The Torah Accuracy Debate

The question has run through the group since it started in July 2025: can a shliach use AI for Torah without passing on something false? This page follows the argument over time. For the practical guidance it produced, see AI and Torah Accuracy.

Summer 2025: "Don't ask GPT Torah questions"

• An early custom GPT, SinaiGPT, was shared in the group's first week.
• A member announced he was building ChatTorah to answer from real sources.
• Members warned that even search-based tools like Perplexity still "invent nuance."
• A Chabad.org article by its managing editor, "Can I Use AI to Write a Devar Torah?", framed the hashkafic side: AI can help organize and polish your own thoughts, but an AI-generated Torah thought is not Torah, and passing it off as yours misleads listeners.

November and December 2025: the problem gets a name

• At the Kinus AI Day, Dicta showed a Chabad Torah AI built to refuse what it cannot cite. A power user said frontier models would stay the engines, and a Torah-specific retrieval layer was the real advance. See Dicta and Iluy.
• A Likkutei Sichos index GPT got its first test question wrong.
• In December, a member asked ChatGPT to act as a "chassidishe rov" and got invented Rebbe answers with page numbers in Igros and Likkutei Sichos. The diagnosis that stuck: "An LLM is not a database. It's a word generator." / "It's a simulation, not a search engine."
• Asked for an AI that is reliable on Likkutei Sichos, the group admin answered: "No. You always have to double and triple check."

January to March 2026: paste the text in, and write rules

• Telling ChatGPT to pull stories only from Chabad.org failed; the links it gave were unrelated.
• A February poll on using AI for Torah learning: 24 often, 4 seldom, 1 never.
• A member drafted nine principles for AI Torah citation in a shared document: how to search, how to cite, Rishon versus Acharon, no practical psak. It was meant to be pasted into any tool. It circulated for comment and was not finalized. (A practical note from that thread: ChatGPT's custom instructions field holds only about 1,500 characters.)
• A strong critique took hold: even when AI does not invent, its Torah is "slightly off," cuts corners and frames things in secular terms. The opportunity is making existing content easier to reach, such as PDFs and recordings, not generating new Torah.
• An AI-written dvar Torah was criticized as plausible but incoherent. The line that summed it up: "use AI for what doesn't need us, use our Chabad for what does."
• The Bina Melachutit framing appeared. See Moshiach and AI.
• In March the builders' group argued for two days about getting the seforim into clean text at all. See Chassidus Copyright and Access.

Spring 2026: the Rebbe's Torah notebook

• A shared NotebookLM notebook loaded with the Rebbe's Torah became the group's default answer to "is there a Rebbe AI?" One member: "He starts telling me things and I'm like, 'the Rebbe said that?!' It's actually accurate."
• It was not perfect. A member said it retrieves only 10 to 12 sources per question from a large library; another found you can push it past 100 by asking for an answer that needs that many. In June a shliach reported Likkutei Sichos sources from it that did not exist. Others answered that you must click through each footnote and read the page in the source panel.
• In May a shliach gave Claude an unedited maamar (בלתי מוגה) and had it pull out the main points for a bar mitzvah boy to say. "It was the biggest nachas." The pattern: you supply the text, the AI finds the nekuda, a person checks it.

Summer 2026: connectors, and a harder line

• On June 18 the Shluchim Exchange Torah MCP launched, and members began using the Sefaria Texts MCP. The pushback: why not just tell the AI to use only Orthodox sources? The answer: an instruction does not give the AI the text; a connector does.
• AI summaries kept adding things. A summary of a sicha opened with a point the sicha does not make (June); an audio overview of a Yiddish sicha tacked on a lesson about AI (July).
• A harder line appeared in July: "Every limud created by AI will ALWAYS be wrong," so use AI for other work and spend the time saved learning. At a yeshivah administrators' kinus a rabbi said AI "doesn't have mesorah." Others held that generating Torah and finding sources are different jobs.
• In August a member proposed a shared "never post without checking inside" rule against invented Rashis, maamar pages and stories.

September and October 2026: Daiah and Chabad.AI

• Daiah, the publishers' Chassidus AI, went from closed beta to the most-recommended tool for finding the Rebbe's Torah.
• The Akeidah test (September 24). Members asked Daiah and the NotebookLM notebook, "Was Avraham upset at the Akeidah?" One found NotebookLM's answer fuller and Daiah's too short, risking a common misconception. Another found every Daiah citation right and several NotebookLM footnotes wrong. A third called both accurate in different styles. The thread did not settle it.
• The launch (October 15). Daiah went public at chabad.ai. Members praised tap-to-source and listed gaps in the library. One argued that "the frontier will always do better than an inhouse system with RAG"; others answered that most seforim are not available to any general AI.
• A member noted in September that ChatGPT now gives sources far more often than six months before. The group's answer stayed the same: check them.

Where the group stands, October 2026

1. Use a tool that holds the text. Daiah for Chassidus, the Sefaria or Torah MCP inside Claude or ChatGPT, or text you paste in yourself. Not raw model memory.
2. Open every source before you quote it. Read around it.
3. Label AI help on anything public, and link the original.
4. Keep people in the loop. AI finds and organizes; a person who knows the sugya checks and teaches.
5. Ask your rav about halachic questions the tools raise.

Related

• AI and Torah Accuracy
• Daiah (Chabad.AI)
• Chassidus Copyright and Access
• Moshiach and AI
• Shiur Prep
• The Kinus Hashluchim AI Day
• Master Prompts