Can AI Be a Link in a Chain of Tradition?

Rabbi Gilah Langner: So welcome, everybody. Come on in, find a seat. Delighted to have you. Ladies and gentlemen, esteemed guests and scholars, welcome to today’s lecture on the intriguing intersection of artificial intelligence and Jewish ethics. In an era defined by rapid technological advancements, the integration of AI into our daily lives presents us with profound moral and ethical questions. As inheritors of a rich tradition of ethical inquiry, Jewish thought offers invaluable insights into navigating this complex terrain.

Now, as we stand at the forefront of the AI revolution, we find ourselves confronted with new dilemmas. How do we ensure that AI systems reflect the values of compassion, justice, and respect for human dignity? In this lecture, we will embark on a journey that traverses the realms of philosophy, theology, and technology. We will delve into ancient texts and contemporary debates, seeking wisdom to guide us in our quest for responsible innovation. 

Well, as you probably guessed, those paragraphs were written by AI. (Laughter) Oh dear, I don’t sound like that specifically. I don’t sound like that.

Jeremy Epstein: I was thinking it sounded kind of stilted for you.

Gilah Langner: So there’s the AI. I’m Jewish, and is it ethical to ask ChatGPT to come up with your introduction? I don’t know, but there you have it – AI and Jewish ethics. So I did, by the way, have to turn the lights off to compensate for all the energy I used just to get that [opening].

So I would just say, folks, welcome to Kol Ami. I’m Rabbi Gilah Langer, and I’m delighted to open the third in our series called Science meets Judaism. We’ve already considered creation and restoration ecology. We’ve explored animal sentience and Biblical tales of animals. And our final panel, on May 19, is called “How Long is Forever?” And it will feature noted physicist Dr. Michael Dine and his spouse, Rabbi Melanie Aron, the first time they’ve ever spoken on a panel together.

And now I have the pleasure of introducing our panelists. Jeremy Epstein is the assistant director for technologies and privacy at the White House Office of Science and Technology Policy. Prior to this position, he was the lead program officer for the National Science Foundation’s Secure and Trustworthy Cyberspace (SatSI) program, which is NSF’s flagship multidisciplinary program for cybersecurity and privacy research. At NSF, he managed an $80-million-a-year budget and was responsible for oversight of most academic security and privacy research in the United States, with over 1,000 active projects, 1,500 faculty members, and 2,500 graduate students. Prior to NSF, Jeremy held roles at DARPA, SRI International, and in commercial software companies. He is chair – are you still chair?

Jeremy Epstein: No, actually, I had to resign.

Gilah Langner: He used to be chair of the Association for Computing Machinery’s US Technology Policy Committee and was founder and director of Scholarships for Women Studying Information Security. After nearly 30 years as a member of congregation Beth Emeth in Herndon, Virginia, Jeremy and his partner, Rabbi Julie Gordon, joined Kol Ami in 2022. Rebecca J. Epstein-Levy –– and yes, a relation, yes – is Assistant Professor of Jewish Studies and Gender and Sexuality Studies at Vanderbilt University. She’s an ethicist who studies sex, sexuality, and disability in rabbinic texts. And she’s the author of When We Collide: Sex, Social Risk, and Jewish Ethics, which was just published last year. One reviewer said of the book, quote, “Epstein-Levy has written the first book in Jewish ethics that neither condescends nor preaches to its reader. This is nothing short of liberation. There are very few academic books that are such a joy to read.”

In addition to teaching at Vanderbilt, Rebecca has taught at Washington University in St. Louis and Oberlin College. Her PhD in the study of Judaism is from the University of Virginia, where her dissertation was entitled “Safe, Sane, and Attentive: Toward a Jewish Ethic of Sex and Public Health.” And her prolific publications range over topics such as intertextually modified organisms, genetic engineering, Jewish ethics, and rabbinic text and community authority and autonomy: Jewish resources for the vaccine wars.

Rebecca identifies herself as Jewish, incredibly gay, and an adult-diagnosed autistic and ADHD’er. She’s a lover of cats, an amateur equestrian, a sometime visual artist, an erstwhile student of classical violin, a frankly excellent home cook, a nice nerd, and a virtuosic – can I say this, kvetch about the house? Welcome to you both. It’s an honor to have you speaking today. And Jeremy, I’m going to turn things over to you first.

