How Synthesis works
Synthesis picks new research papers, explains each one in plain language, and links you back to the original. This page covers where the papers come from, how the explanations are written, and what the app keeps about you.
Where the papers come from
Every paper is real and openly published. Synthesis only uses a paper when the licence on its abstract allows reuse, and credits it wherever its text appears.
- arXiv
- Preprints in physics, mathematics, computer science and more.
- OpenAlex
- An open catalogue of scholarly work, used to find open-access journal articles.
- Europe PMC
- Life-science and biomedical research.
- OSF Preprints
- Preprints across the social sciences, psychology and education.
Many are preprints, which means they haven't necessarily been peer reviewed yet. The paper page says so when that is the case.
How the explanations are written
An AI language model reads each paper's abstract and writes everything you see about it: the headline, a summary at two levels, an analogy, why it matters, key terms, common questions, and what the paper doesn't show. It works under four rules.
- Explanations come only from the paper's own abstract. There is nothing else for the model to drift toward.
- Numbers are copied exactly, never calculated or estimated.
- If the paper doesn't answer your question, it says so instead of guessing.
- Every explanation is labelled AI-generated, and preprint status is shown before you trust anything.
The Bangla version is written in a separate pass, which is given the English explanation alongside the abstract so that both languages say the same thing about the paper.
Reels are narrated by a synthetic voice and illustrated with AI-generated pictures. The artwork on each card is drawn by code, seeded from the paper's id.
The Bangla narration is spoken by OmniVoice, from k2-fsa, in a voice modelled on one recording from Google's Bangladeshi Bengali speech corpus (OpenSLR 37, CC BY-SA 4.0). Numbers are spelled out before they are read, so the voice says them the way a reader would.
Asking about a paper
A question you ask is sent, with the paper's abstract, to an AI model. It answers from the abstract alone and tells you when the abstract doesn't cover what you asked. Questions aren't saved, and each person has a daily limit.
Keyboard
Card deck
- Next paper
- Previous paper
- Explore the paper
- SpaceEnter
- Close a sheet or menu
- Esc
Desktop reader
- Next paper
- Previous paper
- Save, or remove from saved
- S
What the app keeps
Your fields, saved papers, reading level and language are stored in this browser. Sign in with Google and your saved papers are kept with your account too, so they follow you to another device.
The app records anonymous usage, such as a paper being opened, saved or skipped, under a random id. It is used to see how the app is used, and it isn't tied to your name or your account.
The privacy policy has the full detail, including how to delete your account.
The reading study
Part of this project tests whether these explanations are easier to follow than the abstracts they come from. Readers see both for the same paper, without being told which is which, and pick the one that helped them more.
See the results so far