Google's Gemini Notebook, until July 2026 called NotebookLM, is an AI tool that answers only from documents you give it. Its co-creator, the author Steven Johnson, built it around a particular theory of how people should think with AI; a theory not everyone accepts. This walks through the tool itself and the hoarding habit it is meant to break. It looks at the difference between asking AI to think for you, and using it to get more into your own head. It ends with what one study found when students raced GPT-4.
Module 01
A tour of Gemini Notebook
In this module: what a source-grounded notebook is, and what its three panels do
It began at Google Labs in 2023 as an experiment called Project Tailwind, launched that July as NotebookLM, and was renamed Gemini Notebook in July 2026. The idea has stayed the same throughout: you bring the documents, and the AI works only from those.
One notebook, three panels
This is a real notebook, built for this workshop from 60 sources about the tool itself. Click each panel, and the small print at the bottom, to see what it does.
Steven Johnson’s Notebook LM60 sources · Aug 19, 2026
Click a panel to unpack it.
Source grounding: it answers from what you gave it
When you ask a question, the tool does not go looking across the open internet. Your documents are loaded into the model’s working memory, and the answer is assembled from them, with numbered citations pointing back to the passage each claim came from.
Google’s own launch wording from 2023 is careful. Source grounding “does seem to reduce the risk of model ‘hallucinations’”; even so, “it’s always important to fact-check the AI’s responses against your original source material”. Reduce, not remove. The citations are what make the checking fast.
How much it holds
50 sourcesPer notebook, on the free tierPaid tiers raise the ceiling: 100 on Google AI Plus, 300 on Pro, more again on Ultra. The notebook in this deck holds 60.
500,000 wordsPer source, on every tierRoughly the length of War and Peace. Each uploaded file can also be up to 200MB. A single notebook can hold a unit’s entire reading list with room to spare.
Most formatsPDFs, web pages, Docs, audio, video linksThe notebook above mixes journal PDFs, web articles and a recorded interview. The tool reads across all of them at once.
The scale matters. This is not a chatbot with a pasted paragraph; it is your whole reading list, held in working memory at once.
The Studio: reshaping sources into other forms
The right-hand panel turns the same sources into different formats. Click each one.
Click a card. Every output stays grounded in the same uploaded sources.
Module 02
Why we hoard instead of think
In this module: the habit the tool was built to break
Before the pedagogy makes sense, you need the problem it answers. The problem is not that we lack information. It is that we collect it, file it, and never think with it.
Your brain is an association engine, not a filing cabinet
The human brain is very good at recognising patterns, making connections and reading a room, and remarkably bad at storing exact detail on demand. Yet most of us use it as a warehouse: action items, half-read articles, that statistic from a newsletter three weeks ago. “Your mind is for having ideas, not holding them.”
The line is productivity writer David Allen’s, quoted by Tiago Forte in Building a Second Brain (Atria Books, 2022). Forte’s book popularised the term “second brain”: an external, organised store of what you have read and thought.
The cost of the digital pile
Have a guess before you look. How much of a knowledge worker’s week goes to searching and gathering information, rather than using it?
About a day a week. McKinsey’s estimate was 19 per cent of the working week spent searching and gathering information; roughly 1.6 hours every working day. A 2023 Gartner survey adds that 47 per cent of digital workers struggle to find the information they need to do their jobs.McKinsey Global Institute, “The social economy: Unlocking value and productivity through social technologies”, July 2012. An old figure, from a pre-AI workplace, but still the most cited estimate. Gartner press release, 10 May 2023. Both are estimates of scale, not precise measurements.
Saving feels like progress. It is not.
Forte calls the pattern being an information hoarder: a hundred bookmarked articles, hours of downloaded podcasts, forty open browser tabs. The click of “save for later” stands in for actual work. As he puts it, “The paradox of hoarding is that no matter how much we collect and accumulate, it’s never enough.” The pile does not make you more knowledgeable. It makes you more anxious, because you know you have not processed it.
A second brain is meant to end the hoarding by making the pile answer questions. Whether that helps you think, or thinks instead of you, is the argument of the next module.
Module 03
Offloading and uploading
In this module: the intended pedagogy, and the case against it
Steven Johnson is an author of fourteen books who joined Google Labs and helped create NotebookLM. In June 2026 he published an essay giving the tool’s intended practice a name: cognitive uploading. The distinction it draws has become the sharpest available frame for talking to students and staff about AI use.
Two directions of travel
Cognitive offloading
Thinking moves out of your head. The machine does the analysis and you take the output. Johnson’s own example: a student who bypasses actually researching and writing a paper by handing the task over to an AI. The work gets done; the student’s understanding of the topic is unchanged.
The direction most public worry about AI points at, and the one most assessment policies are written against.