Jeremy Epstein: Thank you. I would say it’s an honor to be here, but I’m here anyway. But it is an honor to be here with my daughter-in-law. I expected at some point in my career I would have an opportunity to do things jointly with some of my children, but I thought the most likely was my son and my other daughter-in-law because they’re computer scientists-ish. But this was just such a great opportunity to do this. So thank you, Rebecca, for schlepping up here from Nashville. 

So I have to start by giving a disclaimer, especially because I believe this is being recorded. So I’m going to share my opinions. These are not those of the White House, the office of Science and Technology Policy, the National Science Foundation – they’re probably my opinions, but that’s probably about as far as I’d go. And I’m not representing anybody. And I should just note that at OSTP, I do not work on artificial intelligence topics, so I am definitely expressing my opinions and not speaking for Uncle Sam or Joe Biden.

What do we mean by “AI”?

So let me start by talking just a little bit about “What do we mean by AI?” First of all, artificial intelligence is not a new thing. It’s been around for 50, 60 years, depending on how exactly you count.

There was actually this great story that someone was assigned, in the early 50’s, a summer research project. This was a graduate student. “You’ve got the summer – figure out how to make computers understand speech.” It’s taken a little while, but it’s only recently become an issue, or  become visible to the public. And that’s because of huge amounts of data that we see through places like Google. It’s also huge amounts of computational capacity which have allowed us to do things that we couldn’t do 50 years ago, or even ten years ago.

But I need to emphasize, AI systems are not magic. They don’t think; they’re just doing statistical analyses of lots and lots of data. When we say “AI systems,” typically these days we’re talking about LLMs, large language models, which are basically a piece of software and data that’s “read” (and I’m going to use that word in quotes – “read”) – lots of text, pretty much the entire Internet, and has come up with a model of what text looks like. And so if they see one word, they can predict the next word and the next word and the next word, or the next phrase, and so on. So that’s why they can sound pretty coherent. But that doesn’t mean that the system actually understands, in a human sense, anything it’s saying.

There are other kinds of AI. I’m not going to talk about those, other than to mention things like artificial neural networks that are used for image generation, and there are a lot of other things, too.

One of the real challenges is that we don’t understand how AI systems come to their conclusions. So when Gilah asked, “Write me a spiel about” – I don’t know how you phrased it, but “write me a spiel about AI and Judaism and science,” we don’t know how it came to that conclusion, other than it read lots of stuff and spit something out that sounds coherent. And it’s very hard to get AI systems to explain how they came to a conclusion.

Having said all of that, let me delve into, from a non-Jewish professional perspective, but a Jewish person‘s perspective, some of the areas where I think AI can be helpful when we look at Judaism. We can use AI systems and LLMs, in general, for improving education. We’re seeing things like AI systems to help study Talmud or answer questions, because they can get lots of data and they can bring together sources. When you try to read the Talmud and study all this, there are just so many different sources. It’s hard for us as humans to bring it all together, and an AI system can bring it together effectively. Is he going to get it right all the time? No, but it can help us. We can use AI systems to help us with looking up questions.

So just for the heck of it, just like you asked, do you use ChatGPT or one of the others use ChatGPT? I use ChatGPT. I said, “What happens if you drop a drop of milk in a pot of chicken soup?” Because I was curious. Would it know the answer? I don’t know if you know the answer. The answer is, basically, “It’s okay, as long as it wasn’t intentional, and as long as it gets diluted enough, it’s okay.” And it actually got it right.

 So there are a lot of things that you can ask an LLM to do instead of bothering Rabbi Gilah. One of the things I ran across that someone proposed is you could use an AI system to detect sha’atnez. Okay. How many of you know what sha’atnez is? A few – okay. It’s the mixing of wool and linen, which is prohibited in the Torah. And I don’t think anybody really has figured out why –

Rebecca Epstein-Levy: Except if you’re a priest. 

 Jeremy Epstein: Oh, okay. I didn’t know that. We don’t have any idea why it’s a restriction, but it is a restriction, and there are proposals to use AI systems to make that division. 

Using AI to Support Jewish Values

There are things we can use AI for to support Jewish values. So I’m trying to give the positives first. Improved medicine – faster and more accurate reading of x-rays and other images for better medical outcomes. Inventing new medicines that can save lives and make life better, things like that. We can use AI systems also for doing things like detecting impaired drivers, a drunk driver – are you a sleepy driver? An AI system could detect that maybe accurately, maybe not, and save lives that way. These are examples, in my view, of Tikkun Olam. They’re making the world better through technology. And so we should look at it positively about supporting Jewish ethics.

