AI on Sefaria

Sefaria’s mission is to take the People of the Book into the digital age, and it is our responsibility to explore how to best leverage technologies to create new ways to access, search, discover, and understand Jewish texts. Our founders’ foresight in making Jewish texts machine-readable continues to enable Sefaria — and others — to take advantage of technological advancements such as artificial intelligence (AI) and, specifically, large language models (LLMs).

Pirkei Avot 1:1
Moses received the Torah at Sinai and transmitted it to Joshua, Joshua to the elders, and the elders to the prophets, and the prophets to the Men of the Great Assembly. They said three things: Be deliberate in justice, establish many students, and make a fence around the Torah.

Sefaria is part of this long and storied history of transmitting Jewish tradition, holding sacred the texts in our library. As we experiment with and introduce AI-generated content to the Sefaria library, we are committed to specific guardrails, effectively making a fence around the Torah. We have established the following guidelines for Sefaria’s use of AI:

  1. Transparency: Sefaria will always be transparent about how, where, why, and when AI is being used. This includes clearly marking pages with an AI icon when they include AI-generated content. We currently use one icon to designate AI-generated content.
  2. Learning First: AI content on Sefaria will aim to act as a learning resource and not as a replacement for your rabbis, halakhic advisors, or teachers.
  3. Continued Evaluation: Sefaria will continually monitor new technologies and our use of them to ensure we serve our users in the best possible way.
  4. Feedback: Sefaria is proactively soliciting user feedback regarding AI-generated content. Users are invited to provide corrections and feedback using a dedicated AI feedback form.

In addition, Sefaria agrees to adhere to the Rome Call for Artificial Intelligence Ethics, which has been adopted by organizations around the world, such as the Chief Rabbinate of Israel, Yeshiva University (among other academic institutions), the Rabbinical Alliance of America, and major corporations like Microsoft, Google, IBM, and Cisco.

AI on Topic Pages

You will now find AI-generated content on approximately 1,000 of our 5,000 Topic pages. These pages are marked with an AI symbol (AI) on the top right corner.

On each page, users will find a series of sources (individual selections of texts from the library) related to the topic. These sources have been curated by AI to provide a broad selection of relevant and important sources. Each source includes an AI-generated heading and introduction to make it more digestible to learners of all backgrounds. Editors have given some level of review to both of these elements to ensure accuracy and relevance. Many AI-generated Topic pages also include an overview paragraph written by our learning team.

The basis for this AI work was developed by Sefaria’s learning team, which includes seasoned scholars, editors, and translators, in collaboration with a team of engineers dedicated to this specific project. After developing the framework for these pages, our learning team created detailed criteria and prompts for an AI algorithm to reproduce their content creation process. Of the 15,500 source introductions on AI-generated pages, 14,000 were reviewed and edited by Sefaria team members.

Should you encounter anything that seems amiss, we encourage you to let us know by filling out this short feedback form. Your feedback is crucial to our ongoing efforts to ensure accuracy and uphold our commitment to quality on Sefaria.

ANATOMY OF AN AI-GENERATED TOPIC PAGE

Anatomy of a Topic Page

AI Research & Development

As leaders at the intersection of Torah and technology, Sefaria will continue exploring cutting-edge developments, keeping you informed of our progress. Below is one project we're currently working on; other projects remain in stealth mode.

PUBLIC BETA: PIRKEI AVOT LEARNING GUIDE

Sefaria’s experimental Pirkei Avot Learning Guide aims to make this popular text more accessible to learners of all levels. Sefaria’s engineering and learning teams leveraged AI to create this interactive learning tool.

The Guide summarizes the various commentaries on Pirkei Avot from the Sefaria library, presenting them through an interactive interface. When learners click on any teaching in Pirkei Avot, they are presented with a selection of questions. The interface allows users to explore these questions, providing summarized answers drawn from the commentaries along with links to the commentary in the library. To build this feature, each mishnah (section of the text) and its associated commentaries were input into Google’s Gemini AI model. A prompt was used to identify questions that were a matter of discussion or debate among the commentaries, and to generate summaries of each answer.

Recognizing that even AI-generated content benefits from editorial review, we took all of the output from Gemini and fed it to Claude Opus 3, an AI model developed by Anthropic. We asked Claude to review each commentary alongside the questions and summaries generated by Gemini and judge them on two scales — relevance of the answer to the question, and accuracy of the summary of the commentary, giving each of them a score of one to 10. While all of the content received some level of human review, our learning team was able to use these scores to identify content that needed an especially close look. Anything with a score of five or below received another review to ensure both accuracy and accessibility for learners.

Visit the Learning Guide.