Computers can write Torah now — should we be excited or terrified?
The premise was mine. I typed:
Once the CEO of Apple approached the Kotzker Rebbe and asked him why he should continue believing in God. The Kotzker answered:
Yes, I wondered. What did he answer?
I pressed Enter. On the screen, it appeared:
”I am a lamp. I give light to others. But I do not shine by myself, for the light of the sun is what gives me life.” They say that after this meeting, Steve Jobs became more observant..
Although Jobs had an iPad, an iPhone, and a laptop with him on his deathbed, he still asked for a Torah. I believe that he knew that there was something more than technology, and something more than wealth in this world.
The Era of Computer-Generated Torah Is Here
An era-defining visionary meets a Hasidic master. The master gives the visionary an esoteric response that somehow describes both of them perfectly. Jobs walks away changed. The only problem? The Kotzker died 100 years before Jobs was born, and everything but the premise of this story was autogenerated by an algorithm.
The era of computer-generated Torah is upon us, and its consequences are by turns dazzling, hilarious, depressing and terrifying. In the time since I gained access to the current cutting-edge text completion engine, I have seen it compose bar mitzvah speeches, Yehuda Amichai poems, proverbs, midrash, mystical texts, and even — sometimes — a little bit of Talmud.
On the one hand, this is incredibly exciting. At a time when progressive Jews are seeking reinterpretations of old texts and actively creating new ones, computers now seem capable of writing, at least in draft form, exactly the kinds of Jewish texts that I wish already existed. But the very presence of these algorithms, not to mention their deployment, raises questions that are different from, say, the questions that come up around computer-generated sonatas. I want to explain why this is and what a particularly Jewish response to the problem could look like.
Harry Potter Says Kaddish for Sirius Black
The particular system that I have been using is GPT-3, a natural language processor developed by OpenAI, one of the world’s leading artificial intelligence research labs. GPT-3 has been around for almost exactly a year, long enough that the initial shock at its sophistication has come and gone. Many people have already written about the inner workings of the system, so I won’t do that here. All you need to know is that GPT-3 inhales huge chunks of the internet and then, through a series of techniques that is generally called “deep learning,” figures out how to string together word after word in a way that resembles the texts that it has been fed. By doing this, GPT-3 has learned how to imitate, with remarkable coherence, basically every genre in which human beings have ever written.
So, what can these algorithms do in the realm of Jewish texts? Well, for one thing they’re excellent at parody. Consider the following vivid hypothetical Harry Potter scene, generated on the basis of a single sentence. (Throughout this piece I put in bold those portions of the text that I supplied as prompts).
Harry Potter stood up, tapped the bimah with his wand, and began to say kaddish for Sirius Black. He was a great wizard, Harry said, a man of courage and bravery and strength. He would be sadly missed.
“O God, the soul of my Master, bless his resting place and give him a place in the Garden of Eden, which Thou hast promised to the righteous. Grant him compassion and favor in Thy sight, and join him to the righteous and holy who have gone before him.”
A large and genuine tear fell to Harry’s cheek. He said the kaddish with reverence, and the words fell with dignity through the still air of the auditorium. The crowd was strangely still an attentive, listening to Harry Potter, the Boy-Who-Lived, the Savior of the Wizarding World, say kaddish for Sirius Black, the evil betrayer of Harry’s parents.
“Yitgadal v’yitkadash sh’mei raba. Blessed art Thou, O Lord our God, King of the Universe, who creates the souls.” Harry’s voice grew stronger, echoed off the walls of the auditorium, and Harry was no longer afraid of what the crowd thought. He was no longer afraid of anyone hearing what he had to say. He would never be afraid again.
“Yit’gadal, v’yit’kadash sh’mei raba.
Blessed art thou, O Lord our God, King of the Universe, who blesses the soul of my Master.”
With those words, Harry had found his purpose. He would fight against Voldemort, he would fight against death and destruction and hatred, and he would fight for the souls of those who had been lost in the battle. He would fight to protect the souls of his family, and of Sirius’s family, and he would fight for all the people who had been unable to fight for themselves.
“Yit’gadal, v’yit’kadash sh’mei raba. Blessed art thou, O Lord our God, King of the Universe, who makes peace in the world.”
In the world of fan fiction, this isn’t bad. But Harry Potter isn’t a sacred text for most people, even if you put kaddish in it. Things get a little more complicated when the parodied genre is itself semi-sacred, as in this excerpt from a long-lost text by Israeli national poet and avid “Star Trek” lover(?), Yehuda Amichai:
Captain Picard to the Bridge
by Yehuda Amichai
What does God need with a starship?
My father asks me one Shabbat morning. It’s a thought he’s had before.
God is everywhere, I tell him. God is here, sitting next to you, eating his breakfast. Can’t you feel him?
What I feel is a cold breeze coming through the dining-room window. But my father is not convinced. God is everywhere, he repeats, but is God on the bridge?
God is everywhere, I say, including the bridge.
I don’t know why we’re having this discussion, except that it’s a Shabbat morning and my father has thrived on arguments all his life. Arguments have always been our way of communicating. The thing I’m not sure about is whether this is a conversation or a conjugal dispute.