Cognitive uploading
More moves into your head. The machine is used to surface, organise and question, so that you encounter more ideas, faster, and still do the understanding yourself. In Johnson’s words: “we’ve spent most of our time worrying about what happens when AI does our thinking for us. But we haven’t focused enough on all the ways AI gives us new things to think about.”
The practice Gemini Notebook was designed around.
Steven Johnson, “Cognitive Uploading”, Adjacent Possible, 2 June 2026. The offloading and uploading labels are his.
What uploading looks like in practice
Researching the Watergate break-in, Johnson does not ask the AI to write the history. He asks it to act as a research librarian: “Find me the most relevant primary texts and secondary commentary that I need to understand the Watergate break-in”. The AI curates; he still reads. Then, separately, he turns the tool on his own draft thinking with what he calls negative search space questions: what am I missing? What do I need to know that I do not know about yet?
Both practices are from the same essay. The first speeds up finding the right things to read. The second uses the machine’s breadth to expose the blind spots in an argument you have already formed. Neither replaces the reading itself.
Offloading or uploading? You judge
Four things a student or colleague might do this week. Call each one.
A student pastes the essay question into a chatbot, pastes the response into a document, and edits the wording so it sounds like them.
Offloading. The analysis happened in the machine. Editing the wording changes the surface, not who did the thinking; the student’s model of the topic is unchanged.
A lecturer loads twelve unit readings into a notebook and asks: “Which concepts appear in more than one reading, and where do the authors disagree”? Then re-reads the two disputed papers.
Uploading. The machine surfaced a map of the terrain; the lecturer went back to the primary texts and did the judging. More entered their head than they had time to find alone.
A student generates an Audio Overview of the week’s readings and listens on the bus, then still does the readings, with a sense of the shape before starting.
Uploading, used well. An advance organiser before the real reading. It would tip into offloading the week the podcast starts replacing the readings instead of introducing them.
A course coordinator has a notebook summarise forty pages of feedback into themes and presents the themes at a meeting. Asked what students actually wrote, they cannot say.
Offloading, and it showed. The summary was fine; the understanding behind it was never built. The moment a question probes past the output, there is nothing there. Johnson’s frame predicts exactly this failure.
The objection: struggle is where learning happens
The case against is serious, and it is not new to AI. Cognitive science has long studied how we push memory and thinking onto external tools; the field calls it cognitive offloading. The consistent finding: what the tool holds, the head tends not to. A 2025 survey of 666 people found heavy AI use strongly associated with more offloading and weaker critical-thinking scores, with the youngest users most affected. If the AI curates the readings, summarises the connections and finds your blind spots, how much of the productive struggle that builds understanding is left?
Risko and Gilbert, “Cognitive Offloading”, Trends in Cognitive Sciences, 2016; Gerlich, “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking”, Societies, January 2025 (corrected September 2025). The Gerlich result is correlational: it cannot say whether AI use weakens thinking or weaker thinkers lean on AI. Johnson’s reply, in effect: the struggle belongs in the reading and the arguing, not in the searching. The contested question is whether students can tell the difference.
The bluff that fails at question time
Feed the term’s materials into a notebook, generate a confident summary, present it as your own reading. It works until the first follow-up question; the first “what did the author mean by that”, the first request to connect it to something new. Outputs can be borrowed. Understanding cannot. The gap between the two is invisible until someone probes it. That is worth saying plainly to students; it is also the honest case for doing the reading.
So the pedagogy stands or falls on what the machine is actually doing when it connects your sources. That is the last module.
Module 04
The fox, the hedgehog and the classroom
In this module: how AI connects ideas, where it goes wrong, and what to do with that
When a notebook says two of your sources are related, it is reasoning by analogy: mapping the pattern of one thing onto another. Analogy is also one of the oldest tools in teaching. So it matters that humans and AI are good at opposite halves of it.
Try a famous one yourself
A patient has an inoperable stomach tumour. A ray at high intensity destroys the tumour, but destroys the healthy tissue it passes through on the way in. At low intensity the ray is harmless to tissue, but harmless to the tumour too. How do you destroy the tumour without harming the tissue?
Many weak rays, from many angles, all meeting at the tumour. Each beam is harmless on its own; only at the point where they converge is the dose strong enough to destroy.Duncker’s radiation problem, from 1945, and the classic finding about it: most people cannot solve it cold. In 1980, Gick and Holyoak gave people a story with the same structure first: an army splits into small groups and converges on a fortress. Most still failed to use it until told the story was relevant. Seeing a deep structural match across two surfaces that look nothing alike is hard, even for us.
