Translations are an important component of the Sefaria Library, making Jewish texts — the collective inheritance of the Jewish people — more discoverable and accessible.

But translation isn’t a perfect science. Reconciling two different languages, each with their own subtle nuances, history, and cultural connotations, has its challenges. “Translation is always a mediation,” says Sefaria Translation and Research Specialist Francis Nataf. “No translation ever reaches 100% accuracy, but you always want to come as close to 100% as possible.”

So when he was brought on to experiment with using AI to assist in the translation process, he had his doubts. Was the technology up to the task of interpreting our sacred texts?

Rav Nataf worked with a team to test different tools and various texts for two years before bringing a full AI-assisted translation to completion: the Sefaria Kli Yakar. Released in February 2026, it is the first-ever comprehensive English translation of Kli Yakar, the early-17th-century Torah commentary by Rabbi Shlomo Ephraim of Luntschitz.

Initially generated by Claude 3.7, the translation underwent a six-month review process, during which Rav Nataf compared every word to the original Hebrew, retranslated inaccurate words or passages, and edited to improve stylistic consistency.

Below, we take you behind the scenes of that process — from skepticism to careful confidence, from early experimentation to finished product. In this conversation, Rav Nataf discusses what kinds of Jewish texts are best suited for AI translation, concerns about machine involvement in translation, and why he thinks others should give it a whirl anyway.

I’ve been part of this effort since way before we started working on Kli Yakar. To be honest, when the idea of bringing AI into Sefaria translations first got floated, I wasn’t a fan. The technology, when we first got it, was pretty bad. I don’t think anybody really knows how AI technology will develop, but two or three years ago I thought it did a horrible job of translating. Back then, it would have taken me just as long to edit an AI translation as it would have to do the work from scratch. But over time the technology got much better, and we also got better at identifying which texts would be good candidates for this kind of project.

First and foremost, my mandate was to correct mistakes. We always want to come as close to 100 percent accuracy as possible, although no translation ever reaches 100 percent. I was looking for inaccuracies, and where things were off, I would retranslate — usually small pieces. It was 98 percent there, and my job was correcting that remaining two percent. The work took significantly less time than translating from scratch. Others in the field are talking about AI cutting down their time 50 to 75 percent. In the end, if you put this Kli Yakar translation in the same bucket as completely human-made translations, it’s very good.

Editing is editing. There were certain things I was looking out for, although I wouldn’t say that is specific to AI; all authors or translators have certain strengths and weaknesses. As an editor, it’s your job to pick up on those and look for them.

In the case of this AI translation, two main weaknesses became clear. First, the AI model missed a lot of the Hebrew word play. It may understand one or two meanings of a Hebrew word, but it would often miss additional nuance, like plays off of other rabbinic or biblical phrases, and it was not always able to follow some of the more complicated uses of gematria [interpretation based on numerical values of Hebrew letters].

Second, the model wasn’t stylistically consistent in its output. Large language models (LLMs) lack intrinsic memory, so each operation is a new operation; in this translation, the AI approached things differently from one paragraph to another. Even though style wasn’t my primary mandate, I tried to keep the variation within reason so that the work would be more cohesive.

We tried a few different models from different companies, and we found that Claude was head and shoulders above the others for our purposes. It was impressive! When we started experimenting with AI a couple years ago, it was maybe 80 percent accurate with its translations. By the time we tried Kli Yakar using Claude 3.7, it was around 98 percent accurate. When I saw that, I said, “Let’s go, let’s do it.”

When working with a developing technology, you’ve got to ask whether you should be doing the work now or waiting for it to become better. Three years ago, I think it would’ve been foolish to do this project. Two years from now, I can only imagine how much less cumbersome this will be. So, do we wait? It’s a tricky balance and an interesting question. For me, it’s exciting to be discussing these questions and seeing this technology being used for Torah in a very serious and direct way.

