The Torah Accuracy Debate

Not a single conversation but a recurring through-line spanning the whole chat: what does it take to use AI for Torah without breaking something sacred? See AI and Torah Accuracy for the distilled positions; this page tracks how the conversation unfolded over time.

"Can I use AI to write a devar Torah?"

The earliest framing came from a widely shared Chabad.org article, Can I Use AI to Write a Devar Torah? The group's reading of it:

• AI is fine for brainstorming, editing, and organizing.
• AI-generated Torah is problematic. The article argued that when we innovate in Torah we become partners with the Divine, and questioned whether a computer can create that unity.
• There is an ethical concern of geneivas daas if readers believe the words are the author's own thought.

The hallucination drumbeat

The most persistent complaint: a general chatbot cites Likkutei Sichos, volume and page, and the page does not exist, or the sicha does not say that. Members reported chatbots inventing whole Rebbe answers with fabricated citations to Igros and Likkutei Sichos. The diagnosis crystallized into a working theory:

"LLMs are 'word generators,' not 'databases.' Without structured data, all of Igros indexed, this problem will persist for years. Unless the model providers give access to all the texts and structure them, hallucinations for Torah references will continue."

Building substitutes

Rather than wait, the community started building alternatives, all of which appear in the Toolbox:

• Dicta's Mekorotai, a purpose-built rabbinic AI that refuses what it cannot cite, demoed at the Kinus AI Day but not yet shipping reliably.
• NotebookLM loaded with the Rebbe's Torah, a shared notebook of Likkutei Sichos, Igros Kodesh, Toras Menachem, and Maamorim. One member's reaction: "He starts telling me things and I'm like, 'the Rebbe said that?!' It's actually accurate."
• A plain-text Chassidic corpus that you paste into the model rather than asking it to recall.
• A curated maamorim app with vetted summaries.
• Existing Likkutei Sichos apps and Sefaria.

The "9 Principles" document

A community member began drafting plain-language rules for AI Torah citation, meant to be pasted into whatever tool a shliach already uses:

"I'm working on a set of principles and detailed rules, written in plain language, not code, that you paste into whatever tool you already use. No downloads, no new apps. The rules cover search methods, citation formats, verification steps, and domain-specific instructions for different types of texts."

Circulated as an open document for community feedback, it kept iterating, and is referenced as the closest thing the community has to a system prompt for Torah honesty.

The "Bina" framing

A parallel thread introduced a Chassidic distinction: bina, understanding and processing, without chochma, the sudden flash of insight. One member wrote a full Shabbos-table drosho tying Parshas Mishpatim's Eved Ivri to the nature of AI, ironically composed with the help of a chatbot. Philosophical rather than action-oriented, the thread produced the chat's most-quoted hashkafic framing:

"The machine has the 'bina,' but we have the 'neshama.'"

"Read primary sources, not English translations"

A hardening consensus through early 2026: even when a model sounds right, it is regurgitating English translations of secondary commentary rather than reasoning from primary sources.

"AI has very little access to actual Torah sources. It's skimming basic English sites and regurgitating chabad.org back at you."

"Torah doesn't fit the shape that AI is expecting to read. Layered, subtle, nuanced. The process is lost on it."

Lessons from a magazine project

A member documenting a ten-page Pesach magazine built with an AI assistant surfaced several Torah-relevant lessons:

• Set golden rules at the start of the chat: do not edit my text, Hebrew is right-to-left, no hyphens.
• Testing the model on a maamor outline produced the wrong sicha attribution until it was shown an image, and even then it was unreliable. The takeaway: do not just trust the model.
• Cross-reference everything that names a source.

Specialized translation

Testing eight models for Hebrew translation, the group found one, Qwen, best for pure translation work: etymology, shoresh, and word-by-word transliteration. It was added to the recommended stack for Torah work specifically.

Curricula and shiur prep

The conversation eventually shifted from "can AI translate Torah?" to "can AI organize Torah?", and the answer settled as yes, with vetted sources. A general assistant became the default for sixty-minute shiur outlines with discussion questions, teacher guides, and student handouts. See Shiur Prep. The catch: the input must be vetted text from a curated corpus, NotebookLM, or a typed sicha, never the model's own memory.

The bar-mitzvah maamar

A closing example that captures the working pattern. A member uploaded a maamar (בלתי מוגה) to an AI assistant, which extracted the nekuda and presented it in a way a thirteen-year-old could understand:

"It was the biggest nachas to see a boy from my community chazering a maamar at his bar mitzvah."

The pattern works when you supply the maamar. The AI extracts the nekuda; the human verifies and teaches.

The chat's working answer

By mid-2026 the group's working answer for AI and Torah:

1. Curate the corpus. Use a vetted plain-text corpus, NotebookLM with vetted sources, a curated maamorim app, or Sefaria, never raw model memory.
2. Match the model to the task. A strong general assistant for iyyun and shiur structure; Gemini Pro or Qwen for translation; NotebookLM for citation-grounded Q&A.
3. Always cross-check citations. Do not just trust the model.
4. Buffer layers between AI and talmidim. Curated apps instead of raw AI-generated content.
5. Document principles and share community prompts.
6. Keep the hashkafic framing: AI provides bina; humans provide chochma and neshama. AI does not replace shluchim; it extends their reach.

Related

• AI and Torah Accuracy
• Shiur Prep
• Hebrew and Yiddish
• The Kinus Hashluchim AI Day
• Master Prompts
• Toolbox, for Mekorotai, NotebookLM, and the Chassidic-corpus tools.