I’m enjoying the discussion so far—there’s one angle I haven’t seen mentioned.
Styles of Torah learning never develop in a vacuum. Outside forces keep nudging us. The focus on “yeshivish” masechtos and the deep-dive iyun common in many yeshivos grew, at least partly, as a response to the Haskalah: we had to show that our brilliance matched the best of the secular world.
These pivots aren’t new. The printing press, writing down Torah she-be’al peh, and publishing the Shulchan Aruch all reshaped how we learn.
So where does that leave us? I’m not sure. Like Zach, I want us to keep thinking for ourselves—but we may be standing at the edge of another shift that will reshape Talmud Torah once again.
One last thought: my rebbe, Rav Moshe Stav, once said he has never seen someone studying on a screen and singing along. Torah is called a shirah—a song. Whatever tools we use, let’s ask: does the way I learn make me want to sing?
Fully agree here. I think there is potential for news ways of learning and very interested and curious to see how it all shakes out.
On the last point you made- I agree and not only that, I do think it is much more enjoyable to actually think through and work hard on something and then have it finally all fit together after struggling for awhile- that is what makes learning so geshmak. The struggle + the eventual solution creates the enjoyment. (whereas struggle alone without the resolution is super frustrating, you want to be in that middle zone where you are challenged but is possible to succeed- true for everything). So we also want the derech halimud to always create opportunities for that enjoyment.
As a parallel, when I first started using generative AI for coding it was super fun because I could do so much so quickly. But now that that is the new normal, I now actually get more enjoyment when I manually work hard on a bug that the AI can’t solve and I have to work through it and solve it myself and then when I do that correctly, it is so much more enjoyable.
And agree, this conversation has been super enjoyable for me as well. Really been so meaningfully to talk out these ideas. Thank you Yaakov for the forum for this!
YK Zach -- great article. I think its important to point out a distinction the Maharal makes in the beginning of Tiferes Yisrael. He says that when a person is learning themselves, Hashem appreciates the effort and is willing to overlook mistakes (even fairly heretical ones). But that changes when a person begins teaching Torah (and it seems like this would apply when sharing one's learning with others, as well) - there, effort alone is not enough, the standard is much more outcome-oriented; the material must be objectively correct. So while effort is a central element of talmud Torah, i think its going to too far to say that its the definition of the mitzvah.
I think your quote from Rav Kook at the end illustrates this. I don't understand what it would mean to be 'revealing new dimensions' in Torah if a person struggled through a sugya for an hour and came away with the basic understanding - what new dimension was revealed solely through the effort? Ella, the expectation seems to be that every person has some novel insight they have the capability to uncover - but that's an outcome focus, not an effort focus.
In general, I appreciated the methodology employed of addressing the question by stepping back and asking about the essential goal of talmud Torah. A different answer to that question, which yields a slightly different answer to the AI question, i think might be provided by Rav Hirsch and the Rav in Halachic Man. I think both see the goal of talmud torah to be molding the person learning, teaching him to see life/process experiences through a specific lens. The process of learning is both the way to uncover that lens (outcome focus) and also important for training oneself to adopt that lens (process focus), which is then reinforced by going out and doing the mitzvos about which one is learning when possible. For example, Shabbos is meant to reinforce a specific perspective on our role in the world. We discover what that is through learning, and teach ourselves to use that lens both by keeping Shabbos and by learning more about Shabbos. I think this is a bit more practically oriented than a pure effort focus or the idea from Rav Kook about the soul of the learner revealing new frontiers in Torah.
Relatedly, I think this is something which is going to have to shift in general education as well. When AI can do your homework faster, better, and easier than you can, i think teachers are going to really have to justify why its important for a student to personally know the material they're teaching - for example, literature or poetry as something that can serve as a mental model to process personal experiences.
Thank you so much! I think these are all really good points, and I basically agree with all of them.
I especially liked the perspective you brought from Rav Hirsch and the Rav. I do think it fits nicely with the argument that we should not learn Torha through the prism of AI tools. There are so many angles to approach this from, each with slightly different nuances, but together they really reinforce the idea that both the process, in addition to the outcome, has value. Even if we assume that AI can generate the right outcome, skipping the process means missing out on an essential part of the mitzvah, at least that is the argument I was making.
I also fully agree with the distinction you made in the Maharal between learning for oneself and teaching. The lens I was using for this article was specifically that when I sit down in the Beis Medrash to fulfill my personal chiyuv of Talmud Torah, should I be thinking for myself or using AI to help? I believe there is a a broader discussion about what the mitzvah of talmud torah is, is it learning, is it teaching, or is it both. And you are correct that labeling the essence of the mitzvah to just work on it is incorrect. I should have been more clear that what I meant is that when you are fulfilling the mitzvah when "learning", as opposed to teaching, then this the is how you fulfill it. My presentation might have de-valued the outcome too much, which if so was a mistake, when I really was trying to just show the value in the process.
Obviously, the outcome is extremely important as well. I tried to assume in this article that Ai one day will be able to get the proper outcomes, otherwise the whole discussion becomes pretty boring and feels like it’ll age quickly
I think the discussion of how we treat the outcome is more similar to the discussion of can we use AI for psak, and I linked in the footnotes a great article that Rav Wiederblank wrote on this.