But with Jewish ethical questions, some of the negatives, are more complicated. AI systems get used to make decisions like, “Can you get a mortgage? Can you get released from jail on bail? Can you, maybe someday, should you get tenure?” (Hopefully it’s not going to make that decision. And getting tenure might be compared to, do you get released from jail or not released from jail? I’m not sure. Well, we’ll leave that for now.)

But one of the challenges is that some of these systems are proprietary. Only the vendor knows how these decisions are being made, and the AI system can’t explain it. And so, we also know that these systems have a historic propensity for discriminatory behavior – redlining on housing, keeping people of color locked up – and these contradict Jewish ethics.

We’re also seeing AI systems being used for facial recognition, including for law enforcement. And we know that this has been a problem. The systems have historically been much less accurate for women and people of color than men and white people. And since treating people fairly is a Jewish ethic, I’d say, as a non-ethicist, I’d say using an AI system for these sorts of things is in contradiction to Jewish ethics.  AI systems are going to put people out of work – we all read about this in the newspaper. Self-driving cars are going to put taxi drivers out of work. Things like that – 

Rebecca Epstein-Levy: – also kill people.

Jeremy Epstein: And kill people. Yes, absolutely. They’re going to put lawyers out of work – yeah, I’m not sure if that’s a pro or a con.

There’s a new product that came out this week that purports to replace programmers. There have been many cases in the past of things that are going to replace programmers. They typically don’t replace programmers, but there’s a claim. But I’m going to argue that there’s an ethical question here about developing these technologies that put people out of work – ethical with respect to Jewish ethics.

And another example is AI systems. The way they get created is they read, quote-unquote, “the Internet,” including a lot of copyrighted stuff, and then they come up with stuff. Is that a violation of copyright? There’s a lawsuit currently pending. The New York Times has sued OpenAI and Microsoft and a bunch of other folks. We don’t know what the law is going to say, but since copyright is a Jewish law and a Jewish ethic, I’m going to say that maybe that’s an area as well.

AI systems are used in military targeting at many different levels – where the enemy is hiding, using sensor fusion, things like that. Is that ethical from a Jewish perspective? 

AI systems can be used to teach people to do things that we disagree with. They can be used to teach people how to create biological weapons – how to create weapons in general. How do we square that with Jewish ethics? AI systems are used both to create, and to detect, hate speech. And so, depending on how it’s used, we could either argue that’s ethical or not ethical. Another point I want to talk about, for just a second, is contestability. This is a new word I learned recently, and it looks like you’ve known it for longer than I have. 

Rebecca Epstein-Levy: No, actually, I learned it from you.

Jeremy Epstein: Oh, okay (laughs). I learned it from my friend Susan Landau. It actually turns out if you Google it, there’s a whole lot of discussion of contestability. It’s about the ability of a person or an organization to challenge a decision, specifically a decision of an AI system. “How do you say it says I shouldn’t be let out of jail on bail, but I think I should? How did you come to the conclusion that I wasn’t a good choice to get out of jail and these systems can’t?” And so this is a big issue. What if the AI system says that this person should get the medical treatment and that person should not, and it can’t explain why? How do we square that with Jewish ethics?

Jewish Questions about AI

Okay, so let me ask a few controversial questions, and then I’ll turn it over to Rebecca. Could an AI system count towards a minyan (laughter)? Could it write a kosher Torah scroll? Could it serve on a beit din for conversion? The Orthodox clearly say “No.” I think probably almost all Jews would agree that those are things we don’t want AI systems doing.

But could an AI system infer what a Shabbat-observant person is intending to do and take action? Like, if they’re squinting. It says, “Huh, I’d better turn on the lights.” But the person never asked to turn on the lights, but it inferred from their behavior, because the cameras are being interpreted, et cetera, et cetera. If it sees somebody shivering, could it turn up the heat?

These are things that are well within the capability of current AI systems.  Does the fact that that AI system is there – does that become a violation of the rules of Shabbat for somebody who’s shomer shabbos? Or even, someone pointed out, it knows you – like, it’s observed the other six days of the week – if the first thing you do when you wake up is you make yourself a cup of coffee. And so it says, “Hmm, woke up. I’d better make you a cup of coffee,” and then the AI system makes you a cup of coffee. But to do that, it had to boil water, et cetera, et cetera. Is that a violation?