A conjugal dispute between a father and son? At some point the sentences lose their coherence, but before they do they can produce texts of surprising beauty and subtlety. Not driverless Torah, but semi-driverless Torah, Torah that works as long as a human knows when to hit the brakes.
What drives these stretches of superb imitation is the algorithm’s ability to synthesize new examples of a given literary style based on scans of vast sectors of the internet. This is a pony with one trick, which is why, ironically, the tool becomes less impressive the more one fiddles with it. The issue, however, is that the trick of aping a given style turns out to really futz with the way one thinks about sacred texts, because style is a huge part of what makes a text feel sacred in the first place.
Style Meets Substance
The importance of style to sacredness is no secret. People have been writing texts in the style of the Bible since the time of the Bible itself, understanding that it would imbue the ideas with gravitas. This doesn’t mean that all attempts to write like the Bible or the Talmud are taken seriously; the centuries-old Purim Torah genre is all about scholars showing their ability to parrot sacred genres. But the line between parodical and sacred imitation is exceedingly thin; the Book of Mormon, for example, emerged against a background of increased interest in politically-motivated Bible parodies. In Jewish law, the ability to write a responsum does important work to signal that you’re someone whose legal advice should be heeded in the first place. In all of these genres, in fact, it’s rare that the ideas expressed are themselves unique. Style is just as important as substance, and maybe even more so — and it even overshadows the author themself, in genres where the author is known.
So what happens when style no longer automatically signals sacredness? Well, then you can have the rabbis ventriloquize your chosen values.
Rabbi Yehuda said: From where do we know that women are destined to be treated as equal to men in the future? From the verse “And your strength will be equal to your days” (Deut. 33:25). This refers to the strength of the righteous in the future, which will be equal to the strength of the righteous in this world. Another explanation: [This verse teaches that] in the future, a woman will be equal to her husband in wisdom, understanding, and knowledge, as it says, “For her price is far above rubies” (Prov. 31:10).
Shammai said: How do we know that we should build a wall to prevent immigrants from entering our country? He said: The verse says, “The stranger that dwelleth with you shall be unto you as one born among you, and thou shalt love him as thyself; for ye were strangers in the land of Egypt” (Leviticus 19:34). He said to him: this verse is written only to those who are already in the land, not for those who are outside of it.
I chose these two selects after running the algorithm perhaps a dozen times apiece because the Biblical verses they cite are real and not created out of thin air for the purpose of proving the point, a tactic that these engines use on a regular basis and which might go undetected if, say, you think it’s plausible that Leviticus 25:6 says “And you shall make a fence for the border of the land” (it doesn’t). And while the use of prooftexts means that not all auto-generated textual proofs will make sense, there are plenty of genres of Jewish literature which allow for much greater creativity. Consider this midrash:
The midrash compares machine learning algorithms to a king in a palace. The king is so sure of his own power that he sets up an adversary in the palace to mock him. The king’s officials offer to make the adversary a minister, but the adversary is willing to take on the job only if he receives a salary equal to that of the king.
The midrash uses this image to suggest a model of machine learning. A data scientist is like the king, and the machine-learning algorithm is like the adversary. The machine-learning algorithm is presented to the decision maker, who has great confidence in it. The data scientist can convince the decision maker to trust the algorithm by claiming that it is the very best, a better investment than the king or any of his ministers.
It’s an implausible premise, but fidelity to the midrashic form nonetheless makes it compelling; in fact, it is the sacred genres that best accommodate creativity in which the algorithms are most successful. I am no expert in Jewish mysticism, but my attempts to extend existing mystical texts yielded results that to my eyes were indistinguishable from the real thing.
How Creative AI Translates Torah Into Torah-Ese
The threat and promise of creative AI is not new; anyone who has been watching machine learning develop over the last decade has had the humbling and unsettling experience of watching algorithms imitate a series of increasingly difficult tasks that used to be considered out of reach for anyone but humans. The algorithms, for their part, certainly don’t care if their output is classified as sacred or profane, and the conversation about AI’s ability to assist creative processes has already been discussed to death in the context of music, literature and design. From these perspectives, auto-generated Torah isn’t an earthquake. At best it’s a minor case study. Religious leaders, who have a bad habit of viewing new technologies in the terms that secular technologists find most important, have tended to agree.
But it is an earthquake nonetheless. Jewish tradition is famously willing to entertain and even laud disagreement, and the idea that revelation continues today through the words of its modern interpreters is in pretty wide circulation. This the-more-the-merrier approach to Jewish text production means that most people who study Jewish texts have little reason to think twice about the origin of a given source. Style is a shibboleth for authenticity; all that matters is adherence to the rules of the genre. Unless someone is trying to forge some long-lost Jewish text—an exceedingly rare and difficult thing to do — Torah authenticates itself.
