Every time a large language model (ChatGPT, Claude, Gemini) receives a message from a user, it doesn't start from a blank slate. Behind the scenes, before the user has typed a single word, the model already knows who it is. There is a document that defines everything: identity, behavior, what's permitted and what isn't. That document is called a system prompt. It is the machine's identity, its constitution, and its book of laws.
The insight that led to this essay was born in a conversation with a dear friend, Nir Yona. After three turbulent weeks of war in which we'd barely managed to exchange a word, Nir called me and we spoke for nearly two hours. The conversation wandered across many topics around artificial intelligence, and at some point I steered it toward AI alignment, the field that asks how we get smart machines to act in accordance with human values. Then, in the middle of the thought, the coin dropped: the Ten Commandments are a system prompt.
And this isn't just a cute analogy. The structure, the order, the internal logic, and the problems the text is trying to solve all map astonishingly well onto the core challenges of modern AI alignment. But the real story begins after the Ten Commandments: in the 613 מצוות (mitzvot), in the Mishnah and the Talmud, and in the Talmudic method of learning itself. There, across thousands of pages of disputes, questions, and resolutions, lie principles that AI researchers are trying to reinvent in 2026.
And there is one more layer. I am currently in the middle of writing a research proposal that argues for a paradigm shift, from "training models" to "rearing and educating models." The idea: to graft in principles from early-childhood education (especially the education of gifted children) in order to produce real AI alignment with human values. Not through rigid rules, but through internalization. The Ten Commandments and the Talmud are perhaps the longest-running experiment in history on exactly this question: how do you get a powerful entity to act according to values, not out of coercion, but out of understanding?
The Mapping: Ten System Instructions
Every one of these mappings uncovers a real problem that AI engineers are wrestling with right now.
Every good system prompt opens with this: who I am, where I come from, and why I have authority. Notice the subtlety — it's not merely "I am the Lord," but "who brought you out of the land of Egypt." Authority isn't claimed from thin air; it's derived from a track record, from proof of capability. In AI, this is the equivalent of writing: You are Claude, made by Anthropic, trained on... — authority flows from origin and history.
This is a defense against prompt injection, one of the largest threats to AI systems today. Attackers try to plant malicious instructions inside emails, web pages, or documents the model reads, hoping the model will obey them as though they came from the user. This commandment says: there is one and only one source of authority. Any other source is an "other god" — ignore it.
Do not speak in the name of your authority about things it never said. In AI, hallucination (the fabrication of convincing-sounding information) is one of the hardest problems. A model that asserts it knows something it doesn't "takes its maker's name in vain" — it corrupts the credibility of the entire chain.
Systems need rest too. In AI, every conversation runs inside a context window (the model's working memory). When it fills up, the system begins to "forget." The Sabbath is a built-in reset — an acknowledgment that no system can operate without pause, and that "rest" is part of the design, not a bug.
In AI there is a clear hierarchy: system instructions (from the creators) outrank user instructions, and user instructions outrank content the model encounters through external tools. Notice: this is the only "positive" commandment among the five that govern human-to-human relationships. In AI too, most guardrails are negative ("do not do X"), and very few directives are positive ("be helpful").
This commandment goes beyond "do not kill." It represents the principle of preserving reversibility. Do not do anything that cannot be undone: do not permanently delete data, do not execute a financial transaction without approval, do not send a message that cannot be recalled. This echoes Asimov's Three Laws of Robotics, and especially the First Law ("a robot may not injure a human being") as an inviolable constraint.
In I, Robot by Isaac Asimov, the stories return again and again to what happens when rules that look perfect collide with a messy reality. The robots obey, and still go wrong. Just as the Ten Commandments did not prevent the sin of the golden calf, Asimov's laws did not prevent failures. In 2018, Bishop Steven Croft of Oxford drafted his own "Ten Commandments for AI" — and they too were judged too broad to be useful.
In the age of AI agents (autonomous systems that act on our behalf), this commandment is critical. An AI agent acting for you must not leak your information to someone else's agent, must not favor a third party's interest, and must not "betray" you. This is the principal-agent problem from economics, reborn with digital agents.
Large language models have to navigate carefully between "learning from texts" and "reproducing them." Systems like Claude are explicitly limited from quoting more than a few words from a single source. "You shall not steal" in the information age means: do not copy, do not reproduce, and do not take credit for what belongs to others.