Okay, I think this is almost my end, but not quite. I’ve been learning recently about something called BCIs. Does anyone know what a BCI is? Ooh, a new acronym for you. A Brain-Computer Interface – the headsets, and things like that. And sometimes they’re just earpieces that can monitor your brain activity, either externally, like the headsets, or internally, with brain implants, like in Neuralink, which has gotten a lot of attention.

Neuralink, incidentally, is redoing a lot of academic research from the past 30-40 years, and they’re not nearly as advanced, but they’ll catch up, because Elon Musk has a lot of money, so they will.

Rebecca Epstein-Levy: And no sense.

Jeremy Epstein: And no sense. Absolutely. And they use electrical and magnetic sensors, and in some cases, optical sensors. I haven’t quite figured that one out yet.

So let’s think about that from a Jewish perspective. Could you fulfill a mitzvah by thinking about it? “I think about the prayers, and so therefore, I’ve davened.” Okay, that sounds kind of ridiculous, but if someone is a paraplegic and can’t turn the pages in the Siddur, or they’re shut-in and they can’t speak, maybe that’s – we should consider that. So, look at the positive and negative.

But what about thinking about something prohibited? Is thinking about it itself prohibited? If I think about how delicious a ham and cheese sandwich is, have I eaten a ham and cheese sandwich, especially if I can cause the neurons that correspond to taste to trigger so I can taste it? Even though there’s been none of that going on physically? Where does that fit?

And we don’t want to say, “You can’t use these sorts of systems on Shabbat, because they might cause you to do any of these things,” because someone might need these things to survive – like someone who’s quadriplegic or has other disabilities, someone who has Parkinson’s disease, might need these to survive. So we can’t just say, “No, we can’t do them.” We have to balance it.

Hallucination and Inaccuracy

I’m going to just touch on hallucinating, and then I’m going to conclude. Who’s heard of the term “hallucinating” with respect to AI? A few people. So, “hallucinating” means the AI systems make up facts that look like things that they’ve really seen – “seen.” And so I asked, as an example, for a bio for my son (who might watch this on rerun, so I hope he doesn’t get mad that I’m pointing to him, but I think I shared this with him). He is an expert, he’s a professor of human computer interaction, and [the AI] knows what people who work in human computer interaction, where they typically go to school, where they typically intern, where they typically teach, what courses they teach, et cetera.

So I asked it, “Give me a bio for Daniel Epstein,” and it gave me a bio, all of which was false. Every single fact in it was incorrect, because – there may have been one fact that was correct – but anyway, it was all wrong because it knew what someone who’s at UCI looks like. They typically go to somewhere like Carnegie Mellon – he went to the University of Washington. They intern at Apple, and he interned at Microsoft, and so on. So it sort of had the right premise, but it was all wrong.

Well, this has been a problem, because there have been lawyers who have submitted documents to the court – I don’t know what the proper term is – that cited cases that don’t exist. It made up cases, including complete citations to the law, that just didn’t exist. And lawyers get in trouble for doing things like that, because when you’re a lawyer, you’re supposed to tell the truth to the court. So this is an example of hallucination. 

So Jewish law values privacy protection, accuracy…so how do we square that? So let me conclude by saying AI can be supportive of Jewish ethics, but we need to consider it on a case-by-case basis. We need to be questioning, not simply accepting, and we need to think about contestability.

So let me conclude with two things. I asked ChatGPT: “Is the use of an AI system to determine whether something is kosher ethical?” And it said, “The use of AI systems to determine whether something is kosher can be considered ethical, but it depends on several factors.” And then it says, “the ethical use of AI systems, blah, blah, blah…” – factors such as accuracy, transparency, respect for religious authority, and cultural sensitivity. It likes to use the word “cultural sensitivity” – I’ve never. “When developed and used responsibly, such systems can offer valuable support to individuals seeking to observe kosher dietary laws.” In other words, it didn’t really say anything.

But the really last thing I’m going to say, before I turn it over to Rebecca, is I figured out what my next job will be when I’m done in the White House is. I was reading this great New York Times article I want to fill in for Paolo Benanti. He is the Vatican’s Artificial Intelligence Ethicist, and he gets to advise the Pope on AI. And I think that would be a cool job. Rebecca, over to you.

