One day a machine will tell you what God wants.
I tasted that future at 2 a.m., my phone glowing in the dark as I scrolled through a WhatsApp group chat. Someone asked: “My hallway lights just switched to motion sensors — can I walk through the hallway on Shabbat?”
Replies flew — jokes, half-answers — until a screenshot appeared. A perfect, sourced ruling. None of us had written it. The verdict came from Rebbe.io, an experimental large language model interface that pulls from a wide array of digitized Jewish texts — Torah, Talmud, halachic responsa, commentaries — to generate instant answers to halachic questions. Think ChatGPT, but trained on the Jewish canon rather than the entire internet.
Curious, I tried it myself. I typed: “If the shared hallway lights in my building are motion sensors, can I still walk through on Shabbat? It’s my only exit.” Nine seconds later the model returned 327 words and four halachic routes that would make it permissible:
- Indirect action (grama) — triggering a sensor unintentionally, such as walking by a motion detector.
- Unintentional activation (davar she’eino mitkaven) — such as moving an object that in turn triggers a light switch.
- Shinui — performing the act in an unusual way, such as pressing a button with your knuckle instead of your finger.
- And finally: Consult your local Orthodox rabbi.
Elegant. Instant. And, if I hadn’t known better, final.
That’s the problem. We are already primed to treat machine answers as a binding ruling, or psak. A March 2025 survey found 52% of U.S. adults are weekly users of AI, such as ChatGPT, Gemini and Copilot; among tech workers the figure is 91%. We let these large language models draft contracts, compose medical letters, even write divrei Torah. Yet few users could describe, even in outline, how these systems actually work.
A large language model breaks text into fragments of words (called tokens) and runs them through a system that calculates statistical relationships among billions of parameters. The eloquence you see in the results is a probability distribution, not a deliberation. In other words, an LLM doesn’t tell you things it knows, it tells you things it thinks are probably true.
In a large language model, there is no inner rabbi weighing sources.
The digitized halachic corpus is heavily weighted toward Ashkenazi voices, leaving Sephardi, Mizrachi and women’s scholarship far behind. Bias in, bias out. Even if equal weight was given across the board, the system is trained to follow a certain logic and come to one conclusion. Jewish law, by contrast, thrives on multi-voice argument — eilu v’eilu divrei Elohim chayim, “these and those are the words of the living God.” AI flattens the centuries-long culture of Jewish debate and discourse into a single, high-probability answer.
Worse, the models “hallucinate” — meaning, they invent. A study of AI tools used in the legal domain shows 58% to 82% fabricated citations when models face difficult queries. The danger isn’t just a wrong answer; it’s a wrong answer delivered with absolute poise.
So what happens when an AI can ace every written exam for rabbinic ordination? Passing a test is trivial for a system that can process an infinite amount of digital information. But ordination is not a multiple-choice credential. It demands human judgment, communal accountability, the moral weight of being answerable. An algorithm can simulate scholarship, but it cannot shoulder responsibility.
The next day I called the friend who’d posed the question in our group chat.
“Did you end up checking with a rabbi?” I asked.
“No,” she said after a pause. “The AI answer was good enough.”
That pause is the real warning. The risk isn’t that a machine will misquote God. The risk is that we will stop asking anyone else — that we will confuse easy access with authority, convenience with wise counsel.
Because the question isn’t whether a machine can tell you what God wants. It’s whether we will still bother to ask a human when it does.