A reading room for papers, with the AI on tap

Reading a paper with a chatbot in the next tab means copying passages back and forth and hoping the model is answering about the paragraph you are actually looking at. Reader puts the two in one room: the document on the left, the AI on the right, and a selection that both of them agree on.

Open what you already have

Pull a document out of your asset library or upload one. The bytes stay in the shared library either way — Reader points at them rather than making a second copy, so a paper you annotate here is the same file the graph, the Files Editor and the rest of the workspace already know about.

Text and pages, together

You get selectable extracted text and the rendered page images. The text is what makes selection, search and read-aloud possible; the pages are what makes a figure still look like a figure. Reading position travels with the document, so closing the tab is not the same as losing your place.

Questions that are grounded in what you highlighted

Select a passage and ask about it, and the answer is anchored to that passage rather than to a vague memory of the whole PDF. Summarize condenses the document; Ask answers a question about the selection; Slides turns a section into a deck you can present. Each of those three settles its credits up front, the same way the Trip Planner does, so a long document cannot quietly run a tab.

Read-aloud costs nothing

Listening rides the browser's own speech synthesis over the extracted text. No model, no queue, no credits — it is your device reading your document to you. That is the whole reason the extracted text is a first-class citizen here and not an implementation detail.

Notes that stay with the document

Notes are attached to the document, not to a session, so they are still there next month when you open it again to check what you thought the first time.

Why selection is the whole interface

Asking an AI about a document usually means asking about the document — all of it, at once, with the model choosing which parts to attend to. That works for “what is this about” and fails for everything more specific, because the model's idea of what is relevant is not yours.

Highlighting a passage and asking about that removes the guesswork. The from selection chip on an answer is not decoration; it is the claim that this answer came from those words. When you disagree with an answer you know exactly what it was looking at, which is the precondition for arguing with it productively.

Text and pages, and why you need both

Extracted text is what makes selection, search and read-aloud possible. Page images are what keep a figure looking like a figure, a table like a table, and an equation like an equation.

Most readers pick one and compromise the other. Keeping both is why the reading room can offer a text-driven feature like read-aloud without turning your paper into a wall of reflowed prose where the diagrams used to be.

Listening costs nothing

Read-aloud rides your browser's own speech synthesis over the extracted text. No model, no queue, no credits — it is your device reading your document to you.

That is worth stating plainly because text-to-speech is usually metered, and metering changes behaviour: people ration a feature they are charged for. Here you can listen to a whole paper on a walk and it costs the same as not listening to it.

What does cost, and when it is settled

Three actions reach a model: Summarize, Ask and Slides. Each settles its credits up front, the same way the Trip Planner does, so a long document cannot quietly run a tab while you work through it.

Everything else — opening, scrolling, selecting, highlighting, taking notes, listening — is free.

Notes and position travel with the document

Notes attach to the document rather than to a session, and reading position is stored with it. Close the tab and come back next month to the same page with the same annotations.

The bytes stay in the shared asset library either way — Reader points at them rather than making a copy — so a paper you annotate here is the same file the canvas, the Files Editor and every other surface already know about.

Reading is not the same as extracting

It is tempting to treat every document as something to be processed. Some are, and document ingest or grounded drafting are the right tools when the goal is to get something out.

But a paper you are trying to genuinely understand rewards being read, in order, at your own pace, with the figures in view. The AI here is on tap rather than in the way: it answers when you highlight something and is otherwise absent. That restraint is the design.

Frequently asked questions

What is the difference between Reader and ingesting a document on the canvas?

Reader is for reading: page images, selection, notes, position, read-aloud. Document ingest is for thinking about structure: the file becomes a typed graph of roles and edges you branch from and question. Read in Reader, think in the graph — and both point at the same file in the library.

Can I read something I did not upload myself?

Anything in the shared asset library opens here, including files another surface produced.

Does Slides produce something I can present?

It turns a section into a deck. The useful workflow is to select the part that matters rather than asking for slides on a whole paper, for the same reason selection improves answers: a deck about everything is a deck about nothing.

Can I upload a document, or does it have to come from the library?

Either. Upload directly or pull from the shared asset library — and an upload joins the library, so it is available to the other surfaces afterwards without a second upload.

Does read-aloud work offline or on any language?

It uses whatever voices your browser and operating system provide, so availability and language coverage are your device's rather than ours. The upside is that it costs nothing and works without a round trip; the downside is that quality varies by platform.

Can I search inside a document?

Yes — that is one of the things the extracted text layer is for, alongside selection and read-aloud. It is also why a scanned PDF with no text layer needs OCR first: without extracted text there is nothing to search, select or read aloud, however good the page images look.

Is Reader available on every plan?

It is a paid surface gated by plan entitlement, and it ships in English for now — matching how Watch and the Trip Planner launched, with catalogues and a tutorial to follow.

Open Reader with the paper you have been meaning to finish.