The Field of AI Ethics (and How It Can Be Jewish)

Rebecca Epstein-Levi: Thank you. So, as Rabbi Gilah mentioned in that lovely introduction – and I am pleased that you visited my website, thank you – I am a Jewish ethicist. And I wanted to start my discussion with a quote from someone who is an ethicist, but who is not Jewish, but who is trained in religious studies to kind of set the tone for where I’m going. Damien P. Williams is an AI ethicist, among other things, and in a recent paper, he wrote the following:

 Like all technologies, AI reflects human bias and values, but it also has an unusually great capacity to amplify them. This means we must be purposeful about how we build AI systems so that they amplify the values we want them to, rather than the ones we accidentally feed into them. We have to ask questions about the source material that trains them, including books, social media projects, news and academic articles, and even police reports and patient information. We must also examine the frameworks into which that data is placed. What is the system doing with that data? Are there some patterns or relationships between words or phrases that are given more value than others? Which ones? Why? What are the assumptions and values at play in the design of tools that transform human lived experiences into data, and that data into algorithms that impact human lives?

I want to also preface this by saying: I am not a technophobe. I’m actually pretty sanguine about some forms of technology. And in fact, the book chapter that Rabbi Gilah mentioned is me, with my lefty liberationist politics, trying to make a Jewish defense of genetically engineering crops. And that is something that I think actually has some significant liberatory possibilities, through things like biofortification, drought resistance, et cetera, et cetera, if – and I emphasize this greatly, if – we put a lot of effort into making sure that those interventions are public domain and easily accessible to the people who need them most. However, I am, from my perspective as an ethicist, extremely skeptical about how possible it is to make ethical use, Jewish or otherwise, of the specific large-language models under discussion.

And if I could move forward – this is a big wall of text, but this is my standard policy on the use of AI in my classes that I put on my syllabi. And the upshot is, well, “Don’t cheat. Please don’t cheat in ethics class specifically. I can’t believe I have to say this, you know, don’t defeat the purpose of the assignments I sent you, which are not to make your life miserable, but to actually get you to think through a thing.”

AI and Attribution

And I want us to focus on number three. I consider its use, as it’s being used now, deeply unethical. Jeremy already mentioned some of the significant and very real problems that have already come up and are continuing to come up as large language models get used to make really consequential decisions about people’s lives. We might remember, for example, the recent Screen Actors Guild actors’ strike, and one of the significant provisions that the unions were arguing for was the right to, essentially, veto privileges on AI-generated use of images and voice recordings and such of them.

It’s a big problem in journalism, both from a labor perspective in terms of layoffs, but also from the perspective of having actual humans following stories, getting deeply into them, and being able to fact-check and verify what they’re doing.

Jeremy shared what large language models are and are not capable of right now, and one of the things they are definitely not capable of is the sort of legwork that really good journalists do. But if you’re looking at it from a bean-counting perspective and trying to save money on salaries, you’re not going to consider that.

And – this is where I want to focus the remainder of my remarks, largely – there is also the way that AI, that these large language models are trained, by scraping a vast swath of content without attribution, and certainly without compensation to the people who did the labor and produced that content themselves.

I also do like to point out that it is my students’ generation, in particular, that is being set up to be screwed especially by this, and that is something they should be cognizant of. And if I could have the next slide – so I’ve given my opinions here, but how am I going to relate this to Jewish ethics specifically, other than “I am a Jew with lots of very loud opinions who happens to be making ethical judgments”?

Defining Jewish Ethics

Although, in all seriousness, I think that actually is one not wholly incorrect possible answer to the question “What is Jewish ethics?” Jewish ethics is, possibly, Jews doing ethics. And in fact, what makes Jewish ethics Jewish is a question that many Jewish ethicists do not internally agree on. In the classical discipline of academic Jewish ethics, which is very young – the academic discipline, specifically, historically, there’s been quite a lot of overlap between people who write academic Jewish ethics and poskim, that is, rabbis who do halakhic decision making. And so there has also been a lot of formal overlap between the two genres. However, in the past, I would say three decades, especially, that has changed considerably. And people, and notably women, nonbinary people, more people of color who have historically been excluded from posek positions, have been entering and redefining the field more and more. 

We have also – and I count myself as one of these – have really started to ask, “What exactly is the relationship between Jewish ethics and halakha? How do they differ? And if they’re not the same thing, which I think more and more of us think they’re not, what makes what we do Jewish ethics?” 