When students raced GPT-4 at exactly this
The AI: sees everything, believes too much
GPT-4 showed high recall, low precision: in the authors’ words it “did not miss valid analogies” but “often surfaced spurious, even if internally coherent matches”. It maps surface likeness as eagerly as deep structure, and argues fluently for both.
Like a metal detector: it beeps at everything under the sand, gold and bottle caps alike.
The humans: miss a lot, misuse little
The management students showed low recall, high precision: they “frequently overlooked valid analogies” but “rarely misapplied them”. When they did claim a match, they had checked the causal logic underneath it, and were usually right.
The archaeologist: slow, selective, and able to tell the gold coin from the rusty can.
Sen, Workiewicz and Puranam, “Can LLMs Aid Analogical Reasoning for Strategic Decisions? A Comparative Study”, Strategy Science, 2026. Their proposed division of labour: use the AI as an “expansive retrieval engine”, and keep the judging of causal fit with the human. Some retellings dress this up as the fox and the hedgehog: the broad generator and the careful diagnostician. The picture helps, though the label is a gloss, not the authors’ own.
What this settles about the pedagogy
The study puts a floor under both sides of the module 3 argument. The uploading case is real: the machine will surface connections across sixty sources that no one has time to find alone. The objection is real too: a fluent, internally coherent connection is not evidence of a true one. The human checking step is not politeness; it is where the accuracy comes from. Cut the reading and the judging out of the workflow and you have not streamlined it; you have removed its error correction.
Which turns the tool from a threat into a rather good teaching design: let it retrieve, and assess the judging.
Five uses that survive the evidence
CurateBuild a unit notebook from your own readingsLoad the actual reading list, not the open web. Students question a corpus you chose, and every answer carries citations back into material you set.
PreviewAudio Overviews as advance organisersA generated discussion of dense readings, listened to before the reading, lowers the barrier to starting. Steer it: tell it who the audience is and what to focus on, or it will pitch a general-interest episode.
ProbeTeach negative search space questions“What is missing from my argument?” is a better student prompt than “summarise this”. It uses the machine’s breadth and leaves the thinking with the student.
VerifyAssess the checking, not just the findingHave students ask a notebook for connections across sources, then defend or reject each one against the passages it cites. That is the precision skill the study says humans bring.
DiscloseMind what you uploadA notebook full of student work or unpublished research is a data decision, not just a teaching one. Our workshop on caring for your data covers where chats and uploads go.
Why this matters
Gemini Notebook is the most classroom-shaped AI tool yet built. It is grounded in sources you choose and cites what it says; it is wrong less often, but still sometimes wrong, as its own footer admits. The pedagogy behind it asks for something specific and teachable: use the machine to get more into your head, never to avoid the work of understanding. Treat every connection it offers as a lead to check, not a fact to repeat. That is not a workaround for AI in education. It is a defensible model of it, and the argument about whether students can hold that line is worth having in the open, with them.
Sources used. Product: Google, “NotebookLM is now Gemini Notebook”, blog.google, 16 July 2026; Google, “Introducing NotebookLM”, blog.google, 12 July 2023 (the source-grounding and hallucination wording quoted in module 1); Google support pages on plan limits and Studio output types, accessed 22 August 2026; Google, “NotebookLM now lets you listen to a conversation about your sources”, blog.google, 11 September 2024; Google DeepMind, “Pushing the frontiers of audio generation”, 30 October 2024 (the speech technology behind Audio Overviews); the notebook shown in module 1 is a real NT World Ink notebook, screenshots taken 19 August 2026. Steven Johnson: “Cognitive Uploading”, Adjacent Possible, 2 June 2026; Google Workspace blog interview with Johnson, 29 May 2024 (his role at Google Labs). Second brain and hoarding: Tiago Forte, Building a Second Brain, Atria Books, 2022; the “mind is for having ideas” line is David Allen’s, quoted by Forte. Search-time figures: McKinsey Global Institute, “The social economy”, July 2012; Gartner press release, 10 May 2023. Analogical reasoning: Sen, Workiewicz and Puranam, “Can LLMs Aid Analogical Reasoning for Strategic Decisions? A Comparative Study”, Strategy Science 11(1), 2026 (quotes are from the published abstract); Duncker, “On problem-solving”, Psychological Monographs, 1945; Gick and Holyoak, “Analogical problem solving”, Cognitive Psychology, 1980. Offloading research: Risko and Gilbert, “Cognitive Offloading”, Trends in Cognitive Sciences 20(9), 2016; Gerlich, “AI Tools in Society”, Societies 15(1):6, 2025, corrected version. This deck draws on a recorded dialogue about Johnson’s notebook; claims from it that could not be verified against a primary source were left out, including a “Ferrari” analogy widely misattributed to Forte and a claim that Johnson refuses to publish AI-written books under his name.