AI and Torah Accuracy
Why AI invents Torah sources and Rebbe quotes, the tools that cut the problem down in 2026 (Daiah, the Sefaria and Torah MCPs), and a five-minute check before you quote anything.
On this page
Ask a general AI tool where the Rebbe says something and it will often give you a volume and page of Likkutei Sichos that does not exist, or a real page that says something else. It does this with full confidence. The group has documented it since 2025, and it is still true in late 2026 for any AI tool that answers from memory.
Why it happens
An AI chatbot writes the most likely next words. It is not looking anything up unless it has been connected to a library. As a member put it in December 2025: "An LLM is not a database. It's a word generator." Torah sources make this worse:
- Most sifrei Chassidus are not available to general AI tools as clean text. See Chassidus Copyright and Access.
- What the models did read is mostly English websites and translations, not the seforim.
- Torah is layered and compressed. A model flattens it into something that sounds right.
Even when nothing is invented, a member observed in February 2026, AI Torah writing comes out "slightly off": it frames ideas in mainstream secular terms and skips steps.
What changed in 2026
In 2025 the only fix was to paste the text in yourself. In 2026 tools arrived that hold the text and show you where each point comes from:
| Tool | What it holds | Cost |
|---|---|---|
| Daiah (Chabad.AI) | The Rebbeim's seforim from the Chabad publishers, with tap-to-source citations. Public since October 15, 2026. | Free plan; paid from $12/month |
| Sefaria Texts MCP | Sefaria's library inside Claude or ChatGPT, with citations | Free; needs a paid Claude or ChatGPT plan |
| Shluchim Exchange Torah MCP | Torah and Chabad sources plus shluchim tools inside Claude or ChatGPT. Since June 2026. | Free during beta; needs an account |
| NotebookLM with the Rebbe's Torah notebook | PDFs a member uploaded. Renamed Gemini Notebook in July 2026. | Free |
| Iluy from Dicta | A Hebrew Torah AI with a Chabad library setting | Free |
| ChatTorah | Tanach, Shas and halacha, with clickable sources. Private beta. | Free tier; paid from $13/month |
These reduce invented sources. They do not end them. Every one of them has been caught giving a wrong or stretched source in the group.
Failures the group caught in 2026
Real cases, so you know what to look for:
- Footnotes pointing to the wrong sicha. NotebookLM gave Likkutei Sichos sources with parsha and page that did not exist (June 2026), and in September cited a Lech Lecha sicha for Simchas Torah. A member noted that a footnote's page number can be confused with the page of the quote, so scroll the source panel to the actual page.
- Bold headings that look like sources. In September 2026 an AI answer on a Simchas Torah topic came with bold "Likkutei Sichos" headings and page ranges. The member who posted it called the headings a "hack." Formatting is not a citation.
- An added point. An AI summary of a sicha about the lesson of the computer opened with an idea that "doesn't seem to be referenced in the Sichah at all" (June 2026). The real horaah was the second point.
- An invented "modern takeaway." A NotebookLM audio overview of a Yiddish sicha summarized it well, then added a lesson about AI that was not in the sicha and not about the parsha (July 2026).
- A wrong Gemara location. ChatGPT placed a pasuk in the wrong spot in the Gemara, then kept moving it. Claude said it was not sure and sent the member to Sefaria (March 2026).
- The wrong maamar. An outline request produced a summary of a different maamar until the member uploaded a photo of the first page. The same answer mixed up the Mechaber and the Rema (February 2026).
- "Improved" OCR. AI text recognition quietly rewrites the text it reads "to be written better." Tell it to leave every letter untouched, and check anyway (May 2026).
- No sense of time. AI tools lose track of how long ago you spoke and get dates and time zones wrong. Check zmanim and dates against a reliable source; a wrong shkia time can lead to chilul Shabbos.
Before you quote a source or a Rebbe quote
This is the group's current routine, written for a shliach on a Friday afternoon:
- Ask for the exact source. Volume, page and section, or the date of the sicha or letter.
- Open it. In Daiah, tap the citation. In Sefaria, use exact-match search. Otherwise, the printed sefer.
- Read around it. A few lines before and after. For a letter, find the question it answers.
- No source, no post. A quote you cannot find does not go in the newsletter, on the flyer, or in the WhatsApp status.
- Label AI help. If AI wrote a summary of a sicha, say so on the page and link the original.
This applies to Rebbe-quote images that circulate on WhatsApp too. In July 2026 a member suggested a tool that checks quote memes for a source. Until one exists, run the five steps or don't forward it.
A prompt that helps when you do use a general assistant:
I am preparing a class. Answer only from the text I paste below.
After each point, give the line it comes from in quotation marks.
If the text does not say something, write "not in the text."
Do not add sources, stories, or lessons that are not in the text.
Do not change any Hebrew.
[paste the sicha or maamar here]
Does the model matter?
Less than what you feed it. The 2025 advice in the group (Claude for iyyun, Gemini for popular Chassidus, Qwen for Hebrew word work) has faded as models changed every few months. What has held up:
- A model connected to real text beats a stronger model answering from memory.
- Claude has a record in the group of saying "I'm not sure, check Sefaria" instead of guessing.
- One member noticed in September 2026 that ChatGPT now gives sources far more often than six months earlier. Given sources still have to be checked.
- In the October 2026 launch thread for Chabad.AI, members argued whether a general frontier model will always beat an in-house tool. See Daiah for both sides.
AI images in Torah material
- Check AI-made Torah graphics closely. A member warned in October 2026 about a cross appearing on the crown of a Sefer Torah and in other details.
- Members disagree about AI images of the Rebbeim, "accurate or not." The question was left open.
Halachic and hashkafic threads
- AI-written divrei Torah. Chabad.org's managing editor, Rabbi Menachem Posner, wrote "Can I Use AI to Write a Devar Torah?". His view: AI can help organize and polish your own ideas, but presenting AI-generated insights as your own misleads listeners, and an AI-generated Torah thought is not Torah.
- "AI doesn't have mesorah." A rabbi at a July 2026 yeshivah administrators' kinus put it that way. A shliach went further: "Every limud created by AI will ALWAYS be wrong," and the time AI saves should go to learning. Others answered that feeding it specific sources is a different thing from letting it generate.
- Shabbos. See AI and Shabbos.
- For any practical question of halacha, ask your rav. These are summaries of a discussion, not rulings.
Other Torah AI sites shared in 2026
Members posted these; ChabadLabs has not tested them. Descriptions are the sites' own.
- TheWholeTorah.ai: classical texts as structured lessons in several languages, with cited sources.
- MenorAI: a Torah study companion with attributed daily sources.
- Yochai: a Jewish learning companion.
- Daat Torah: Shulchan Aruch study in French.
- Rav Pinson AI: described by the site as "AI-powered conversations with Rav Pinson."
Open questions
- Will Daiah and the other Chabad tools offer connectors, so Claude or ChatGPT can cite them directly?
- Rashi script and handwritten Hebrew (ksav yad) are still read poorly by most AI tools. One shliach reported in October 2026 that a newer personal agent did well at OCR of sichos; nothing systematic yet.
- The community's "nine principles" for AI Torah citation, drafted in February 2026, were circulated for comment and not finalized.