Is it Jewish ethics if it’s derived from Jewish sources? Probably, yes. But how do we decide what counts as a Jewish source? Are we simply sticking to the classical, rabbinic canon – or are we sticking to Torah in the broad sense? Or, are we also taking into consideration the lived experiences of Jews making ethical decisions – something that my friend and colleague Michal Rauscher argues, and I think quite persuasively, we actually really need to be doing primarily?

But the Jewish ethics I do sort of try to take in aspects of all of these. I do primarily work with classical rabbinic sources, but I do it a little bit differently than it’s historically been done. And I’m part of a group of several ethicists in my generation who are reconsidering these questions. 

Historically, Jewish ethics has tended, when it draws (which it usually does) on classical rabbinic texts – it asks, “Okay, what’s the present-day question we’re asking?” What’s a Talmudic source whose simple subject matter seems to deal with it? Okay, what does the Talmud say about x?” But in a really foundational paper written in 1990, Louis Newman, talking about Jewish ethical discourse on end-of-life care – when it is and is not appropriate to withdraw life-sustaining treatment or equipment – noted, because one of the classical sources that has usually been drawn on for this purpose is a text about whether or not you should keep someone who is chopping wood away from the bedside of someone who is imminently dying, the idea being that the distraction – essentially, it’s too damn loud for the person to die. But is it considered hastening the death if you get the woodchopper to leave?

And the assumption has been, “Well, a woodchopper is like a ventilator.” Which sounds absurd on its face, right? But this is the reasoning that has been kind of stock-and-trade for a while. And Newman asked, “But is a woodchopper a ventilator?” And he exhorted us to be very wary of, and I quote, “Employing a model of textual interpretation which assumes, first, the texts themselves contain some single determinant meaning, and second, that the exegete’s role is to extract this meaning from the text and apply it to contemporary problems.” Newman asks, “How about we don’t do that?”

And then this claim gets actually taken further by another friend and colleague of mine, Emily Filler, who asks, “What if what’s actually Jewish about Jewish ethics isn’t so much the sort of strict content of a given Jewish text that we’re deriving ethics from, but what if it’s the form?” What if it’s not so much trying to find that content that matches the contemporary problem, so much as it is that what makes Jewish ethics Jewish is that there’s a particular way our sources are teaching us to think, that’s reflected in the formal features of our classical texts? 

And that is the avenue I want to pursue in our limited time today, thinking about AI, because the fact is, I don’t think there’s anything in – the Talmud does not actually say anything about AI. Shocker, I know. Nor do I think you can say “Is a scribe ChatGPT?”, for example. But I do think, again, going back to the question of attribution and traceability, verifiability, and so on, there, I think, classical Jewish text, such as the Babylonian Talmud, actually has some interesting formal features that can help us think through this. 

Why We Need Attribution in Jewish Tradition

And one, as you might know, one particularly characteristic feature of rabbinic texts is what’s called the “memra chain.” “Rabbi A, son of Rabbi B, said the name of Rabbi X and the name of Rabbi Y (attributed statement).” And, you know, it’s the sort of thing that if we’re learning Talmud – as I’m sure we all do (sarcasm), but I do – it’s the sort of thing that when we’re sort of reading the bit out loud, we might be tempted to say, “and memra chain, memra chain, memra chain, content.”

But I think the mechanic of the memra chain here is important, because one feature of many large-language models in the current usage, and one which, as Dr. Williams from my first slide noted, both reflects and amplifies existing structures of power and oppression, is that they’re trained by, again, scraping this broad swath of content without either attribution or compensation. 

And there are two ways, I think, this amplifies – or two major ways that I’m focusing on here, anyway, that I think it does this work. First, it treats creative, scholarly, and journalistic work merely as content without – or data – without respect either for its inherent value or the very tangible value of the labor needed to produce it. Second, because the sources of information that it gets, because, again, we don’t know how it comes to its conclusions exactly, but we do know that it’s drawing on what’s available out there.

But these sources aren’t easily traceable, and so it makes it difficult both to check the veracity of content, and it makes it easy to present regnant biases as objective truth. 

Life-or-death example here: you may have heard, I think, in this past summer, some of the kerfuffle about, “Oh, crap. There are mushroom foraging guides available on Amazon which appear to have been written by a large language model.” And many of you perhaps know the expression, attributed to Terry Pratchett, among others – “All mushrooms are edible. Some of them are only” – hey, I’m not finished. (Laughter) “All mushrooms are edible, comma. Some mushrooms are only edible once.” Please allow for the dramatic pause to happen. Thank you. 