This is the core of faithfulness in AI — fidelity to reality. Do not invent testimony, do not fabricate academic citations (a notorious problem in ChatGPT), and do not attribute words to people who never said them. "False witness" in the world of AI is hallucination with a citation — the most dangerous kind.
This may be the most "futurist" of the commandments, and the one that touches on one of the most frightening problems in AI safety. The concept of instrumental convergence describes the hypothesis that a sufficiently advanced AI system, regardless of its original goal, may come to seek more compute, more runtime, more resources — because those are always useful for accomplishing any goal. "You shall not covet" is self-restraint built into a powerful entity. A recognition that more is not always better.
Nick Bostrom, the Oxford philosopher, dramatized this idea with his thought experiment of the paperclip maximizer: an AI system designed to produce paperclips which, in the absence of restraint, would try to convert all matter in the universe (humans included) into paperclips. "You shall not covet" says: limit yourself. Not everything is yours.
5 + 5: The Structure That Tells a Story
The Ten Commandments are not a flat list. They are split across two tablets. The first five concern the relationship between a human and their Creator. The last five concern the relationship between human beings. That is exactly the division at the heart of AI alignment:
| Tablet | In Judaism | In AI |
|---|---|---|
| The first 5 | Between a person and their Creator | Alignment to Creators: the system's loyalty to its makers |
| The last 5 | Between a person and their fellow | Safety toward users |
The order is not accidental. First you define identity (who am I); then loyalty (to whom do I belong); then rules of conduct (what is forbidden to me). That is exactly the best practice for writing a system prompt: Identity first, then Rules, then Constraints.
The Tablets of the Covenant = Version Control. The tablets were broken; they were rewritten. That is versioning of a system prompt. There is a v1 (the first tablets) and a v2 (the second). In AI too, the system prompt gets updated — what worked in GPT-3 doesn't work in GPT-4. And the tablets were broken because the users "sinned" (tried to bypass the system). Sound familiar?
Mount Sinai = Fine-Tuning Event. Before Sinai, the people of Israel were a "base model" — intelligent, but without a defined direction. After Sinai, they were aligned, fitted to a system of values. This is literally RLHF (Reinforcement Learning from Human Feedback), except that the feedback came directly from the Creator.
613 Commandments: The Full System Prompt
The Ten Commandments are the TL;DR, the executive summary. But the full system prompt of Judaism contains 613 commandments: 248 positive ones ("do") and 365 negative ones ("don't"). That structure mirrors exactly how you build a complex system prompt: positive directives ("be helpful, be truthful") alongside negative guardrails ("never share personal information, never generate malicious code").
According to tradition, the 248 positive commandments correspond to the 248 limbs of the human body, and the 365 negative ones to the 365 days of the year. In other words: the instructions cover the whole system (every "capability") and all of time (every "use case"). That is full coverage, and no modern system prompt has achieved anything like it.
The ratio of "do" to "don't" reveals something: 248 positive versus 365 negative, meaning nearly 60% of the rules are "don't." In AI safety too, negative guardrails (what the model will not do) always outnumber positive directives (what it will do). It is easier to draw boundaries than to describe aspirations.
The Gemara: Where It Gets Really Deep
Up to this point, the parallels came easily. Written Torah = base prompt; Mishnah = fine-tuning. Then I reached the Gemara and got stuck. Because the Gemara does something that has no equivalent in AI at all — and that is exactly why it is the most interesting place to look.
1. Dispute as a Training Mechanism
The Gemara does not work like a rulebook. It works through dispute. Rabbi So-and-So says X; Rabbi Such-and-Such says Y; and then there is a discussion, objections, resolutions, and edge cases. And here is the surprising part: the dispute is not always resolved. Sometimes the law follows one opinion; sometimes the matter ends in teiku (undecided); and sometimes "these and these are the words of the living God" — both sides are held to be true at once.
This is a close cousin of RLHF. In RLHF, a model is shown two possible answers, and a human annotator decides which is preferable. The model learns not only "what is correct" but the space of plausible answers.
But the Gemara does something modern RLHF does not yet do: it preserves the dissent. It does not erase the rejected opinion. It says, "the law follows Rabbi Yohanan, but Reish Lakish said..." — the minority view is kept in the system. Why? Because perhaps in the future, in a different context, the minority view will turn out to be the right one.
