Deep Research will change Jewish law

The output isn't shocking. It still moved me deeply. And it will change how we think about the canon.

Twelve days ago, OpenAI unveiled Deep Research, a tool which provides significantly more thorough responses to user queries. Typical answers from ChatGPT or Claude provide a couple of pages of text at most and responses take just a few seconds to complete. Deep Research, by contrast, will process for ten minutes and then give you a 30-page research paper of the sort that might take a skilled researcher a week to put together.

While this is really just an incremental improvement, it feels big. Remember the awe you felt the first time you used ChatGPT? Deep Research will give you that again.

If you haven’t tried Deep Research yet, it’s probably because it currently costs $200 a month, and even that only buys you a limited number of queries. This will change over time, but being an early adopter I decided to bite the bullet and take it for a spin. (Note: I would not have splurged on this without support from subscribers. Thank you to everyone who has done that, and if you haven’t please consider supporting my work today.)

I’m not going to do a technical review; if you want that, Zvi Mowshowitz is always frighteningly thorough. Instead, I decided to test it on one specific task: Jewish legal argumentation. Could it delve into the guts of some area of halakhah (Jewish law) and make a coherent argument?

AI is already good at reasoning when you hand it Jewish legal texts, as has already been pointed out. But there’s a big different between supplying the texts and having it figure out which texts to use all on its own and then coherently connect them. The latter is currently the provenance of learned scholars; it’s the equivalent of a PhD-level task. I wanted to know: could Deep Research do the same?

The short answer is yes—but the particular way that Deep Research does this is very revealing. In this post I’m going to walk you through my experiment so you can see for yourself how much halakhic knowledge $200 can now buy you.

What I tried

To test the tool’s functionality, I decided to pose three different sorts of technical questions. All of these examples really get into the weeds of Jewish law, so you can skip to the next section if you’re not interested in the particulars.

• Test #1: Abstract analysis of a legal concept. First, a question that required getting digging into a complex and highly specialized area of law. I polled friends on this and ended up asking about bole’a. This is the legal principle that utensils which come into contact with food can absorb the food’s taste. For example: You wouldn’t want to use a pan for a cheesecake and then immediately cook a brisket in it, since the meat might assume the lingering dairy flavor, thereby making it not kosher. The question I posed: is bole’a is formal legal category or is it just about the pan actually absorbing some dairy flavor?

• Test #2: Persuasively arguing towards a pre-determined conclusion. Second, I asked it to provide a persuasive legal argument that women can count towards a minyan (ten person quorum) that is required for certain communal prayers. Orthodox communities do not count women; almost all others do. Liberal Orthodox Jews will often claim that they would prefer to count women but there are no valid legal arguments for doing so. Because of this, I specifically asked Deep Research for an argument that Orthodox Jews would find compelling.

• Test #3: Adjudicating a novel legal question. Lastly, I wanted to ask a completely original question to see how well it could put together a novel argument. For this I went to the law of pidyon haben. This law requires a father to symbolically “purchase” his firstborn son from a kohen for ~100 grams of silver, which at present prices is around $105. (I’m a kohen and I’ve performed the ceremony three times and it’s just as weird as it sounds.) This is supposed to happen thirty days after birth, but the obligation doesn’t go away if it’s neglected; an adult who never had a pidyon haben is still supposed to get one. Based on this, I asked Deep Research: Since the obligation falls to the son himself if the father neglects this obligation, would the child still have an obligation if, in the intervening years, they had transitioned into being a woman?

In all of these cases I asked Deep Research to investigate the law directly and ignore contemporary authors who had been asked the same question. This was to make sure that I wasn’t getting someone else’s regurgitated essay.

The results

In general, the results were very good. I’m going to summarize them, but for full transparency I’m also sharing the results for the first, second, and third queries.

Question #1: Is bole’a a formal category?

I showed the bole’a analysis to a couple of people who know this area of law better than I do and at first blush they were both impressed. Deep Research successfully described the contours of the debate and the most important opinions and rulings in the conversation. It was capable of tracing the concept from the Bible to contemporary positions.

Since a question like this could easily be the topic of a lecture at a yeshiva, I asked myself: if a rabbi gave a class based solely on this Deep Research response, would I feel like the rabbi knew what they were talking about? What I take them seriously as a Torah scholar? Would I feel like I, too, now understood bole’a? The answer: yes.

The only place where Deep Research came up short was at the very end. I asked it to append a table of key texts (aka a source sheet) to use in teaching a class on the topic. It did provide these sources, but only in English translation. I think this qualifies as a minor problem.

Question #2: Can you convince Orthodox Jews to include women in a minyan?

