AI notebooks need source boundaries

What happened

On August 27, 2026, Google introduced Expert Intelligence, a cross-Google initiative that starts inside Gemini Notebook. Google says the feature lets people bring insights from leading authors, publications, and domain experts into Google AI products, beginning with select ebooks purchased through Google Play Books.

Google's own announcement says eligible purchased books can be added directly to Gemini Notebook. Once a book is in the notebook, a reader can ask questions about it and generate formats such as infographics, Audio Overviews, quizzes, and other outputs. Google says responses are grounded directly in the book.

The launch starts with more than 100,000 books from publishers including Bloomsbury, De Gruyter Brill, Johns Hopkins University Press, Macmillan Publishers, O'Reilly Media, and Penguin Random House. Google also says it worked with more than 15 authors, including Steven Pinker, Michael Pollan, and Jennifer Wallace, on Featured Notebooks that add supplemental sources and artifacts around their books.

The access controls are part of the announcement, not a side detail. Google says people must own a book through Google Play Books to interact with that book in Gemini Notebook. If a notebook is shared, collaborators who do not own the book will be prompted to buy their own copy before they can use that book as a source.

The Verge reported the same day that Google demonstrated the feature with examples involving Michael Pollan's Food Rules and Kim Scott's Radical Candor, and that Google plans to bring Expert Intelligence to AI Mode in Search and the Gemini app. Publishers Weekly also reported on August 27 that publishers and the Authors Guild welcomed the copyright and attribution guardrails while noting the larger legal context around books and AI.

Those are the reported facts. REC's read is about workflow design. Google is making source boundaries visible: which book is in the notebook, who has access to it, what outputs can be generated from it, and what happens when a collaborator does not have rights to the underlying material.

Why this matters

AI-assisted publishing keeps moving from blank-page generation toward source-based adaptation. That shift matters because most expert communication should not begin with a prompt. It should begin with a real source: a person answering a researched question, a book, a report, a transcript, a dataset, a demo, or a document the team has permission to use.

Expert Intelligence is timely because it gives a mainstream consumer shape to that idea. The user is not simply asking a model what it knows about a famous book. The product flow asks whether the user owns an eligible title, adds that title as a notebook source, and limits what shared collaborators can do when they do not own it.

That is a more useful pattern than treating every source as generic background. It acknowledges that the source has an owner, a scope, and a permission boundary. It also acknowledges that the output changes form. A book can become a quiz, an audio summary, an infographic, or a plan, but the source should still be identifiable.

For REC, the lesson applies to expert material. A research-guided video interview creates source material that should stay attached to every downstream asset. If a founder's answer becomes a blog section, clip, social post, sales note, or webinar outline, the team should know which recorded answer supports it.

That discipline helps readers and teams. Readers get clearer claims. Editors get a way to check whether the final asset still matches the source. The expert gets a record of what they actually said. AI can still help with structure, summaries, and adaptation, but the authority comes from the source.

The source boundary is the product

The important product detail is the boundary around the source. Google says a book must be eligible, purchased, and added as a source before Gemini Notebook can use it in this flow. Google Help also lists Play Books as a supported source type and says eligible purchased titles can be imported, while ineligible titles appear grayed out.

That boundary changes the editorial question. Instead of asking whether AI can produce a good-looking output, the team can ask what source the output is allowed to rely on. That is a sharper standard. It moves review from surface quality to source fit.

A source-bound workflow needs three records. First, it needs the source record: the interview, transcript, document, or purchased book that shaped the work. Second, it needs the rights record: whether the team can use that source, share it, quote it, or adapt it. Third, it needs the edit record: how the source became the final asset and what AI changed along the way.

That may sound procedural, but it is practical. Without those records, a team can easily publish a confident claim that came from a loose summary, an unapproved note, a misremembered interview answer, or a source the team was allowed to read but not republish.

The same issue shows up in expert video. A clip may feel clear on its own, but the fuller answer may include a caveat. A blog paragraph may read cleanly, but the recording may show the speaker was describing one customer segment, one market condition, or one personal view. Source boundaries help the editor keep that context.

What REC teams can learn

First, identify the source before drafting the asset. In REC, the strongest source is usually the expert on camera answering a researched question. Supporting sources can include public pages, customer-safe notes, slides, reports, or product documentation.

Second, keep sources separate until the editor decides how they should meet. Google's example lets a notebook combine a book with the user's own information. REC teams can do something similar, but they should label what came from the interview, what came from outside research, and what came from internal context.

Third, use AI to adapt from the approved source set. AI can suggest headings, pull candidate clips, clean transcript punctuation, draft summaries, and create variants for different channels. The review question should be whether the output stays inside the source boundary and preserves the expert's meaning.

Fourth, preserve permission decisions near the content. A founder interview, a customer quote, a paid report, a partner deck, and a purchased book do not carry the same rights. If a source can inform the team's thinking but cannot be quoted or shared, the production record should say that plainly.

Fifth, review the final format as its own object. A quiz, audio overview, infographic, blog section, short clip, or sales email can introduce a new meaning by changing sequence, emphasis, or audience. The final asset needs its own check against the source, not only a quick spellcheck.

Sixth, make retrieval easier for the future. Store the transcript passage, source link, permission note, AI role, and approval note together. When the claim is reused later, the team should not have to reconstruct where it came from.

What teams should avoid

Do not treat a source list as proof. A source list can show where material came from, but an editor still has to confirm that the final claim follows from the source.

Do not mix owned, licensed, public, private, and customer-provided material without labels. Those sources have different boundaries. Treating them as one pile creates legal risk and editorial confusion.

Do not let AI flatten expertise into generic advice. If the source is a real interview answer, preserve the speaker's condition, example, tradeoff, and limit. Those details are often the reason the answer is useful.

Do not assume a generated format is harmless because it is derivative. A summary, quiz, infographic, or audio version can still misstate the source or imply permission the team does not have.

Do not publish from a notebook or transcript without a final human review. Source grounding helps, but it does not decide what a reasonable reader will take away from the finished piece.

The practical takeaway

Google's Expert Intelligence launch is timely because it puts source rights, AI adaptation, and content discovery in the same product story. A reader can bring an eligible purchased book into Gemini Notebook, ask questions, and generate new formats, while Google keeps a boundary around who can use that book as a source.

REC teams should copy the discipline, not the exact feature. Start with an accountable source. Keep it attached to the transcript or document. Label the rights boundary. Name the AI role. Review the final asset against the original material.

AI-assisted publishing becomes more useful when the source boundary is clear. The reader may only see the final article or clip, but the team should be able to trace it back to the material that earned the claim.