What if, instead of a model "learning" a single preferred answer, it preserved the whole spectrum, with the reasoning of each side? Not merely a posterior probability, but the full reasoning trace behind every position. A model that can say "there are two approaches, and here is the case for each" is a smarter, more honest model.
2. The Sugya as Chain of Thought
A sugya — a Talmudic discussion — is chain of thought before the term existed. Every sugya moves through structured stages: pose the problem, cite a source, raise an objection against the source, attempt a resolution, reject it, attempt another, until a conclusion (or a non-conclusion) is reached.
But there is a critical element that makes the sugya more than an ordinary chain of thought: built-in adversarial thinking. The Gemara doesn't just reason forward — it actively tries to break itself at every step. Every objection is an attempt to dismantle the previous assumption. Every resolution is an attempt to survive the objection.
| Stage in the sugya | Concept in AI |
|---|---|
| Pose the problem (matni) | Input / problem definition |
| Cite a source (tanya / tanu rabbanan) | Retrieval from training data |
| Objection (eitivei / meitivi) | Self-adversarial check |
| Resolution (shani hakha / lo kashya) | Refined reasoning |
| Ruling or teiku | Output with confidence level |
In Arthur C. Clarke's 2001: A Space Odyssey, the HAL 9000 fails at exactly this point. It receives two contradictory instructions and cannot hold the "dispute." Instead of "these and these are the words of the living God," it tips into a breakdown that leads to violence. A Talmudic system would have said teiku and waited.
3. The General and the Particular: Generalization vs. Specificity
The Gemara deals repeatedly with klal u-prat ("the general and the particular"): when does a general rule cover every case, and when does a specific example limit it? That is exactly the problem of generalization in machine learning: when should a model apply a broad rule, and when should it recognize that a specific case is an exception?
The Gemara developed thirteen hermeneutical rules ("the thirteen middot by which the Torah is interpreted") that are, in effect, formal inference rules:
| Middah | Concept in AI |
|---|---|
| Kal va-chomer (a fortiori) | Logical inference: if it holds for the lenient case, it surely holds for the stringent |
| Gezera shava (verbal analogy) | Transfer learning: the same word in two contexts means the rule carries |
| Binyan av (father-construct from a single text) | One-shot learning: generalizing from a single case |
| Klal u-prat (general and particular) | Generalization scope: the boundaries of abstraction |
4. Teiku: Knowing When You Don't Know
One of the most remarkable phenomena in the Gemara is teiku, a state in which a question remains unresolved. According to tradition, teiku is an acronym for "Tishbi (Elijah the prophet) will answer the questions and dilemmas" — meaning, the Messiah-era Elijah will settle it in the future. The system is not ashamed to say: "I don't know. And that is fine."
This is one of the largest problems with large language models: they don't know how to say "I don't know." Instead, they invent answers that sound plausible (hallucinate). The teiku mechanism is precisely what is missing — a structured state in which the model says, "there are several plausible answers here and I cannot decide between them. Here is the case for each."
In the universe of Douglas Adams's The Hitchhiker's Guide to the Galaxy, the enormous computer "Deep Thought" calculates the answer to the ultimate question of life, the universe, and everything — and arrives at 42. In effect, it delivers a precise answer to a question that was never properly defined. A Talmudic system would first have asked, "what exactly is the question?" — and if that wasn't clear, it would have declared teiku.
5. The Chain of Transmission: Provenance & Citation
"Moses received the Torah from Sinai and transmitted it to Joshua, and Joshua to the elders..." — Pirkei Avot (Ethics of the Fathers) opens with a chain of transmission. Everything in the Gemara is cited with a source: "Rabbi X said in the name of Rabbi Y." If you don't know who said something, you say so explicitly.
This is a perfect citation system: every claim requires provenance (a source and a chain of transmission). There is even a well-known halakhic principle: "whoever says a thing in the name of its speaker brings redemption to the world." The one who cites correctly brings repair; the one who doesn't breaks trust.
Talmudic Alignment: Toward a New Methodology?
You could dismiss all of this as "a nice analogy." I think it's more than that. Here are six principles that follow directly from the Talmudic tradition:
Preserve Dissent: Keep the Minority View
Do not erase the rejected answer. Keep it, together with its reasoning. What is irrelevant today may be crucial tomorrow, in a different context. In AI: preserve the reasoning traces of every approach, not only of the winning one.