As for the Rav Kook piece, I agree it doesn't fit so well with the idea even if you get something wrong there is value. But I do think I also do think it shows there is an aspect just besides the correct outcome being generated and that was the point I was trying to get across.
And this idea in Rav Kook might also connect back to how we view the outcomes that AI generate. There is a great shiur by Rav Twersky on the need for humans in the development of Torah (ran it through Sofer.Ai here: https://app.sofer.ai/transcript/c5b540f9-da28-4daf-b0bb-a323b468548a). (side point, I think it is so cool that this shiur was given almost 20 years ago, and almost two decades later gains so much relevance because realities that were not around at the time)
I think with this in mind (Rav Kook/need for human development), you can make the argument that even if you are going to work and toil in the beis medrash the same amount, but in your process you use Ai to generate ideas that you then analyze and work one level of abstraction higher, that this would not be ideal. I would make the argument that what the mathematical, algorithmic, lack of Neshama model cannot generate a "cheftzah shel torah". I did not take this approach in the article, but I think it is a good argument as well. I can see another side to that, but that is what my intuition tells me and I want to look more into this.
And agree about education as well. I think it is clear that the most efficient way to work would be to learn how to use AI tools properly. But if you just do that, you lose so much value in the process of actually working on the solution manually. Both in how it affects you as a person, but even just in the domain knowledge and ability to use the tools better. I am really curious where we will land on this question.
Thank you again for your comment. I think it is a fascinating discussion and there are so many angles to take. Agree with all the points you made, just wanted to flesh out a little more my thoughts in the context of points you made.
והנה כל הלומד תורה הוא מוציא מהכח אל הפעל את מציאות חכמתה מצד נפשו, ובודאי אינו דומה
האור המתחדש מצד חיבור התורה לנפש זו לאור הנולד מהתחברותה לנפש אחרת, ואם כן הוא
מגדיל התורה ממש בלמודו.
I do think you could understand this two ways. Is it that he is bringing a new valid חכמתה מצד נפש, which would indicate that really it needs to be that new novel and correct insight need to be thought of to create this.
Or is the focus חיבור התורה לנפש זו, the connection of the soul of the person interacting with Torah, which might indicate that even if no new correct ideas were thought up, that even so, just by this persons soul interacting with Torah brings a new light into the world.
Not sure myself. But thought would just put it out there.
It's a great article and definitely thought-provoking but I'd like to maybe raise some additional points that I've been grappling with that perhaps is a slightly different perspective on the subject. I think we have to be careful about this romantization over effort. I wrote about effort being critical for learning in my book. I'm not down-playing it. But there are also lots of sources about the focus on outcome (i.e. that understanding matters. That learning is a means as well as an ends). We also don't actually live like that the sweat is the be all and end all of Torah. You mention that there is a difference between Chipus and Yegiah. But in practice we take it further than this. We mostly don't frown upon people learning Rashi or Mishna Berura or Kehati, or listening to the Daf or learning with Nekudos or with the Mesivta. Yes there are some who push back on these tools and each person should reflect on how and when it's appropriate to make use of these, the costs and benefits of doing anything that makes learning easier. But I think most people would agree that generally speaking, we should try not to push too hard on the 'elitism' button but allow the promotion of wider access to Torah. Why should we draw the line at AI and say well that's too far? Why shouldn't it be acceptable to ask AI for creative questions on the Gemara we never thought of and then reflect on those points throughout the day. The same can be said of any number of use-cases in which AI could be used to make learning more accessible and prominent in our lives. Even asking AI for Chiddushim or writing a neat summary of a Sugya could be used to great effect. Isn't this still using our soul if we allow it to? Why should we have individual gatekeepers deciding what my Ameilus B'Torah should or shouldn't be? I think we have to expand our minds as to what is possible with AI and to also realize that the net result could still be a greater internal connection to Torah even if the methods change and the perceived effort reduces. Just a few thoughts...
Great points — thank you. Here are my thoughts, and I appreciate you helping me sharpen them.
On Romanization over effort
While my use of words like “effort” and “toil” in the article might sound like a romanticization of effort (which could be debated whether that is good or bad) I do believe the core point stands: we should be process-focused, not outcome-focused, especially when it comes to Talmud Torah (and I’d argue outside of Torah as well). Even people doing Daf Yomi say "It’s not about the daf, it’s about the yomi." I do believe we generally value spending an hour working through a seder more than someone who covers the “same material” in 15 minutes.
Does learning with the Mishnah Berurah or Kehati undermine this? I don’t think so. Even though they are later Achronim, they are Torah sources worth engaging with, you are still being osek in Torah.
It’s a separate question whether learning more primary sources represents a higher level of learning. For example, is learning the actual words of the Parsha more inherently valuable than reading your Rabbi’s derasha? That’s a real discussion, but it’s a different one.
Distinction between ArtScroll
I think there’s an important distinction between ArtScroll and AI tools that reason for you. ArtScroll was created to help those lacking certain skills access Torah, like someone who can’t read Aramaic and therefore can’t learn Gemara. It filled a skill gap, making Torah accessible to people who otherwise couldn’t engage with it.
If you want to use AI to help translate a word or clarify a phrase you don’t understand, that’s totally fine. If you even want it to find information for you, give you background on an idea, create a chart, or present information in a way that is helpful to you, that is amazing (assuming there verifiable accuracy, which is also a major point we emphasize at SoferAi - everything needs to be verifiable)
But using AI to “think” for you is different. If you’re capable of understanding complex ideas that an AI gives you, then you’re also capable of thinking through them yourself, and you should.
