LARGE LANGUAGE MODELS (LLMs) — the algorithms that underlie generative AI chatbots like ChatGPT and Gemini — are transforming daily life. Some people already use them as a matter of routine instead of traditional internet searches; many students use them (even when they aren’t supposed to) to complete assignments; many companies have dabbled in replacing work traditionally done by humans — especially writing and design — with material produced by them. And a smaller but growing number of people turn to these chatbots for more intimate needs, spanning everything from friendship to therapy. This extends to spiritual and religious life as well. Is it okay to ask a chatbot a halakhic question, and rely on its reasoning for an answer? Are LLMs accurate sources of Jewish knowledge? And what does it mean to ask a machine for Jewish advice rather than turning to books or rabbis?
Babylonian Talmud, Avodah Zarah 7A
The Sages taught: In the case of one who asks a question of a Sage with regard to an issue of ritual impurity and the Sage rules that the item is impure, he may not ask the same question of another Sage and have him rule that it is pure. Similarly, in the case of one who asks a Sage a halakhic question and he deems it forbidden, he may not ask the question of another Sage and have him deem it permitted.
In a situation where there were two Sages sitting together and one deems an item impure and the other one deems it pure, or if one deems it prohibited and the other one deems it permitted, the questioner should proceed as follows: if one of the Sages was superior to the other in wisdom and [his view was shared by a majority], one should follow his ruling, and if not, he should follow the one who rules stringently. 1
1 THE TALMUD WAS WELL AWARE that, by the time of its redaction in the sixth century or so, there were many competing halakhic opinions circulating in society at large, as well as in the study halls where debates were taking place. A system was required for assessing what one should do in the face of this plethora of perspectives. And so Talmudic rabbis arrived at one major litmus test: the authority of the person issuing an opinion is crucial. Whom you ask matters.
On this view, someone becomes authoritative on the strength of their lineage—whom they learned from—as well as demonstrated facility and virtuosity in their thinking. All of this has a great impact on how we rely on AI. From where does a chatbot derive its authority? Has it been trained on authoritative texts? (We don’t really know; all chatbots are proprietary, and the companies that create them do not generally release information about the data they have been trained on.) Has it demonstrated repeated moments of brilliant legal analysis, or at least a consistent pattern of citing sources accurately, and without hallucination?
Maimonides, Mishneh Torah: “Laws of the Sanhedrin” 2:7
[A judge] must…possess seven attributes: wisdom, humility, the fear of God, a loathing for money, a love for truth; he must be a person who is beloved by people at large, and must have a good reputation. All of these qualities are mentioned explicitly in the Torah. 2
2 HERE WE SEE a second major litmus test: trust. If we think of authority as akin to a credential— proof that someone has attained a certain level of learning — then we can think of trust as the continuing reliability of their judgment. A religious leader is authorized to rule on something because they are ordained; they are trusted to rule on something because their ordination hasn’t been revoked, their past rulings have been accepted as sound and not overturned, and their reputation remains untarnished. Here, Maimonides sums up the position of many earlier sources about what constitutes trust in a halakhic authority. These traits are hallmarks of someone who will offer a sound judgment of a case or a clear reading of the texts.
Rabbi Yehuda Herzl Henkin: “Qeri’at Ha-Torah by Women: Where We Stand Today,” Edah Journal 1:2 (2001)
Where does all this leave us? Regardless of the arguments that can be proffered to permit women’s aliyyot today — that kevod hatsibbur can be waived, that it does not apply today when everyone is literate, that it does not apply when the olim rely on the (male) ba`al qeri’ah and do not themselves read — women’s aliyyot remain outside the consensus, and a congregation that institutes them is not Orthodox in name and will not long remain Orthodox in practice. In my judgement, this is an accurate statement now and for the foreseeable future, and I see no point in arguing about it. 3
3 HERE WE SEE a real-world example of the dynamics of trustworthiness and authority, and of how they play out in a halakhic framework. This is an excerpt from a response to Rabbi Mendel Shapiro, who published a responsa ruling that, even using exclusively traditional sources, a woman may read from the Torah in a regular Orthodox service. Henkin’s reply: even if Shapiro were interpreting all the sources he cited correctly (which he doesn’t think was the case), no one person can establish a communal halakhic practice on the basis of an untested theory that is not accepted by the majority of rabbis or practitioners.