The short answer is that simpler texts will yield a better translation. Also, the better known the referenced texts are, the better the work will be. Take the Torah, for example. AI has lots of resources with which to understand the Torah, so working on a Torah commentary gives you a better chance for an accurate translation.

We did some experiments that didn’t go well. For example, we tried the Mishnah Berurah, a commentary on the Shulkhan Arukh. At Sefaria, we prioritize functionality around interlinking of texts, but what we found was that a text with lots of references to other texts was too complicated for a machine to understand and translate. The more strings you have, like in halakhic texts, the harder it is for a machine to work with. It also had a hard time with esoteric texts and with Aramaic.

Kli Yakar has relatively linear thinking, as far as Torah commentaries go. In contrast, you have some Chasidic texts for example that fly from one topic to another, which would be much trickier for the AI to follow. In this commentary, the author tends to stay with one idea. He does reference some outside texts, like the Midrash and Gemara, although the Gemara is pretty accessible for AI at this point. There are many other texts that have these qualities, many of which would be good candidates.

The AI tools we have at this point in time are not ideal writing instruments for professional writers. It’s the small things, details I’d be embarrassed to have in my own translation. This translation isn’t a fully AI-produced artifact, though — it’s a collaboration between myself and the LLM. That crucial difference is why I feel very confident about the content of this edition. In fact, I’d say this work is extremely accurate, better than the vast majority of translations out there.

I’m excited about putting out this translation. First, because it’s something that enhances the field of translation by bringing AI into the work in a conscious way. Second, because it means we can make more translations available across the board — not just in the Sefaria Library, but in general. I hope this project serves as a model for others. I see this as a pioneering project. Different platforms, including Sefaria, are going to have to decide what sort of interaction makes sense going forwards.

I’d encourage people to try it out. I know people are worried about hallucinations, when an AI injects invented text into something it produced. In terms of Torah, that’s a serious concern. We can’t allow hallucinations to go into the content. If you’re working with a solid AI assistant, hallucinations are fairly easy to check for as long as you know what you’re doing. My mandate is to make sure the translation is as close to 100 percent accuracy as possible.

Of course, translators are also worried about their jobs, as are many other people. That’s a legitimate fear. Companies can now get the same product for a lower cost, and that’s a problem everywhere. When it comes to a professional translation, though, there’s no way that AI can give you a final product. A human eye is still required.

There’s an interesting parallel to consider: A Sefer Torah has to be written by a human, not a machine. Although, then you have Rabbi Yitzchak Abadi [a prominent 20th–21st-century posek] who disagreed with that and found a way to use technology to write a Sefer Torah. In any case, translation is a different story. There’s no halakhah that translation has to be handmade. Bottom line, we’re interested in the meaning of the words. And if a tool, like a hearing aid or a screen reader, can do that, then I don’t think it's problematic to look at other forms of mediation. The thing is, translation is always a mediation. Whether that mediation is human or not, ultimately the processes are comparable — as long as you have the accuracy, and accuracy is the concern at Sefaria in terms of translation.

I don’t know what we’re going to do in the future, but I think it’s commendable that we took this project very seriously and tried to figure out a responsible way to work with advancing tools. I feel really good about this as our first product, and I’m happy to put my name on it, because I think it will please people who are rightly concerned about the sanctity of Torah. We did a good job with this translation, including from a Torah perspective and a halakhic perspective.

People are already using AI to translate and for other Jewish learning purposes. If we don’t use it, others will — and probably not as well, or as carefully, as Sefaria. So I think we’re doing the right thing.

From when I was in college or high school, I knew that computers could do powerful things and were going to do more and more things with time. In that sense, I don’t think I would be so surprised. But also 40 years ago I didn’t think I’d be a translator or a writer. Life has its interesting twists, and this is an interesting place to end up.

Technology is always pushing the envelope in new ways, and that’s been true for hundreds of years. AI raises massive questions we need to deal with. My hope is we should deal with it well and use it well.