I’d be shocked if there’s a person who has an hour seder set aside, who can read a Gemara without ArtScroll, but would see more by using ArtScroll, that any Rabbi would tell him to use it because this would allow him to cover more ground. He’d lose his skills. And more fundamentally, you understand and connect more deeply to the things you work hard on.
The goal in the beis medrash isn’t to come up with the “best Torah.” If it was, we’d let the gedolim learn and just listen to them. The goal is to invest in the process. Skipping that by outsourcing your thinking to AI isn’t making Torah more accessible; it’s bypassing the very thing we’re meant to be doing.
Not Binary
I positioned the argument as binary, either we are thinking learners or tool users. I think you rightfully point out that there’s a middle ground, where we use reasoning AI tools as part of our own learning process. I think that’s the best argument. My pushback here, though, would be a little more technical (which I tried to avoid in the article) and less essential.
Even if AIs no longer hallucinate and we are assuming much higher level intelligence than they have now- just like humans, once they’re not quoting a source word-for-word, they’re applying their own interpretation. There’s no reason to believe that interpretation is correct, and there’s good reason to believe, given how these models are designed, that biases creep into the results.
So let’s say you ask a complex question and it gives you a complex mehalach based on three sources. You have to look up those sources anyway. You can’t accept the AI’s summary as truth. So why let a non-human, black-box system shape your initial perspective? Instead, let it gather and present relevant sources, that’s useful. And that my argument to begin with.
That’s what want our Maishiv tool to do. You can ask a complex question, and it will return sources it finds relevant, but it won’t give its own perspective. I would call that search, not reasoning (even if there’s complex reasoning behind the search results). And I’m fine with it doing sophisticated layout, walking you through which sources to see, pointing out which ones raise questions, which ones support an idea, and so on. That can be incredibly helpful. Just don’t let it think for you.
I wasn’t saying these tools cannot be helpful in our learning process, I would argue the opposite. But I strongly believe that even if they get more intelligent than they are today, we should not outsource any of our actual thinking to them.
Thanks so much for clarifying your position. I broadly agree on most of your points there. But I still think you are thinking somewhat narrowly with this idea of skills vs. Outsourced thinking.
I think highlighting your point about Artscroll is a good way of showing why and how I slightly disagree on where you put the goalposts or whether we should even have goalposts at all. You frame ArtScroll as a tool that helps people access Torah when they’re lacking specific skills, and AI as something that can cross the line by “thinking” for the learner, once it goes beyond enhancing skills or looking things up. If I understand you correctly, that's one of your main points. But I wonder if this frame of: you're either outsourcing or thinking, might miss some of the nuance in how tools like ArtScroll are actually used and therefore why AI is not too dissimilar in its application.
There are millions of legitimate ways to use an ArtScroll beyond skills building. Yes some people rely on it for translation or building skills. Others use it for structure of a Sugya. Some people skim the English just to confirm their understanding. Others could read the footnotes deeply and engage with the further discussions and questions that it raises. In some ways, ArtScroll already does quite a bit of what you call 'thinking' work, whilst still allowing a person to be learning in a way that isn't considered just cutting corners. Learning Rashi is the same thing. You could have got to the same outcome with a more effortful, thinking-full process. But Rashi did the heavy lifting for you and that's ok. There are other reasons to use Rashi even though you might simply use it to skip thinking what the Gemara means. And it's all good. I'm not Chas Vshalom trying to compare Rashi to AI. I'm just highlighting that the real-life line between “using a tool” and “outsourcing thinking” isn’t so clear-cut and in many cases, using ArtScroll actually allows people to engage more deeply with Torah ideas, not less. We therefore keep the artscroll door open nice and widely for a wide variety of people and contexts.
The same could be said for AI. It all depends on how it’s used. If someone asks AI to summarize a Tosafos and then just copies the idea without thought or uses that as a replacement for actually learning the Tosfos. Sure, that’s a weak form of learning, by most standards. I'm not denying that. But many other contexts aren't as obvious and so why should we limit the use-cases because of the remote potential (that also exists with the current tools) that we might just skip straight to the outcome? What if someone uses AI to generate three interpretations of a sugya and uses that as a springboard for their own analysis or to challenge their assumptions? Isn’t that a form of thinking too? We've always tended to use use Sefarim quite judiciously making sure to look up the sources and challenge misaligned interpretations. This is the learning process. It's the same way we ask our Chavrusa for their perspective, knowing that what they may say may be total nonsense. The same way AI has biases and misinterpretations, human written sefarim and conversations are subject to the same, if not more, fallibility. Why shouldn't AI play a greater role than your suggesting it does in how Torah is learned and is developed? We might ask AI for a Chiddush. We don't accept it blindly but we use it to provoke further reflection and study. Is this outsourced thinking? Is this not OK? Conversely, If one listens to a shiur to prepare for delivering a Shiur. Is this not outsourced thinking too? Maybe yes, maybe no. Where's the line between help and replacement? I have no idea. But you seem to be drawing it with clear boundaries. What exactly is “thinking” in Torah learning? Is it only arriving at a sevara from scratch or does it also include adapting and integrating? Most of what we do in the Beis Midrash isn’t raw thinking. It's an integrated, dynamic process. We respond to the Rishonim, we challenge the Acharonim, we test the boundaries of ideas. Isn't that too thinking even if we've used shortcuts along the way? Maybe the goal isn’t to eliminate help, but to make sure we’re still present in the process. A tool (whether ArtScroll or AI) that lets me explore Torah with more depth could actually amplify that presence, not undermine it. Hope this sharpens my own view.