This has real value for our topic in two ways. It confirms that a responsa isn’t useful unless it is well supported by accepted authorities: don’t think that because you’re brilliant, the rest of the rabbinic community will automatically accept your ideas. And it shows that, even if an LLM were to demonstrate that it did not hallucinate and always provided proper responses that conformed with accepted law, unless the general consensus of authoritative sources accepted that this LLM could be trusted, it would remain a suspect source of wisdom.2 HERE WE SEE a second major litmus test: trust. If we think of authority as akin to a credential— proof that someone has attained a certain level of learning — then we can think of trust as the continuing reliability of their judgment. A religious leader is authorized to rule on something because they are ordained; they are trusted to rule on something because their ordination hasn’t been revoked, their past rulings have been accepted as sound and not overturned, and their reputation remains untarnished. Here, Maimonides sums up the position of many earlier sources about what constitutes trust in a halakhic authority. These traits are hallmarks of someone who will offer a sound judgment of a case or a clear reading of the texts.
Rabbi Daniel Nevins, Halakhic Responses to Artificial Intelligence and Autonomous Machines (2019)
When we say that machines are functioning autonomously, currently we mean this in a very limited sense. Machines are given a task and the capacity to complete the task within certain parameters, usually by following algorithms built on a series of predetermined “if … then” rules. They are not capable of establishing independent goals or refusing to act on orders that fall within their operational parameters. Nor are they accorded legal personhood, no matter how personal people may get in conversations with virtual assistants. Just as it would be absurd to punish a courtyard for “stealing” a goat, so would it be absurd to whip an autonomous vehicle in punishment for “murdering” a pedestrian. Legal standing and free will are essential components to moral stature and liability. At this stage, artificial intelligence functions like a tool, and so moral liability must remain with the principal. Or perhaps the machine is more like an animal, in which case its owner is responsible to a greater or lesser extent depending on typical performance. In any event, the machine is not obligated (בר חיובא), as Ravina puts it, nor does it have free will (דאי בעי עביד), as Rav Sama emphasizes. Without these capacities, liability remains with the principal who appointed the agent— the person, not the machine. 4
4 WHILE THIS RESPONSA, which was adopted by the Conservative movement, is six years old —several lifetimes for AI—it addresses many broad implications of AI for Judaism. In particular, it runs through a series of questions that help us understand what it would look like for a chatbot to become a viable guide to Jewish thought. One crucial element of reaching such a state: taking responsibility for interpretations of Jewish law that a chatbot issues. Rabbis who offer guidance on Jewish matters know the weight of responsibility that comes with offering an answer or solution: they bear responsibility if they make a mistake in reasoning, causing someone else to behave improperly. If chatbots are ever to be viable alternatives, then rabbis must have a hand in training them; there needs to be an autonomous human who ultimately bears responsibility for any advice that is offered.
A. Trevor Sutton, “AI and the Discipline of Human Flourishing,” Religion & Liberty 34:1 (2024)
Long before the advent of generative AI, [recently deceased philosopher of technology] Albert Borgmann…in his book Technology and the Character of Contemporary Life…argued that technology has shaped contemporary life around its peculiar pattern. Borgmann suggested that the pattern of technology becomes particularly harmful when there are no means by which one can “prune back the excesses of technology and restrict it to a supporting role.”
Borgmann makes a distinction between “focal things” and “devices.”
A wood-burning stove is a focal thing: it requires skill and bodily engagement through woodcutting, seasoning wood, and fire building. This thing exists within a context of forest, home, family, and community. It leads to social engagement and focus as multiple people contribute to the process…
A device stands in stark contrast to a focal thing. Devices make no demands of skill, strength, or attention. Devices provide commodities for enjoyment without encumbrance or context. The lack of encumbrance makes the commodious consumption of devices thoughtless and disposable. Technological devices produce a commodity without burdening us in any way. Devices are quick, easy, foolproof, and safe. A furnace or central-heating system is a device. These devices provide warmth without any demand from the recipient.
ChatGPT is also an example of a device. This device provides a commodity — summaries, essays, answers — without any skill, preparation, or demand on the user.