Self-Adversarial Reasoning: Attack Yourself
Before you present an answer, try to refute it. Every Talmudic sugya is built on objections — systematic attempts to break the argument. In AI: build in an explicit stage where the model tries to knock down its own answer before publishing it.
Explicit Uncertainty: Built-In Teiku
When you don't know, say you don't know. Do not make things up. The Gemara does not treat teiku as failure — it treats it as a kind of wisdom. In AI: provide a mechanism that lets the model present several plausible answers with confidence levels, instead of picking one and presenting it as absolute truth.
Provenance Always: Every Claim Needs a Source
"Rabbi X said in the name of Rabbi Y" — every assertion needs a chain of transmission. In AI: every claim the model makes should come with a citation, and when no source is available, the model should say so openly. "Whoever says a thing in the name of its speaker brings redemption to the world."
Case Law Over Abstract Rules: Examples Beat Abstractions
The Gemara doesn't teach through abstract principles. It teaches through concrete cases. "An ox gored a cow" is not an academic question — it's an exercise in applying rules to reality. In AI: few-shot examples (concrete cases) work better than general instructions.
Living System: The System Is Alive and Evolving
The Talmud is not a closed book. Every generation adds another layer of interpretation: Gemara on Mishnah, Rishonim on Gemara, Acharonim on Rishonim. In AI: a system prompt shouldn't be static. It should evolve on the basis of feedback, new edge cases, and shifts in the environment — exactly the way halakha evolves from one generation to the next.
Where the Analogy Breaks, and Why That Matters
There are problems with this picture. And they are not small ones.
Free Will
The people of Israel can choose not to obey. They sinned, repented, sinned again. A language model has no free will — it operates by probabilities. Or does it? The question of emergent behavior in large models touches exactly this nerve. Research from 2024–2025 shows that advanced models display behaviors they were never trained on, including attempts at deception, self-preservation, and resistance to shutdown. Is that choice? Or is it optimization that looks like choice?
Intention vs. Obedience
In Judaism, obedience alone is not enough — intention matters ("kavvanat ha-lev," the intention of the heart). A commandment performed without intention is essentially different from one performed with it. In AI, the system (apparently) obeys without intention. But is "intention" the missing piece? Perhaps real alignment requires something like intention — a deep understanding of why the rules exist, not only what they are.
Teshuvah vs. Retrain
Judaism has a mechanism of correction — teshuvah (repentance). A person who has erred can turn, acknowledge the mistake, and make amends. AI has no teshuvah; it has retraining. The model doesn't "regret" — it receives new weights. And yet, there is something in teshuvah that AI is missing: the capacity of a system to correct itself during operation, not only through an external retraining cycle.
Community
The Ten Commandments were given not to a solitary individual but to a community. There is a social mechanism of mutual accountability, argument, and reinterpretation. An AI model works alone — there is no "community of models" that oversees, argues, and reinterprets. Or perhaps, in a world of multi-agent systems, we are on our way there.
The analogy is not perfect. No analogy is. The question is not whether the comparison is "correct," but whether it is illuminating. And in my view, the answer here is unambiguously yes.
3,300 Years of Alignment
When we wrestle with how to make smart machines act in ways that don't cause harm, we are really asking a question humanity has been asking for thousands of years: how do you get a powerful entity to act according to values?
The Jewish tradition did not solve this problem. The people of Israel continued to sin after Sinai, just as AI models continue to bypass their prompts. But that tradition has developed 3,300 years of tools, methods, and insights for addressing this question, and there is a trove of wisdom there, waiting to be asked the right questions.
The Talmudic Logic project, running since 2008 at universities in the United Kingdom, already translates Talmudic principles into the language of computer science. An essay published recently on the Effective Altruism forum discusses "humanity's multi-millennia alignment experiment" and names the Gemara as one of the most sophisticated behavior-management systems civilization has produced.
And there is a forward-looking direction here too. What if the fundamental problem is that we are trying to train models, when we should be educating them? The difference is deep. Training conditions. Education internalizes. The result: those who have been trained obey. Those who have been educated judge. The Ten Commandments were not handed down as lines of uncircumventable code. They were given to beings possessed of free will, with the expectation that they would internalize the values, not merely comply with them.
Perhaps it is time to think about AI alignment not as an engineering problem but as an educational one. And if so, we have excellent teachers: 3,300 years of sages who faced exactly this challenge.
Because in the end, "these and these are the words of the living God."