgate. He couldn’t confirm whether “horns” was right. He didn’t read Hebrew, had no access to the source, and had in front of him a text produced by a competent authority, with every appearance of truth. All he could do was trust it.
And there’s a trap here that gets worse over time. The better the models get — the more fluent, the better-written, the better-reasoned their outputs — the more convincing they’ll also be when they’re wrong. Jerome wasn’t an amateur. He was the leading expert of his era, and that’s precisely why his choice carried so much weight and lasted so long. An error stated with competence travels further than an error stated with hesitation.
Faith is the price of scale
The temptation, once you notice this, is to respond with more control. Check everything, approve every step, let nothing pass without inspection. Except that doesn’t scale. If you have to read the Hebrew of every sentence, you might as well translate the whole book yourself.
So, sooner or later, you arrive at an uncomfortable conclusion: to scale, you have to accept a certain degree of faith in the output.
It’s not a new idea, and that’s what reassures me. Every translation has always been an act of faith by whoever reads it. Every delegation of work is faith in whoever does it. Entire institutions — science, law, the press — rest on results that almost no one verifies in full, and that we still choose to trust. We’ve never managed to verify everything. We built the world anyway.
The task, then, was never to eliminate error. It was to decide where to place trust and where to go look.
Knowing which word to check
Error is going to keep being part of this, the way it has always been part of human history. The horned Moses stood for centuries, and the world went on its way regardless. It wasn’t the end of anything. It was just a word, poorly chosen, that no one checked in time.
What changes now is the speed and scale at which these choices happen. An agent makes more invisible decisions in an afternoon than Jerome made in a lifetime.
That’s why I don’t believe the skill of the future is checking everything. It’s knowing which word to go check. Where a silent choice has big consequences, and where it doesn’t hurt at all to be wrong. Where the contact form is, and where a margin’s width is.
Moses’s face was radiant. Someone read horns. The distance between the two is all the work that’s left for us.
Sources
The story of St. Jerome’s translation and Moses’s horns is documented in various sources on art history and translation history. For an introduction: Horns of Moses (Wikipedia) and The reason why Michelangelo’s Moses has horns (Aleteia). Michelangelo’s Moses was commissioned in 1505 by Pope Julius II and is housed in the Basilica of San Pietro in Vincoli, in Rome.