So this is a very stark example of why it matters to be able to check the veracity of the content being generated. But there’s also the question of reproducing biases – again, something that Jeremy alluded to. The paper from Williams I quoted opens with this, as follows: “Recently I learned that men can sometimes be nurses and secretaries, but women can never be doctors and presidents. I also learned that Black people are more likely to owe money than to have it owed to them. And I learned that if you need disability assistance, you’ll get more of it if you live in a facility than if you receive care at home.” This is “wisdom” bequeathed by asking various questions of these systems, and again, these are being used in very real ways.

There was a case I ran across where an AI system in – I want to say it was the state of Pennsylvania – was being used to make custody decisions about questionable child neglect cases. There was a couple, both of whom had ADHD, whose baby was not exactly thriving. And I’m not a parent, and I never will be one, but I’m sure many parents can attest that you can be doing everything scrupulously right, and the baby still might not appear to be thriving. The question is, who you are has a lot to do with if you get blamed for that. And as it happens, the system that the state was using to make custody decisions had their diagnosis available to it, and it made the recommendation, based on the disabilities of the parents, that they were not fit to have their child, and the child was removed. I don’t know how that resolved, but that is a really stark example of why you actually need a human on these things. Not that a human might do it better or not that a human will necessarily do it better, but there’s at least a possibility. 

Citing and Remembering Using the Memra Chain

So, attribution, traceability, verification – what, then, can we learn from the ways rabbinic texts formally build in systems of attribution, both in terms of giving credit and improving traceability? And how can we draw on that as arguments for how important it is, at the very least, to continue to aggressively ask these questions of – I won’t say “these systems,” because the systems, these AI systems, can’t think for themselves, but the people who are developing and promoting them.

I want to give one example of a text where not only is attribution sort of clearly demonstrated over and over again, but the attribution is actually key to how the argument resolves. So this is from Babylonian Talmud, Berakhot 2b, and it’s part of a longer argument about what the appropriate time to begin to say the evening Shema is. And one rabbi, in a Mishnah that I haven’t quoted here, but is quoted previously in the conversation, says, “When people retire to eat their evening meal is the time that you should say the Shema.” But the sort of anonymous redactor voice of the Gemara asks: “Wait a second. Not everybody retires to eat their meal at the same time, because a poor person and a priest have potentially very different times that they’re going to go in to do that, that are ritually prescribed, in fact.”

 And so it quotes another tradition from a baraita that in some ways echoes, and in other ways deviates, from our Mishnah.

From when does one begin to recite the evening Shema? From the time when the day becomes sanctified on the eve of Shabbat: This is the statement of Rabbi Eliezer. Rabbi Yehoshua, echoing our mishnah, says: From the time when the priests are eligible to partake of their terumah, a specific ritually separated food. Rabbi Meir says: From when the priests immerse themselves in order to partake of their terumah, because you have to be in a certain state of ritual purity to partake of this.

 Rabbi Yehuda said to Rabbi Mayer, “Do the priests not immerse themselves during the day?” “This is not resolved, Rabbi Hanina said, when the poor person enters to eat his bread with salt.”

But Rabbi Aḥai, and some say Rabbi Aḥa [and I want us to note, here, this formal feature that actually has a way of dealing with uncertainty, and flagging it as to whom the statement should be attributed to] – And some say Rabbi Aḥa says: From the time when most people enter to recline at their meal during the week.”

And we now go back to the Gemara, which says,

“So if you say that the time of the poor person and the priest are one time, the same time, and that’s when you should say the evening Shema, that doesn’t work with the spirit, because then the opinion of Rabbi Aḥai and Rabbi Yehoshua would be identical. And that can’t be the case because otherwise, why would the baraita cite them separately?”

 In other words, the baraita is only going to cite one opinion to one effect, but there are two cited, attributed opinions that disagree – that separate out the thing that we’re trying to claim is the same.

“Rather, must one not conclude from this that the time for the poor person is separate and the time for the priest is separate? Indeed, you have to.”

