NotebookLM Was Supposed to Help Us Think

The distinction, between consumption and cognition, is worth paying attention to right now. Especially when it comes to NotebookLM.

Credit where it is due. When Google Labs launched NotebookLM in 2023, the underlying instinct was sound. The tool was built to sit between you and your own documents, not the open internet. It would not hallucinate wildly the way a general chatbot might because it was grounded in whatever you uploaded.

Steven Johnson, the developer who has worked on the project from the start, described it as “a tool for understanding things.” That is a genuinely interesting design philosophy. Most AI products at the time were optimising for generation. NotebookLM was, at least in theory, optimising for comprehension.

The source-grounding constraint was its defining feature and its most pedagogically promising one. Think of it like a private tutor who has read every book on your shelf but refuses to make things up. The tutor can only work with what you have given it. That limitation forces a different kind of interaction. You are not asking the universe for answers. You are interrogating your own material.

In an educational context, this is valuable. It means a student cannot simply prompt their way to an essay. They have to supply the sources first, and in doing so, they have to at least begin the process of curation.

The feature creep nobody really noticed

The 2026 version of NotebookLM is a substantially different product. It now generates slide decks with image generation, infographics, comparison tables, mind maps, video overviews, and fully structured reports. It can export to PowerPoint. The Studio panel offers suggested formats based on your sources, so you do not even need to decide what kind of output you want. The new “Join” feature lets you interrupt the Audio Overview hosts mid-sentence through your microphone, turning a passive podcast into something closer to a live conversation.

Each of these features is, in isolation, clever. Some of them are genuinely useful. But taken together, the trajectory is unmistakable. The tool that was designed to help you understand your sources has become a tool that produces finished outputs from them. The cognitive middle, where understanding actually happens, has been gradually compressed.

This is not a criticism unique to Google. Every AI product is under the same competitive pressure to shorten the distance between input and output. But it is worth being honest about what that compression costs, especially when the product is marketed to students and educators.

The effort-reward cycle and why it is worth protecting

There is a useful concept in cognitive neuroscience called the effort-reward cycle. When you struggle with something difficult, your brain’s reward system activates. The dopamine hit that follows successful effort reinforces motivation and deepens engagement. It is the reason a hard-won insight feels different from one that was handed to you. The struggle is not a flaw in the learning process. It is the engine.

NotebookLM’s original design, the one that asked you to upload your own sources and then interrogate them, preserved some of that productive friction. You still had to choose what to upload. You still had to formulate questions. You still had to decide what the tool’s answers meant for your own work.

The newer features compress that friction. An auto-generated mind map tells you what the key themes are before you have formed your own view. A one-click report structures arguments you have not yet made. An Audio Overview summarises readings you have not yet wrestled with. Each of these outputs looks, from the outside, like evidence of understanding. But the understanding belongs to the model, not the student.

A more useful way to think about this

The most helpful framing is probably not “NotebookLM is bad” but rather “NotebookLM is misaligned with how learning actually works, unless you deliberately resist its default path.”

The default path is production. Upload, generate, export. The Studio panel is front and centre. The features that encourage slower, more deliberate engagement, like the chat’s new Thinking Mode, which shows the reasoning chain behind an answer, are there but secondary. They require the user to consciously choose depth over speed.

This is a design problem, not a moral one. AI tools should be designed to enhance motivators for critical thinking, such as quality standards and skill-building, and to mitigate inhibitors, such as time pressure and low awareness. AI should not be treated as a tool for delivering answers but as a “thought partner” that can also act as a “provocateur.”

That language is interesting, because it is almost exactly what NotebookLM was supposed to be. The original vision, the one Steven Johnson described, was closer to a provocateur than a production line. Somewhere along the way, the product outgrew its philosophy.

What educators (and the rest of us) can do

The good news is that the source-grounding constraint still exists. NotebookLM still will not wander off into the general internet. It still anchors its outputs to the material you give it. That constraint makes it more honest than most AI tools, and it creates a genuine opportunity for educators willing to work with the grain rather than against it.

The trick is to use the tool upstream rather than downstream. Use the mind map as a starting point for discussion, not a substitute for it. Have students generate an Audio Overview and then critique it, identifying what the hosts got wrong or oversimplified. Use the comparison table as a first draft that must be revised, not a finished product. In other words, treat NotebookLM’s outputs the way you would treat a student’s first attempt. As raw material for the real work of thinking.

The students who benefit most from AI tools will not be the ones who use them to skip the hard part. They will be the ones who use them to find the hard part faster.

NotebookLM is not a bad tool. It is, in many respects, an impressively engineered one. The source grounding is genuine. The interface is clean. For someone who has already done the reading and wants to check their understanding or generate a quick reference materials, it is genuinely useful.

The problem is that this is not how most people use it, and Google knows that. The trajectory of every feature update moves in the same direction. Towards output. Away from process. Towards product. Away from thought.

And that’s something we have to be cautious about.

Note: I have created free ebooks on how to maximise the use of NotebookLM without losing the “thinking process”. Feel free to download from https://chuahkm.com/ebooks