I think you are making great points and definitely see your perspective and think there is a lot there.
I think the most fundamental point of agreement is how to view what the AI generates. Is that a “cheftzah shel torah?”
I would argue that it’s not. For something to be Torah, I believe it must be developed by a human. That’s the intuition I developed based on the Torah I’ve learned, the messages I’ve absorbed from my Rebbeim, and key sources like “Lo bashamayim hi”, “eilu v’eilu”, the Gemara in Shabbos 88b where Moshe explains Torah is meant for humans not angels, the Rav Kook I quoted at the end of the article, and this amazing shiur from Rav Twersky I heard explaining the role of human thought in developing Torah: https://app.sofer.ai/transcript/c5b540f9-da28-4daf-b0bb-a323b468548a.
I walked you through how I came to this conclusion to show that the way humans generate torah is very different than whatever a model generates. I used my tzelem Elokim, the God-given ability to think and form an opinion through digestion and analysis of Torah.
When a model “thinks,” it’s not actually thinking. It’s just predicting the next most likely word based on patterns. There’s no real understanding, and definitely no tzelem Elokim. It is mimicking thinking. There are many tricks to make this mimic thinking even more closely, but at the end of the day there is no actual real thought. (Unless you assume human thought works similarly, but that is some tricky hashkafic territory I believe)
When you read an ArtScroll footnote, that’s a cheftzah shel Torah. A human/humans wrote it, and you’re engaging with it and developing the Torah further. You are engaging with human analysis of Torah which is itself Torah.
But when you ask AI to generate a chiddush, all it’s doing is generating a sequence of likely words. Those words aren’t random, but they’re also not objectively Torah.
And even in the assumption that Ai will become more intelligent and fully hallucination free (big assumption), anything that is not a direct quote is an interpretation, and besides the trust issues, what makes the model’s interpretation a valid piece of Torah? There was no thought actually put into it, just bits being flipped that lead to this sequence of character being shown. So relying on this would be outsourcing your thinking to software to then develop your higher abstraction of Torah on a level that has not been thought through.
However, if you then trace the ideas back to real sources and think it through yourself, that’s valuable and that would be proper torah learning and a good use of Ai to help with you develop ideas. But I’d call that using it as search. You’re using the model to find the actual Torah you want to engage with. And because of this, I believe the tools should be designed with this use case in mind.
This is a great discussion. Your responses are great. So thank you. I once again agree wholeheartedly with almost everything you said, especially the core point: AI doesn’t produce a cheftza shel Torah because it lacks the human / Tzelem Elokim element. Torah must come through human thought, using our intellectual, emotional and spiritual faculties. I'm fully aligned with that. And perhaps the comparison to Artscroll or Rashi was a poor one.
Where I’d push a bit is on the application of your idea. AI isn’t generating Torah. An AI generated Chiddush isn't actually part of the Torah corpus. I agree. But can it support actual Torah learning more than just acting as a “search tool”? I think yes for sure.
So I wouldn’t call AI outputs Torah. They are more like a mirror that throws things back at you, provokes, sparks, challenges. But then it's on you to leverage that for actual Torah learning purposes. A person building off AI-generated content can still be learning Torah and perhaps even in a more optimal dimension, if the process is reflective and engaged. It's obviously the person not the content that needs to do the learning or thinking. But why cant that interaction with this mimicking un-human-like dummy little model take place to push the conversation further. Why can't AI be in the room to probe and to refine like nothing we've have before, even if you say it's nothing more than a really good language predictor.
Overall I agree with how you've categorised it. But I still think the applocation is broader than merely a search function. You're underselling it's possibilities. Why shouldn't it push and stretch us in ways we could never have reached otherwise?
Yes, really enjoying this discussion and think it is really cool to see how it is really refining and clarifying the points of views.
And I think this last point you are making is getting to the actual nafka minah in reality. I think you make a really good point and definitely hear the argument and think there is validity to it.
I think this is really strong argument in particular: "They are more like a mirror that throws things back at you, provokes, sparks, challenges. But then it's on you to leverage that for actual Torah learning purposes."
I do think it is a slippery slope and very hard to draw the line. It will take a lot of intellectual honesty to use it exactly in this way. (To clarify, I would not say it should not be done on the argument of it being a slippery slope, but just pointing out. I think it is a slippery slope to stop innovation on the argument of slippery slope ;) )
I think my feeling is that once you are operating on the level of this analysis, there are so many other human generated levels of Torah that you can interact with that will expand your thinking and would just be better to do that, and maybe just let Ai to find them for you the best pieces to challenge your thinking. But as I said, I do see the value in what you are proposing.
And I definitely think through this back and forward we see that in reality, the practical question where we disagree is a lot smaller than where agree - which is often the case in machlokes, and why it is always good to discuss and speak out.
Again, thank you so much, as I said I really appreciated this discussion and it really think it was a great refinement of our points and feelings.
I’m enjoying the discussion so far—there’s one angle I haven’t seen mentioned.