Things require skilled and active human engagement; devices require no focus, engagement, or context. Things require practice; devices invite consumption. Things constitute commanding reality; devices procure disposable reality. 5
5 A LUTHERAN PASTOR writing in a conservative think tank journal, Sutton warns of the dangers of relying too much on AI by pointing to the potential loss of human creativity. The distinction he describes here, between things and devices, is tricky to maintain: try telling your grandmother that a washing machine is just a device and if she doesn’t “engage” with the laundry she risks all of human creativity. Put another way, the extent to which any individual relates to an object as a device, rather than a thing, may have as much to do with novelty and with cultural expectations as anything inherent to the objects themselves. The printing press, which automated writing, and may seem analogous to Sutton’s furnace supplanting the wood stove operator, not only created a massive explosion of Jewish creativity and thought by allowing many more people to read and study written works, it created whole new cohorts of creative skills, like type designers and printers.
Chatbots, arguably, have the same potential. Sutton is right in saying we need to be vigilant and monitor how their use is adopted and evolves. We have to look only to social media to see how our use of a technology can go awry. He is also right that we do not want our generation to be the last to know what human creativity looks like. But he neglects the changing cultural context: just as a person can engage creatively in making or interpreting a printed book as much as a calligraphed scroll, our relationship to chatbots—and the extent to which we are able to engage with them creatively and skillfully—is likely to change over time.
Rabbi Joan Friedman
Responsa committee chair, Central Conference of American Rabbis
I HAVE BEEN USING AI in many ways as a rabbi. In particular, I’ve been using it to create poetry that I use as part of our minyan. For example, I have created poems for the Asher Yatzar, for Kaddish, etc. These poems are the result of my prompts, and carefully selected themes that I wish to incorporate. I do not believe someone without a sophisticated understanding of Judaism could create what I create. If I write a story or poem with AI, can I take credit for the work, or do I need to indicate that it was created using AI?
We found that using AI is like using a tool, but not in the sense of a hammer. We asked ourselves whether we regard this as if somebody says I built a house, does that mean that they physically built it, or that they hired a contractor to build it? And for us, it still seemed closer to having hired a contractor to build this house: to claim something written with AI as one’s own, we found that misleading and g’neivat da’at (intellectually dishonest).
We said we are not making grand pronouncements on AI and large language models, but in this particular circumstance, there is a difference between a person being completely original and a person using AI to create something. And we think that there’s a difference, there’s a lack of originality that needs to be acknowledged, and not to [acknowledge it] is being deceptive.
AI cannot do what any rabbi does, or any face-to-face person is doing, which is to take into account the circumstances of the questioner and to ask the questioner: Why are you asking that? What do you need? And I would not necessarily trust it to carefully capture all of the nuance that a person could.
THERE ARE TWO ISSUES underlying the question of whether chatbots should have a role in shaping people’s practice of Judaism. One has to do with an LLM’s capacity to provide accurate information with clear sourcing, backed up by human authorities — that is, with its capacity to issue answers and interpretations that are authoritative and trustworthy. This is a matter of the substance of a chatbot’s answer. On that score, there is reason to believe that, eventually, LLM-based tools will exist that can credibly answer specific questions about Jewish practice and thought.
But there is a somewhat separate issue that is perhaps at least as important, which is a psychological one: What does it mean for a person to be in relationship to a machine, to turn to it with their spiritual, religious, and deeply personal questions? There is real danger in humans developing deep attachments to chatbots — attachments that superficially mimic relationships to other humans, and which can come with all the attendant feelings of dependence and vulnerability. AI cannot and will not ever be a replacement for a real-life relationship with a rabbi, mentor, study partner, or friend. Judaism exists in community. Halakha exists in community. A rabbinic answer, by necessity, takes into account the person asking it and the cultural milieu where it is being asked. Insofar as Judaism evolves, it does so because of social forces and human innovation. This is what Orthodoxy calls mesorah (tradition), which values the relationship between teacher and student and seeks to adapt to the changing world. This is the force that drove others to develop their own thriving forms of Judaism. If we hand the reins to AI, then the Judaism of the future will no longer be the living, breathing tradition that we have known for millennia.
