What is SAI?

SAI aims to unlock superhuman research by developing AI co-scientists in a human-centered way. The first service we provide is an AI review service. You upload a document, or a package that includes code and data, and SAI returns a detailed, referee-style report and detailed inline comments. It is built to do the kind of careful, substance-level review more thorough than human reviewers typically do.

What research is SAI based on?

SAI grew out of research at the University of Chicago. We strive to deliever the best possible co-scientist experience. Learn more about our research and publications on the Research page.

How does a review work?

You upload your document or package and pick a review mode. SAI runs thorough checks throughout the document with customized AI agents. Their findings become one referee-style report and detailed inline comments.

What are the review modes?

Three, increasing in depth:

• Paper-only: reviews the manuscript and its narrative, logic, and claims on their own.
• Paper + code + data: reviews the paper together with the code and data behind it.
• Full run: our most in-depth review, which replicates the paper end to end. Because this re-runs the actual work, its cost depends on the paper and is estimated before the real run.

How is this different from just asking ChatGPT or Claude?

A chatbot gives you one pass over the paper and usually generic feedback. SAI checks it with customized agents validated by our research and, in the full-run modes, re-runs your code and data to confirm the numbers and claims hold. Our research shows that SAI review is clearly better than simply asking Claude or ChatGPT to review your document.

What do I get back?

Two linked views of the same review. The report opens with an overall summary, then strengths and weaknesses, followed by recommended changes. The inline view shows detailed comments beside the passages they point to. You can also share the review with others, and download the report as needed.

How accurate is it, and what are its limits?

SAI is a strong assistant, but it can make occasionally make mistakes. Treat its suggestions as input and make the final call yourself. We will keep improving SAI, and we welcome your feedback on any errors you find.

Will my paper or data be used to train AI models?

No. The third-party AI providers we use run under zero-data-retention terms, which means they do not keep your content or use it to train their models. Full details are in our Privacy Notice.

What happens to my uploads and reports?

Your uploads and reports are stored in your account so you can return to them, and they stay there until you delete them.

Which AI models does SAI use?

SAI uses leading third-party models, currently from providers like Anthropic and OpenAI, all under zero-data-retention terms.

What file formats and sizes can I upload?

PDF, TeX, Markdown, DOCX, and .zip packages (for a paper together with its code and data). Very large submissions may be subject to size or length limits. We will provide pro-rated estimates for very long documents.

How do I get started, and what does it cost?

Create an account and you get a free review to try it. After that, reviews use credits, which you can buy in packs, and the deeper modes use more credits than a paper-only review. Exact prices may change during the beta; current pricing is always on the pricing page.

Is SAI a finished product? What is coming next?

SAI is an early beta and is improving quickly. The review service is live now; Literature, Experiments, and Ideation features are on the way.

Who is behind SAI, and how do I get in touch?

SAI is built by SAI Labs, Inc. For any question, including privacy or data requests, email contact@sai.science.