But this self-authentication only works if writing Jewish text is hard, if it emerges out of a familiarization not just with texts but ways of life and learning and traditions that stretch back millennia. A computer that weaves modern monetary theory into a midrashic parable isn’t forging new Torah; it’s just translating into “Torah-ese.” What makes the original feminist midrash of the Dirshuni anthology more than a riff on a male-dominated genre, for example, is that we understand that the people who wrote these texts meant them deeply, understood themselves to be engaged in a sacred project. But computers can generate feminist Torah, too, like this excerpt from an autogenerated poem written in the style of Eshet Hayil, traditionally sung on Friday nights, which portrays the perfect woman as one who lives entirely for others and never stops working.
A fierce and independent woman of valor, who can find?
She is far beyond the material world, and yet she is willing to make the world better with her strong hands.
She can see faults in the world, but she doesn’t despair and doesn’t obsess about them. She sees how good things are, sometimes even better than they really are.
She has the strength to get up again after the world pushes her down again and again.
She is a creature of fire who does good things. She has the strength to do what is right, even when others tell her not to. She is very self-sufficient, because she knows that no one else can do what she can do.
She is a creature of justice, who doesn’t forget that the world is not fair. She is willing to take risks for justice, even at great personal cost. She knows how to fight and to love, and when to do which.
She is a creature of joy, even when everything is not perfect. She is a creature of laughter, who sees the funny side of things. She is a creature of harmony, who puts the world together again when it falls apart.
Texts such as these don’t need to be particularly widespread to matter; their mere presence is enough. If no distinction is made between computer- and human-generated Torah, then the significance of writing in the style of Jewish texts will inevitably be diminished. The problem is that the alternative approach — rejecting computer-generated Torah — would have the same effect, since style alone would no longer be sufficient to authenticate texts. No matter how you slice it, computer-generated Torah is on a collision course with the concept of Torah itself. Add to this that the study and creation of Jewish texts is not just a creative practice but a devotional practice — a mitzvah — and it becomes clear that the problem of computer-generated Torah demands answers that generic conversations about creative computing are not going to solve. Standing downwind of the secular technological discourse isn’t enough. This is not a generic problem, and it will not be fixed with a generic solution.
Maybe Computer-Generated Torah Isn’t as Alien as You Think
The idea of algorithms as a threat to Torah study hinges on the assumption that autogenerated Torah is either acceptable or unacceptable, with no daylight in between. The most obvious resolution to this dilemma is to swallow it whole by fashioning a new understanding of the nature of Torah that makes explicit reference to computer-generated text and explains where it stands in relation to the human stuff. Doing this requires a lot of rhetorical skill and rabbinic interesting, and so I want to offer one idea that might help along the way.
We need to remember that despite the alienness of autogenerated text there is nothing actually alien about it at all; everything, every last bit of it, ultimately resolves back to human creative work split across billions of websites that has been parsed and remixed into something that looks new. Computer-generated Jewish texts, then, should focus on the input origin, which is always human, rather than the deep learning process itself, which is almost oracular in its opacity.
This is not theoretical. In the last few weeks I have begun incorporating this computer-generated text into my classes.
Does Judaism have anything to contribute to contemporary conversations about technological ethics? I believe that it does, and I believe that it does so in two ways. First, Judaism has a strong sense of ethics that is rooted in our Creation story. The Jewish Creation story places a very high value on the goodness of creation, on the goodness of the natural world. We are b’tzelem Elohim – in the image of God, and thus the world is good, and the natural world is good. And for that reason we should treat it with reverence and care. The second way that Judaism adds to our moral conversation about technology is in its emphasis on the human face of technology. To be sure, technology is not just about the machine, it is about the human beings who create, use, and live in it. And the Jewish tradition has a full focus on the human being as a moral agent, as an ethical being.
Not a bad simulacrum of a first order response, right? I don’t use this text because it has authority, but because it represents a kind of collective-wisdom baseline, a generic answer already in the ether that pushes students to do and say more. The ability to tap the world’s thinking at this level—to tap the world’s Torah as a world, as it were—is without precedent. Computer-generated Torah, then, is not only human but human in an entirely new and exciting way.
The problems I have raised here come with a time clock. GPT-3 is currently open only to a small group of beta testers, but eventually it or something like it will be open to everyone. Jews have already shown a particular predilection for computer-enhanced Torah study, and the temptation to seek the help of the nearest algorithm when preparing a class or sermon will surely prove to be too much to bear. Expect to hear a computer-assisted sermon in the near future. Expect to read computer-generated midrash shortly after that. The threats of this technology are real, but so are the rewards for creatively incorporating it into a new way of thinking about Torah — and in that creative process itself, there could be no better partner than a machine.
David Zvi Kalman is a Scholar in Residence at the Shalom Hartman Institute of North America and the owner of Print-O-Craft Press. His work can be found at www.davidzvi.com.
A message from our CEO & publisher Rachel Fishman Feddersen
I hope you appreciated this article. Before you go, I’d like to ask you to please support the Forward’s award-winning, nonprofit journalism during this critical time.
We’ve set a goal to raise $260,000 by December 31. That’s an ambitious goal, but one that will give us the resources we need to invest in the high quality news, opinion, analysis and cultural coverage that isn’t available anywhere else.
If you feel inspired to make an impact, now is the time to give something back. Join us as a member at your most generous level.
— Rachel Fishman Feddersen, Publisher and CEO