AI and Torah Lishmah

If Jewish educators know one thing, they know that Torah study isn’t going to make you rich. For most students, it will have little impact on their economic outcomes at all. When it comes to AI, this gives Jewish educators a distinct advantage.

When AI entered the general studies classroom, two concerns immediately popped up. The first one was about plagiarism and the need to reconfigure assignments. However, behind this concern was a deeper question: How do you teach the skills that students will need for the economy of the future? I sympathize with these concerns; they will surely take time to resolve, and many painful mistakes will be made along the way.

But the world by and large does not care how much Torah you know. The point of Torah learning is learning. Whereas their general studies counterparts must optimize their lessons so that their students can compete in an AI-boosted marketplace, Jewish educators have the freedom to optimize the experience of Torah study itself. Unlike most educators, Jewish educators can afford to ask: How can AI improve Torah lishmah, Torah study for its own sake? How can AI enhance learning when learning is an end in itself?

The answer comes down to fluency. Jewish education is about imparting particular pieces of information, but the real prize is giving students the critical mass of skills and confidence and a roadmap that they need to navigate the sources directly, allowing them to take ownership through exploration. This ability—knowing how to learn and what to learn—involves a combination of textual skills and a sense of the Torah corpus itself.

Thanks to translation and digitization, that critical mass is smaller than it used to be. To study a page of Talmud by oneself or in havruta once required a tremendous amount of knowledge. Today, with the entire Talmud elucidated and translated, it involves much less. Digitization has also made it much easier to navigate the corpus. Knowing which texts were available and where to find them was once a skill unto itself, and without a well-stocked beit midrash most people (and almost all women) had no chance to access any but the most important. Today, this has all changed. Not only is every major Jewish text freely available online (along with translations), but many digital Torah databases come pre-loaded with connections to commentaries, parallel texts and references. These tools help learners build their internal roadmaps of what Torah contains. They encourage people to explore.

Can AI lower the bar to entry for Torah lishmah even further? Well, not yet. Large language models like ChatGPT do indeed know quite a bit about Judaism and Jewish law, but their training data is not sufficiently focused on Hebrew or Torah; they can’t teach you how to read the Talmud, their Q&A interfaces don’t encourage wide exploration, and they don’t (yet) enhance the classroom.

AI Developments on the Horizon

But AI is moving fast, and these problems are solvable. Decades of building Torah databases have yielded exactly the data set needed to design an AI that could, indeed, be a tutor, classroom aid and textual guide all wrapped up in one. More than being convenient, these tools might imbue learners with the confidence to ask hard, direct questions.

All of these developments are stops on the way to a grander prize: the development of a conversational AI that embodies the entirety of Torah. Such a model, presented as a “person,” would have the ability to speak on behalf of the Jewish tradition as a whole, or whatever part of it happened to be of greatest interest. Imagine an AI embodiment of Deuteronomy, or Rabbi Meir, or Glikl of Hameln—and imagine having them speak to one another. AI as Torah personified would quickly make its way into the classroom and could be a valuable third partner in any havruta. Such an AI might spur a great deal of textual exploration and high-level understanding of different parts of the tradition. Add in analysis and use of sound and images, and the possibilities are almost limitless.

I have confidence that these AIs will be built; Jews have long been at the forefront of digitized religious learning, and I expect this to continue. The AIs I am describing have the ability to open up Torah lishmah to people who might otherwise never have been interested, and they will create new possibilities for those already invested.

Three Guiding Principles

That said, the devil is in the details. It is important to implement these AIs in a way that emphasizes that Torah study is not like other forms of study, does not leave users overly reliant on the AI as the font of all Torah knowledge, and inspires the creation of new ideas. To that end, here are some specific suggestions for how an AI programmed to enhance Torah lishmah might be productively designed.

Better to say too little than too much. The current cutting-edge AIs all suffer from “hallucinations,” which means they make up false information. Given that Torah study involves thinking carefully about each and every word and phrase, the stakes for making up new Torah are unusually high; in addition, the things not said are often as interesting as the things said. To that end, Torah AIs must be scrupulous not to fabricate information and to make the user aware of the limits of their knowledge. This also gives space for users to develop their own ideas without falsely believing that someone else has already had that same thought.

Center the sources. Some AIs show their work; others do not. A key concern for a Torah AI is that its answers may be so abstract that the source texts themselves seem unnecessary. This could alienate learners from the texts and leave them feeling like they know how to navigate the AI but not its underlying sources. To counter this, Torah AIs should be designed to provide a mixture of abstraction and direct textual analysis, so that learners are still regularly being exposed to the source material. The AIs must also be able to answer complex philological and grammatical questions, should the user be interested.

Make all Jewish knowledge freely accessible. A major problem of AI development is bias; skewed data sets can lead to skewed results. For Jewish wisdom, this means that far too much contemporary Jewish studies research is locked away in places where AIs cannot find them. If AI becomes the main entry point for Torah knowledge, it is crucial that this portal not have arbitrary gaps in its knowledge. To that end, researchers should think about publishing in places where their ideas can be consumed by AIs, and existing publishers should consider giving AIs access to their publications. For users, this knowledge will also hint at the kinds of ideas that no AI could assist them in developing. Some users may become intrigued and want to study more.

AI is like water: It is seeping into everything. Jewish educators, with their unusual religious incentives, have a particular opportunity to develop AIs that inspires students and encourages them to expand Torah further.

Return to the issue home page:

AI and Tech

Fall 2023