You are talking to an AI. A language model writes the answers on Sh'elah, and nobody reads or checks an answer before it reaches you. This page explains how that works, where it can go wrong, and what the AI does not do.
1. Which AI models answer your questions
Sh'elah sends your question to Google Gemini first. If Gemini is not available, Anthropic's Claude answers instead. After Gemini fails twice in a row, Sh'elah pauses it for 45 minutes and uses Claude alone, then tries Gemini again on a later question.
| Provider | Role |
|---|---|
| Google Gemini | First choice for every answer |
| Anthropic Claude | Answers when Gemini is not available |
An answer does not say which of the two wrote it. Source passages that come without a translation can be translated by separate online services (MyMemory and Google Translate), not by these models.
Sh'elah does not use your questions or answers to train AI models without your explicit consent. What each provider receives is listed in the Privacy Policy, §4 and §9, and each provider's own terms cover what it does with that data.
2. Where the answers come from
Before the model writes anything, Sh'elah searches for sources and gives what it finds to the model along with your question. The technical name for this is retrieval-augmented generation, or RAG. The search covers:
- Sefaria, for the text of the Torah, Talmud, halachic codes and commentaries.
- Halachipedia, a wiki of halachic sources and discussion.
- Community customs, from a set of entries that Sh'elah keeps in its own database.
- Hebcal, for dates, holidays and zmanim.
- Wikipedia, only as a last resort when the Jewish sources have nothing on the question.
- A short list of Orthodox sites (Chabad.org, the OU, Star-K, the CRC and a few others), only as a last resort for a real-world fact no source covers, such as whether a product is kosher. The AI sends a short search phrase to Anthropic's web search, which can reach only those sites. The answer then says it used the web and shows each page as "From the web". You can turn this off under Web search in the chat menu.
The model also sees the earlier messages of the same conversation. If you are signed in, it sees short summaries of your two most recent earlier questions as well. It does not see your name or email.
Starting from real texts keeps answers closer to the sources than a chatbot answering from memory. It does not make them correct. The search can bring back the wrong passage, or one that only touches the topic, and the model can still misread a good one.
3. Where it can go wrong
Language models sometimes state something confidently that is wrong, made up, or not in the sources they were given. On Sh'elah that usually looks like one of these:
- A citation that looks real but points to the wrong place, or to a text that does not exist.
- One opinion presented as the only one, when the sources disagree.
- A passage that matches the topic but not the question.
- Meaning that shifts when a Hebrew or Aramaic text is put into English, or the other way around.
Some things help. The model is told to cite only sources it is sure exist. A check on the finished answer catches explicit content and leaked instructions, though not factual mistakes. Citations link to the source text where Sh'elah can match it, so you can read the original yourself. These reduce the risk and do not remove it. Open the source before you rely on an answer (see Terms of Service §6).
4. What the AI does not do
- It does not give a psak halacha. An answer is study material, not a ruling.
- It does not replace your rabbi or posek.
- It does not give legal, medical, mental-health or financial advice.
- It is told never to claim to be a rabbi or a person.
- It does not answer when a message sounds like an emergency, self-harm or abuse. Sh'elah shows a fixed message instead, pointing to emergency services, a crisis line or a trusted adult, and the model is never called. Everyday health questions with a halachic angle, such as fasting with a medical condition, are answered like any other question.
- It declines hateful, violent or explicit requests, and subjects with no Jewish angle.
5. Your rabbi makes the call
Every answer comes with a note: AI-generated educational information, not a halachic ruling, ask your rabbi for practical guidance. The note stays on screen and cannot be turned off. Sh'elah is meant to help you find and understand sources faster. If you plan to act on an answer, your rabbi decides.
6. Age and sensitive topics
You must be at least 13 to use Sh'elah, and at least 16 in the EU and UK (see Terms of Service §3). The AI is told to assume the reader could be 13. On sensitive subjects such as niddah, mikveh and intimacy, it explains the halachic framework in clinical language, cites sources, and sends you to a rabbi, teacher or parent for practical details. If an answer comes back with explicit content anyway, a check replaces it with a short explanation and a referral.
7. Contact
Questions about this page, or an answer that seemed unsafe, wrong or inappropriate: write to the address below. The thumbs-up and thumbs-down under an answer reach me too.
Contact: Akiva Yevdayev
Email: contact@shelah.org
See also the Terms of Service (§6–§8: AI output, not professional advice, assumption of risk) and the Privacy Policy (§4 and §9: what is sent to AI providers).