Styles of Torah learning never develop in a vacuum. Outside forces keep nudging us. The focus on “yeshivish” masechtos and the deep-dive iyun common in many yeshivos grew, at least partly, as a response to the Haskalah: we had to show that our brilliance matched the best of the secular world.
These pivots aren’t new. The printing press, writing down Torah she-be’al peh, and publishing the Shulchan Aruch all reshaped how we learn.
So where does that leave us? I’m not sure. Like Zach, I want us to keep thinking for ourselves—but we may be standing at the edge of another shift that will reshape Talmud Torah once again.
One last thought: my rebbe, Rav Moshe Stav, once said he has never seen someone studying on a screen and singing along. Torah is called a shirah—a song. Whatever tools we use, let’s ask: does the way I learn make me want to sing?
Fully agree here. I think there is potential for news ways of learning and very interested and curious to see how it all shakes out.
On the last point you made- I agree and not only that, I do think it is much more enjoyable to actually think through and work hard on something and then have it finally all fit together after struggling for awhile- that is what makes learning so geshmak. The struggle + the eventual solution creates the enjoyment. (whereas struggle alone without the resolution is super frustrating, you want to be in that middle zone where you are challenged but is possible to succeed- true for everything). So we also want the derech halimud to always create opportunities for that enjoyment.
As a parallel, when I first started using generative AI for coding it was super fun because I could do so much so quickly. But now that that is the new normal, I now actually get more enjoyment when I manually work hard on a bug that the AI can’t solve and I have to work through it and solve it myself and then when I do that correctly, it is so much more enjoyable.
And agree, this conversation has been super enjoyable for me as well. Really been so meaningfully to talk out these ideas. Thank you Yaakov for the forum for this!
YK Zach -- great article. I think its important to point out a distinction the Maharal makes in the beginning of Tiferes Yisrael. He says that when a person is learning themselves, Hashem appreciates the effort and is willing to overlook mistakes (even fairly heretical ones). But that changes when a person begins teaching Torah (and it seems like this would apply when sharing one's learning with others, as well) - there, effort alone is not enough, the standard is much more outcome-oriented; the material must be objectively correct. So while effort is a central element of talmud Torah, i think its going to too far to say that its the definition of the mitzvah.
I think your quote from Rav Kook at the end illustrates this. I don't understand what it would mean to be 'revealing new dimensions' in Torah if a person struggled through a sugya for an hour and came away with the basic understanding - what new dimension was revealed solely through the effort? Ella, the expectation seems to be that every person has some novel insight they have the capability to uncover - but that's an outcome focus, not an effort focus.
In general, I appreciated the methodology employed of addressing the question by stepping back and asking about the essential goal of talmud Torah. A different answer to that question, which yields a slightly different answer to the AI question, i think might be provided by Rav Hirsch and the Rav in Halachic Man. I think both see the goal of talmud torah to be molding the person learning, teaching him to see life/process experiences through a specific lens. The process of learning is both the way to uncover that lens (outcome focus) and also important for training oneself to adopt that lens (process focus), which is then reinforced by going out and doing the mitzvos about which one is learning when possible. For example, Shabbos is meant to reinforce a specific perspective on our role in the world. We discover what that is through learning, and teach ourselves to use that lens both by keeping Shabbos and by learning more about Shabbos. I think this is a bit more practically oriented than a pure effort focus or the idea from Rav Kook about the soul of the learner revealing new frontiers in Torah.
Relatedly, I think this is something which is going to have to shift in general education as well. When AI can do your homework faster, better, and easier than you can, i think teachers are going to really have to justify why its important for a student to personally know the material they're teaching - for example, literature or poetry as something that can serve as a mental model to process personal experiences.
Thank you so much! I think these are all really good points, and I basically agree with all of them.
I especially liked the perspective you brought from Rav Hirsch and the Rav. I do think it fits nicely with the argument that we should not learn Torha through the prism of AI tools. There are so many angles to approach this from, each with slightly different nuances, but together they really reinforce the idea that both the process, in addition to the outcome, has value. Even if we assume that AI can generate the right outcome, skipping the process means missing out on an essential part of the mitzvah, at least that is the argument I was making.
I also fully agree with the distinction you made in the Maharal between learning for oneself and teaching. The lens I was using for this article was specifically that when I sit down in the Beis Medrash to fulfill my personal chiyuv of Talmud Torah, should I be thinking for myself or using AI to help? I believe there is a a broader discussion about what the mitzvah of talmud torah is, is it learning, is it teaching, or is it both. And you are correct that labeling the essence of the mitzvah to just work on it is incorrect. I should have been more clear that what I meant is that when you are fulfilling the mitzvah when "learning", as opposed to teaching, then this the is how you fulfill it. My presentation might have de-valued the outcome too much, which if so was a mistake, when I really was trying to just show the value in the process.
Obviously, the outcome is extremely important as well. I tried to assume in this article that Ai one day will be able to get the proper outcomes, otherwise the whole discussion becomes pretty boring and feels like it’ll age quickly
I think the discussion of how we treat the outcome is more similar to the discussion of can we use AI for psak, and I linked in the footnotes a great article that Rav Wiederblank wrote on this.
As for the Rav Kook piece, I agree it doesn't fit so well with the idea even if you get something wrong there is value. But I do think I also do think it shows there is an aspect just besides the correct outcome being generated and that was the point I was trying to get across.
