Checked quotes

Why “it gives sources” is not enough​

Most tools that cite something cite a document. You get a filename, sometimes a page, and the implicit promise that the claim is in there somewhere. Checking it means opening the file and reading until you find the part the model meant — which costs more than reading the source yourself, so nobody does it, so the citation stops working as a check and becomes decoration.

A checked quote is narrower and more useful: not “this document supports this answer” but “this claim rests on these exact words, here.” And because the words are matched against the document by the software rather than vouched for by the model, a quote the model smoothed or invented cannot pass as a real one.

What is checked, and what is not​

The quote must be in the source. The quoted text is matched against the document it cites. Models copy imperfectly — a word swapped, punctuation tidied, a clause dropped — so a near miss is replaced with the source’s own wording, which keeps the citation exact. A quote that cannot be found at all is not quietly shown: in a chat answer it is flagged at the end, and in an AI annotation the note is sent back and redone.

Where it lands is part of the citation. Clicking a quote opens the document there with the sentence highlighted: the page in a PDF, the chapter in an ebook, the page of a presentation, the heading in a Word file.

Whether the quote supports the claim is not machine-checked. A real sentence can still be attached to a conclusion it does not support, and matching text cannot detect that. This is the part that stays with you — but it has stopped being expensive. Reading one highlighted sentence is a check you will actually do; re-reading a chapter is not.

Disagreement is shown, not averaged. When two of your sources say different things, the answer is written to set them side by side, each with its own quote, rather than smoothing them into one bland sentence. That is usually the interesting part.

An answer about an ebook. The citation opens on the quoted words in context, and the same passage is highlighted in the book on the left.

What it changes for you​

You stop having to decide whether to trust an answer as a whole. You check the two or three claims the rest of your work will rest on, a click each.

For long material this compounds. A summary of a book you have not read is unfalsifiable — you cannot tell a good one from a plausible one. A summary where every line points back at its words is something you can audit in ten minutes.

Where you meet it​

• Reading PDFs and ebooks — answers about the document you have open
• AI annotations — every note points at the passage it is about
• Questions across a folder — each claim names the document it came from, as well as the words
• The canvas — cards built from your documents cite the same way, one click from the passage

Earlier versions of this page called the idea a claim graph. The name changed because it described more than the product checks; the behaviour above is what it does.

From the blog: Why AI invents citations, and the check that catches most · What “verbatim” means when the source is a PDF