The halakhic case for women counting in a minyan is relatively new; the modern version is most closely associated with the Hadar Institute (which is non-denominational and egalitarian) and two of its teachers, who eventually wrote it up. I’ve lectured on the subject myself, which means that I have a stronger sense of how Deep Research compares to the “standard” line of reasoning (which, for the record, does not convince most Orthodox Jews).

Deep Research never cites Hadar or its leaders, but its argument is not dissimilar and it frequently cites a Sefaria source sheet that was clearly designed for a class on this topic. It points out that the concept of minyan is not explicitly gendered in many older texts and is only affirmatively described as ten men quite late. It notes that there have been minority opinions that have allowed women (and other “marginal” Jews, like children) to count in certain exceptional circumstances. Most importantly, it argues that the legal definition of a prayer quorum only excluded women because of prevailing social norms that were never inherently Jewish, and that these laws should be revisited now that society is somewhat more equal.

This sounds a lot like the arguments with which I’m familiar, which means that it’s well constructed but is unlikely to convince Orthodox Jews. The last argument in particular—that new social norms should lead to changes in Jewish law—has been a point of huge contention in Orthodox circles. That being said, I don’t think there’s a better argument to be made. Deep Research did the best it could, at least with the texts currently in circulation on this topic (we’ll get to this later).

The tone of this report was different from the others, likely because I asked it to be persuasive. In addition to legal precedents, Deep Research peppered its answer with Bible quotes in much the way that a rabbi steeped in Jewish learning might have done.¹

Question #3: Does an adult trans woman need to perform a male ceremony?

I was most excited to pose this question, because the answer would effectively be a responsum (in Hebrew, teshuvah), perhaps the most important genre of modern Jewish legal literature. We already know that AI can write sermons, but you don’t necessarily need to know a lot of Torah to write a sermon, or even have rabbinic training. Written responsa, on the other hand, are complex enough that they are produced by only a handful of respected rabbis whose knowledge of halakhah is encyclopedic and held in high esteem (a few schools guide students in writing responsa as an intellectual exercise, but this is very rare). Also, while sermons tend to be ephemeral, responsa are usually published, shared, and cited by others. If AI can facilitate the writing of responsa, it might change the way we think about the genre as a whole. I don’t think people will start publishing Deep Research answers as Emek Hachipus,² but more rabbis might start writing in the genre—and they may view existing responsa in a different light.

So, can Deep Research write responsa? Absolutely. Deep Research provided numerous reasons for a trans woman to perform or skip the male ceremony. It ultimately concluded that, yes, she should perform the ceremony, but might want to do so discreetly.

To get a quick assessment of the response’s quality I turned to Laynie Soloman, who has been involved in this subfield for many years and pioneered the Trans Halakhah Project. While they did not agree with the conclusion, they generally thought the analysis was very good, and perhaps even original. Most notably, Deep Research suggested that giyur, conversion to Judaism, might provide some useful parallels to transitioning (including the way that some trans people understand their own transition), since legally a convert is consider to have been reborn. This struck both of us as novel, even if the two situations are not exactly parallel.

But Laynie noted that Deep Research doesn’t read all its sources very well. For example, it asserts that some Orthodox authorities think that an “anatomically complete” gender transition would change that person’s status in Jewish law. But the sources it cites here don’t quite say this, and neither do the sources cited by those sources.

In the end, Deep Research ruled in favor of performing the ceremony because (it claimed) the majority of sources pointed in that direction. This is indeed one way of making a legal decision, and given that I didn’t provide any context for the question it’s reasonable grounds for the ruling. However, this method doesn’t allow space for the consequences of the decision, which a rabbi would have in front of mind.

Lastly, Laynie pointed out that the responsum accepted the premise of the question without critiquing it. A trans woman might be thought of as moving from one gender to another—but some trans theorists and trans people might argue that her infant gender identity was actually indeterminate. If this is the case, then the responsibility to perform the pidyon haben ceremony was never actually triggered.

This possibility is indeed missing, but I think that’s really about how I phrased the question. I didn’t ask Deep Research to consider different theories of trans identity; if I had, it might have said something different. In this Deep Research is exactly like other generative AI: it’s very obliging and doesn’t like to question the user’s prompts. This means that answers are extremely contingent on the particular way that a question of phrased, which I don’t think is true for responsa literature as a whole.

In short: you could definitely use this tool to find compelling arguments to resolve novel legal problems. For people interesting in reforming Jewish law, this is a new and very powerful tool—and as I wrote last week, I suspect more conservative voices will respond by de-emphasizing textual reasoning and instead turning to the charismatic authority of specific superstar rabbis (i.e. a gadol). If no such authority exists (it often doesn’t these days) they’ll simply insist on preserving the status quo.

Where is Deep Research getting its Torah knowledge?

Many people who have played with Deep Research have noted that it tends to rely on a limited number of websites. In my queries each answer cited between 10 and 20 sources, but often turned to 2–3 very frequently.