And this idea in Rav Kook might also connect back to how we view the outcomes that AI generate. There is a great shiur by Rav Twersky on the need for humans in the development of Torah (ran it through Sofer.Ai here: https://app.sofer.ai/transcript/c5b540f9-da28-4daf-b0bb-a323b468548a). (side point, I think it is so cool that this shiur was given almost 20 years ago, and almost two decades later gains so much relevance because realities that were not around at the time)
I think with this in mind (Rav Kook/need for human development), you can make the argument that even if you are going to work and toil in the beis medrash the same amount, but in your process you use Ai to generate ideas that you then analyze and work one level of abstraction higher, that this would not be ideal. I would make the argument that what the mathematical, algorithmic, lack of Neshama model cannot generate a "cheftzah shel torah". I did not take this approach in the article, but I think it is a good argument as well. I can see another side to that, but that is what my intuition tells me and I want to look more into this.
And agree about education as well. I think it is clear that the most efficient way to work would be to learn how to use AI tools properly. But if you just do that, you lose so much value in the process of actually working on the solution manually. Both in how it affects you as a person, but even just in the domain knowledge and ability to use the tools better. I am really curious where we will land on this question.
Thank you again for your comment. I think it is a fascinating discussion and there are so many angles to take. Agree with all the points you made, just wanted to flesh out a little more my thoughts in the context of points you made.
On the piece by Rav Kook, here is the language:
והנה כל הלומד תורה הוא מוציא מהכח אל הפעל את מציאות חכמתה מצד נפשו, ובודאי אינו דומה
האור המתחדש מצד חיבור התורה לנפש זו לאור הנולד מהתחברותה לנפש אחרת, ואם כן הוא
מגדיל התורה ממש בלמודו.
I do think you could understand this two ways. Is it that he is bringing a new valid חכמתה מצד נפש, which would indicate that really it needs to be that new novel and correct insight need to be thought of to create this.
Or is the focus חיבור התורה לנפש זו, the connection of the soul of the person interacting with Torah, which might indicate that even if no new correct ideas were thought up, that even so, just by this persons soul interacting with Torah brings a new light into the world.
Not sure myself. But thought would just put it out there.
It's a great article and definitely thought-provoking but I'd like to maybe raise some additional points that I've been grappling with that perhaps is a slightly different perspective on the subject. I think we have to be careful about this romantization over effort. I wrote about effort being critical for learning in my book. I'm not down-playing it. But there are also lots of sources about the focus on outcome (i.e. that understanding matters. That learning is a means as well as an ends). We also don't actually live like that the sweat is the be all and end all of Torah. You mention that there is a difference between Chipus and Yegiah. But in practice we take it further than this. We mostly don't frown upon people learning Rashi or Mishna Berura or Kehati, or listening to the Daf or learning with Nekudos or with the Mesivta. Yes there are some who push back on these tools and each person should reflect on how and when it's appropriate to make use of these, the costs and benefits of doing anything that makes learning easier. But I think most people would agree that generally speaking, we should try not to push too hard on the 'elitism' button but allow the promotion of wider access to Torah. Why should we draw the line at AI and say well that's too far? Why shouldn't it be acceptable to ask AI for creative questions on the Gemara we never thought of and then reflect on those points throughout the day. The same can be said of any number of use-cases in which AI could be used to make learning more accessible and prominent in our lives. Even asking AI for Chiddushim or writing a neat summary of a Sugya could be used to great effect. Isn't this still using our soul if we allow it to? Why should we have individual gatekeepers deciding what my Ameilus B'Torah should or shouldn't be? I think we have to expand our minds as to what is possible with AI and to also realize that the net result could still be a greater internal connection to Torah even if the methods change and the perceived effort reduces. Just a few thoughts...
Great points — thank you. Here are my thoughts, and I appreciate you helping me sharpen them.
On Romanization over effort
While my use of words like “effort” and “toil” in the article might sound like a romanticization of effort (which could be debated whether that is good or bad) I do believe the core point stands: we should be process-focused, not outcome-focused, especially when it comes to Talmud Torah (and I’d argue outside of Torah as well). Even people doing Daf Yomi say "It’s not about the daf, it’s about the yomi." I do believe we generally value spending an hour working through a seder more than someone who covers the “same material” in 15 minutes.
Does learning with the Mishnah Berurah or Kehati undermine this? I don’t think so. Even though they are later Achronim, they are Torah sources worth engaging with, you are still being osek in Torah.
It’s a separate question whether learning more primary sources represents a higher level of learning. For example, is learning the actual words of the Parsha more inherently valuable than reading your Rabbi’s derasha? That’s a real discussion, but it’s a different one.
Distinction between ArtScroll
I think there’s an important distinction between ArtScroll and AI tools that reason for you. ArtScroll was created to help those lacking certain skills access Torah, like someone who can’t read Aramaic and therefore can’t learn Gemara. It filled a skill gap, making Torah accessible to people who otherwise couldn’t engage with it.
If you want to use AI to help translate a word or clarify a phrase you don’t understand, that’s totally fine. If you even want it to find information for you, give you background on an idea, create a chart, or present information in a way that is helpful to you, that is amazing (assuming there verifiable accuracy, which is also a major point we emphasize at SoferAi - everything needs to be verifiable)
But using AI to “think” for you is different. If you’re capable of understanding complex ideas that an AI gives you, then you’re also capable of thinking through them yourself, and you should.