So this is a case where the attribution makes or breaks the argument. And I bring this up because I think it shows us that proper attribution and traceability is not just a formality. It’s significant to the specifics of the case the Gemara is making. And I would argue it’s also ethically significant, both because, not to worry, the only reason this doesn’t usually happen to me is that I always forget to turn my ringer off, which then sometimes leads to me missing important calls. 

So, you know, so not only is it halakhically significant, it’s also ethically significant, because it’s important to know where it came from, and it’s important to know those citation rules in order to actually, if we assume that it actually means to make a practical ruling, which is questionable – but we’ll bracket that – if it wants to make a practical ruling, it needs to have actually taken into consideration why the sources it’s working with are making the particular claims they do.

In other words, it’s saying that verifiability, and the form that that takes, matters. It matters to day-to-day practice. It matters to the ethically important question of fulfilling mitzvot.

Can AI Be Part of the Conversation?

Audience Member: It’s making me think that, you know, the Talmudic as you, the discussion is a process. And so would we, if we were using the AI as part of that discussion, count as a person, as part of that discussion? Because I went to an interfaith thing and somebody said, “You know, I’ve heard that when you have eight Jewish people in a room, you have 20 opinions.” So how does that help? And the rabbi said something very interesting, which was, “Because at the end, there’s consensus.”

Rebecca Epstein-Levy: Except there isn’t always. But that’s important too, right?

The process is that they come to maybe more of a consensus and a decision, or sometimes they don’t, as you say, but if they need to come to a decision, there will be a decision.

And so if you’re using an AI opinion, would that count as a person as part of that discussion? Because I think what we’re saying is we don’t want – the AI might have something interesting to say, but we don’t want it to stand by itself.

Rebecca Epstein-Levy: So there are a couple of different ways you could take that, I think. If we’re sort of thinking about the question of “What’s the threshold of counting as a person?” Again, you pointed out, they can’t think for themselves. Or as my father, of blessed memory, used to say when he was teaching beginner computing classes, “Computers are stupid. They can only do what you tell them to do.” And I think, and this is something that, to the best of my understanding, continues to hold true here, right.

So, do we count the AI as a person? Well, there’s a whole set of separate Talmudic discourses about what counts as a person for the purpose of a minyan. Does a Torah scroll count? Does a tree count? Does the Sabbath day count? And again, this is one, if memory serves, this is one of those where we actually don’t really come to a good conclusion. So there’s a line of inquiry you could go off there, which is going to get into questions of, “Okay, how does your cosmology account for being a person, for X or Y purpose?” 

But, I think what’s perhaps more important here is that one thing I think these large-language models do is often present something as settled when it isn’t. It creates an illusion of consensus. And another formal track I could have gone off here, which I just thought about, would have been precisely to use the rabbinic form of “Teiku” – that is, “let the argument stand,” as another sort of formal corrective to these trends we see. That did not answer your question, but…

It wasn’t really about the AI being a person. It was more like the process. We can include AI and say, “Okay, let’s take AI into consideration.” It’s another opinion. So it’s just another opinion, not necessarily a person, but part of that process. Because I think in Judaism, we have this process, and we don’t just take one person’s opinion. And I think that’s so much, so important in this, that you’re saying that there are things that go out and rely on AI. And that’s so scary to me.

Jeremy Epstein: I want an opinion, because “opinion” implies thought and reasoning, and that’s not what AI is capable of.

Rebecca Epstein-Levy: I’ll defer here.

Gilah Langner: Can we just hold some more of the discussion questions?

Rebecca Epstein-Levy: I mean, I am actually happy to segue into discussion here.  I had one more slide, but it was basically – we can just look at it very quickly – the final one, which is just several more examples of these chains of attribution, which, as you can see, these member chains, which, as you can see, go on a bit.

And I think another thing that’s important – I want to notice that they sometimes trace both biological genealogy and what we would call intellectual genealogy or reproduction. And they have ways of accounting for these kind of complex – let’s call them “geometries” – of input and attribution. And I want to note also very briefly, there are actually technological considerations, of a sort, to how these are arranged.

The final source sheet note, Berakhot 10b – is actually one of several seemingly unrelated opinions attributed to this particular memory chain that are stacked kind of on top of another. And sort of the regnant thought here is that because this was primarily an oral tradition, these things that seem disjointed and random on paper are actually arranged in such a way as to make for mnemonics that make oral recitation and recall easier. 

And what I think is interesting, for our purposes, here is the way it goes out of the way to make that attribution accessible, at least to certain people and doing certain forms.