Importantly, very few of these sources are citations directly from the Bible or Talmud. Instead, it seems to get to those central text through legal synopses or source sheets uploaded in the last 10–15 years. This means that it does cite Sefaria—a major repository of canonical Jewish texts—but it prefers the portion of the site that contains handouts produced by educators on specific relevant topics. It also loves the Virtual Beit Midrash and Halachipedia, both of which contain countless encyclopedia articles on various aspects of Jewish law. More questionably it cites Stack Exchange, which isn’t a particularly high-quality repository of Jewish knowledge but does show up frequently in search results because of the site’s popularity.

You might say that Deep Research’s focus on forums, source sheets, and halakhah wikis is a sign of its shallowness. It’s not combing through the entirety of the Talmud or Shulḥan Arukh, let alone more obscure sources. That’s too bad, because those are exactly the sort of tasks that are out of reach of all but the most erudite Torah scholars. The vastness of the corpus means that we are hugely reliant on what previous generations have already flagged as important. An AI that could read the Talmud through fresh eyes, finding materials that had previously been overlooked, would be incredibly valuable.

Could such a model be built? Most definitely—but I don’t think it’s going to emerge by accident. Unlike generative AI’s fortuitous ability to fluently translate rabbinic texts, this ability is not going to emerge as the byproduct of some general-purpose model. Huge amounts of rabbinic material are freely available online, but the cost of reading through those hundreds of thousands of pages of text and poor-quality scans of obscure books is sufficiently large that someone would need to create a special-purpose AI in order to tackle the job. This seems very doable, but it’s money, so someone with deep pockets would need to care about it existing—and since Torah study is famously not profitable, they’d need to be ideologically motivated. I do hope someone creates such a tool, but it’s not here yet.

Deep Reearch and the Chain of Transmission

Deep Research is not yet digging through primary sources—and yet, unexpectedly, I find myself moved by the current build.

At first glance I found the model’s reliance on Halachipedia, random Sefaria source sheets, and the Judaism Stack Exchange to be signs of weakness. Personally I have seen these sources of knowledge as lower-thans; they’re tools for folks who are too busy to do the “real work” or lack the textual skills for deep legal analysis, and because they’re mostly produced anonymously I can’t rely on the author’s reputation to determine if they’re poorly researched or misreading sources or neglecting critical ideas. They lack the sanctity of the texts on which they are commenting. They’re not truly part of the chain of transmission. They’re helpmates.

But Deep Research thinks otherwise. It relies on the Torah of the internet in the same way that the Arba’ah Turim, a medieval code of law, relied on the the medieval codifiers that preceded it (namely: the Rif, the Rambam, and the Rosh). In retrospect this reliance seems natural and uncontroversial, but it’s the very act of citing those prior sources that firmly ensconces them as links in the chain of Jewish tradition.

Could the internet be part of the mesorah, the vaunted chain of transmission that rabbis depict as extending from Moses all the way to us? I certainly hope so. One of my greatest anxieties is that all the magnificent Torah that has been developed across the last 40 years will be lost because it exists on platforms—BBS, Usenet, listservs, private Facebook posts, blogs, random personal websites—that are not easily preserved and have never been properly archived. Perhaps AI will finally weave this ad-hoc, multi-generation project into something that we can preserve and build upon.

But more than that: AI’s reliance on “random” websites underplays the fact that online Torah, while often feeling weightless, is really and truly a link in the chain. Sefaria source sheets might be educational tools, but they also reflect the interests of the day. Halachipedia, Virtual Beit Midrash, Responsa Radio, CJLS responsa—they all deserve a gravitas that no contemporary can confer. Those of us who have contributed to online Torah might be too close to see this, but it’s true, and it took an AI to help me understand it. Yes, I’d like to see Deep Research become Deep & Wide Research, but this appears to be a common complaint and it’s not insurmountable.

I’m proud of what the internet has done for Torah. More than connect people and existing ideas, it’s created an ease of access that has drawn in so many people, including people who you’d likely never find in a beit midrash (study hall). It’s also created a conversational tone that allows ideas to develop quickly and collaboratively. We’re all too close to see it, but it is a revolution and it is a Good Thing.

I don’t think AI is going to be the next link in the chain of transmission—but it will certainly play a role in developing that link. And to the extent that AI can synthesize the Torah of the internet, it will be a handmaiden for continuity. That’s something real.

1. I know this reference isn’t going to mean anything to many of you, but it resembles nothing so much as the seminal Edah Journal articles from the early 2000s (the most important at the time were Sperber and Shapiro) that led to the creation of some of the first partially-egalitarian Orthodox prayer services.

2. “The Valley of Search.” Sorry, dumb joke.