I’d be shocked if there’s a person who has an hour seder set aside, who can read a Gemara without ArtScroll, but would see more by using ArtScroll, that any Rabbi would tell him to use it because this would allow him to cover more ground. He’d lose his skills. And more fundamentally, you understand and connect more deeply to the things you work hard on.
The goal in the beis medrash isn’t to come up with the “best Torah.” If it was, we’d let the gedolim learn and just listen to them. The goal is to invest in the process. Skipping that by outsourcing your thinking to AI isn’t making Torah more accessible; it’s bypassing the very thing we’re meant to be doing.
Not Binary
I positioned the argument as binary, either we are thinking learners or tool users. I think you rightfully point out that there’s a middle ground, where we use reasoning AI tools as part of our own learning process. I think that’s the best argument. My pushback here, though, would be a little more technical (which I tried to avoid in the article) and less essential.
Even if AIs no longer hallucinate and we are assuming much higher level intelligence than they have now- just like humans, once they’re not quoting a source word-for-word, they’re applying their own interpretation. There’s no reason to believe that interpretation is correct, and there’s good reason to believe, given how these models are designed, that biases creep into the results.
So let’s say you ask a complex question and it gives you a complex mehalach based on three sources. You have to look up those sources anyway. You can’t accept the AI’s summary as truth. So why let a non-human, black-box system shape your initial perspective? Instead, let it gather and present relevant sources, that’s useful. And that my argument to begin with.
That’s what want our Maishiv tool to do. You can ask a complex question, and it will return sources it finds relevant, but it won’t give its own perspective. I would call that search, not reasoning (even if there’s complex reasoning behind the search results). And I’m fine with it doing sophisticated layout, walking you through which sources to see, pointing out which ones raise questions, which ones support an idea, and so on. That can be incredibly helpful. Just don’t let it think for you.
I wasn’t saying these tools cannot be helpful in our learning process, I would argue the opposite. But I strongly believe that even if they get more intelligent than they are today, we should not outsource any of our actual thinking to them.
Thanks so much for clarifying your position. I broadly agree on most of your points there. But I still think you are thinking somewhat narrowly with this idea of skills vs. Outsourced thinking.
I think highlighting your point about Artscroll is a good way of showing why and how I slightly disagree on where you put the goalposts or whether we should even have goalposts at all. You frame ArtScroll as a tool that helps people access Torah when they’re lacking specific skills, and AI as something that can cross the line by “thinking” for the learner, once it goes beyond enhancing skills or looking things up. If I understand you correctly, that's one of your main points. But I wonder if this frame of: you're either outsourcing or thinking, might miss some of the nuance in how tools like ArtScroll are actually used and therefore why AI is not too dissimilar in its application.
There are millions of legitimate ways to use an ArtScroll beyond skills building. Yes some people rely on it for translation or building skills. Others use it for structure of a Sugya. Some people skim the English just to confirm their understanding. Others could read the footnotes deeply and engage with the further discussions and questions that it raises. In some ways, ArtScroll already does quite a bit of what you call 'thinking' work, whilst still allowing a person to be learning in a way that isn't considered just cutting corners. Learning Rashi is the same thing. You could have got to the same outcome with a more effortful, thinking-full process. But Rashi did the heavy lifting for you and that's ok. There are other reasons to use Rashi even though you might simply use it to skip thinking what the Gemara means. And it's all good. I'm not Chas Vshalom trying to compare Rashi to AI. I'm just highlighting that the real-life line between “using a tool” and “outsourcing thinking” isn’t so clear-cut and in many cases, using ArtScroll actually allows people to engage more deeply with Torah ideas, not less. We therefore keep the artscroll door open nice and widely for a wide variety of people and contexts.
The same could be said for AI. It all depends on how it’s used. If someone asks AI to summarize a Tosafos and then just copies the idea without thought or uses that as a replacement for actually learning the Tosfos. Sure, that’s a weak form of learning, by most standards. I'm not denying that. But many other contexts aren't as obvious and so why should we limit the use-cases because of the remote potential (that also exists with the current tools) that we might just skip straight to the outcome? What if someone uses AI to generate three interpretations of a sugya and uses that as a springboard for their own analysis or to challenge their assumptions? Isn’t that a form of thinking too? We've always tended to use use Sefarim quite judiciously making sure to look up the sources and challenge misaligned interpretations. This is the learning process. It's the same way we ask our Chavrusa for their perspective, knowing that what they may say may be total nonsense. The same way AI has biases and misinterpretations, human written sefarim and conversations are subject to the same, if not more, fallibility. Why shouldn't AI play a greater role than your suggesting it does in how Torah is learned and is developed? We might ask AI for a Chiddush. We don't accept it blindly but we use it to provoke further reflection and study. Is this outsourced thinking? Is this not OK? Conversely, If one listens to a shiur to prepare for delivering a Shiur. Is this not outsourced thinking too? Maybe yes, maybe no. Where's the line between help and replacement? I have no idea. But you seem to be drawing it with clear boundaries. What exactly is “thinking” in Torah learning? Is it only arriving at a sevara from scratch or does it also include adapting and integrating? Most of what we do in the Beis Midrash isn’t raw thinking. It's an integrated, dynamic process. We respond to the Rishonim, we challenge the Acharonim, we test the boundaries of ideas. Isn't that too thinking even if we've used shortcuts along the way? Maybe the goal isn’t to eliminate help, but to make sure we’re still present in the process. A tool (whether ArtScroll or AI) that lets me explore Torah with more depth could actually amplify that presence, not undermine it. Hope this sharpens my own view.
I think you are making great points and definitely see your perspective and think there is a lot there.
I think the most fundamental point of agreement is how to view what the AI generates. Is that a “cheftzah shel torah?”
I would argue that it’s not. For something to be Torah, I believe it must be developed by a human. That’s the intuition I developed based on the Torah I’ve learned, the messages I’ve absorbed from my Rebbeim, and key sources like “Lo bashamayim hi”, “eilu v’eilu”, the Gemara in Shabbos 88b where Moshe explains Torah is meant for humans not angels, the Rav Kook I quoted at the end of the article, and this amazing shiur from Rav Twersky I heard explaining the role of human thought in developing Torah: https://app.sofer.ai/transcript/c5b540f9-da28-4daf-b0bb-a323b468548a.
I walked you through how I came to this conclusion to show that the way humans generate torah is very different than whatever a model generates. I used my tzelem Elokim, the God-given ability to think and form an opinion through digestion and analysis of Torah.
When a model “thinks,” it’s not actually thinking. It’s just predicting the next most likely word based on patterns. There’s no real understanding, and definitely no tzelem Elokim. It is mimicking thinking. There are many tricks to make this mimic thinking even more closely, but at the end of the day there is no actual real thought. (Unless you assume human thought works similarly, but that is some tricky hashkafic territory I believe)
When you read an ArtScroll footnote, that’s a cheftzah shel Torah. A human/humans wrote it, and you’re engaging with it and developing the Torah further. You are engaging with human analysis of Torah which is itself Torah.
But when you ask AI to generate a chiddush, all it’s doing is generating a sequence of likely words. Those words aren’t random, but they’re also not objectively Torah.
And even in the assumption that Ai will become more intelligent and fully hallucination free (big assumption), anything that is not a direct quote is an interpretation, and besides the trust issues, what makes the model’s interpretation a valid piece of Torah? There was no thought actually put into it, just bits being flipped that lead to this sequence of character being shown. So relying on this would be outsourcing your thinking to software to then develop your higher abstraction of Torah on a level that has not been thought through.
However, if you then trace the ideas back to real sources and think it through yourself, that’s valuable and that would be proper torah learning and a good use of Ai to help with you develop ideas. But I’d call that using it as search. You’re using the model to find the actual Torah you want to engage with. And because of this, I believe the tools should be designed with this use case in mind.
This is a great discussion. Your responses are great. So thank you. I once again agree wholeheartedly with almost everything you said, especially the core point: AI doesn’t produce a cheftza shel Torah because it lacks the human / Tzelem Elokim element. Torah must come through human thought, using our intellectual, emotional and spiritual faculties. I'm fully aligned with that. And perhaps the comparison to Artscroll or Rashi was a poor one.
Where I’d push a bit is on the application of your idea. AI isn’t generating Torah. An AI generated Chiddush isn't actually part of the Torah corpus. I agree. But can it support actual Torah learning more than just acting as a “search tool”? I think yes for sure.
So I wouldn’t call AI outputs Torah. They are more like a mirror that throws things back at you, provokes, sparks, challenges. But then it's on you to leverage that for actual Torah learning purposes. A person building off AI-generated content can still be learning Torah and perhaps even in a more optimal dimension, if the process is reflective and engaged. It's obviously the person not the content that needs to do the learning or thinking. But why cant that interaction with this mimicking un-human-like dummy little model take place to push the conversation further. Why can't AI be in the room to probe and to refine like nothing we've have before, even if you say it's nothing more than a really good language predictor.
Overall I agree with how you've categorised it. But I still think the applocation is broader than merely a search function. You're underselling it's possibilities. Why shouldn't it push and stretch us in ways we could never have reached otherwise?
Yes, really enjoying this discussion and think it is really cool to see how it is really refining and clarifying the points of views.
And I think this last point you are making is getting to the actual nafka minah in reality. I think you make a really good point and definitely hear the argument and think there is validity to it.
I think this is really strong argument in particular: "They are more like a mirror that throws things back at you, provokes, sparks, challenges. But then it's on you to leverage that for actual Torah learning purposes."
I do think it is a slippery slope and very hard to draw the line. It will take a lot of intellectual honesty to use it exactly in this way. (To clarify, I would not say it should not be done on the argument of it being a slippery slope, but just pointing out. I think it is a slippery slope to stop innovation on the argument of slippery slope ;) )
I think my feeling is that once you are operating on the level of this analysis, there are so many other human generated levels of Torah that you can interact with that will expand your thinking and would just be better to do that, and maybe just let Ai to find them for you the best pieces to challenge your thinking. But as I said, I do see the value in what you are proposing.
And I definitely think through this back and forward we see that in reality, the practical question where we disagree is a lot smaller than where agree - which is often the case in machlokes, and why it is always good to discuss and speak out.
Again, thank you so much, as I said I really appreciated this discussion and it really think it was a great refinement of our points and feelings.
Hav e a good shabbas and yuntif!