I don't think the biggest benefit of AI tools existing is the fact that you can ask a question and get a response within seconds. Or that you can generate an image of your younger self with a single sentence. Or that you can create a slide deck that would have previously taken you weeks to build within an afternoon. Those things are useful, sure, and a few of them still feel like magic. But they're all surface-level wins, and I think they're simply faster versions of things we already knew how to do.

I think the biggest benefit of AI has been the effect it's had on research. Now, I'm not talking about the generic "summarize this article for me" or "give me five bullet points on World War Two." Instead, I'm referring to the research that used to live across twelve browser tabs, a bunch of notebooks, and a Notion dashboard you spent more effort maintaining than using. The kind of research where holding onto all the material you had and making sense of it was a pain.

NotebookLM is a tool I've been using for this since it launched, and while it has a few problems here and there, I've largely been a fan. Claude Projects is one I've found myself leaning more and more on over the past several months. Eventually, I ended up ditching NotebookLM for Claude Projects, and now, after spending real time inside both, the trade-offs have surprised me more than I expected.

NotebookLM and Claude Projects start from the same place

And it's the same place a lot of AI tools should start

On the surface, both NotebookLM and Claude Projects solve the same problem I just described above: making a pile of source material feel like something you can actually think with. A big problem with using AI for research is that it doesn't necessarily have a direction to pull from. You ask a regular chatbot a question, and it answers from everything and nothing at once. Both NotebookLM and Claude Projects fix this by giving the AI a specific, bounded set of material to work from — your material, the stuff you've already vetted and decided is worth paying attention to.

Once you've uploaded this material, each of the tools has a solid direction of what you're working on, and you can sort of connect it to working with an assistant who has actually done the reading. Instead of asking a question and bracing yourself for whatever the model decides to confidently invent, you're asking a question of something that has the same documents open in front of it that you do.

But they treat your sources very differently

And the difference shows up fast

While both NotebookLM and Claude Projects seem to be targeted at solving the same problem, they work in very different ways. NotebookLM is an entire tool built around this problem, and it lets you create what it calls notebooks. These are individual workspaces, where you drop in your sources and chat with them in isolation. Every notebook is really its own little universe, meaning every notebook you create is independent from the others. You can pack a notebook with a bunch of sources, and then ask questions about them or turn them into different Studio outputs.

Claude is a conversational AI chatbot first and foremost, and Projects is one feature inside it. Projects are self-contained workspaces with their own chat histories and knowledge bases, where you can upload documents, set custom instructions, and have focused chats with Claude that draw on whatever you've put into the project's knowledge base. The structure is similar to NotebookLM, but the boundaries are incredibly different.

With NotebookLM, every notebook's data is sealed off from everything else. The model has no broader knowledge to fall back on, no way to pull in something from outside the notebook, no awareness of the wider web. There are features within the tool that now let you use the wider web to fetch sources, but you're still the one in charge of deciding what gets pulled in. The model never reaches outside the notebook to answer a question. Claude Projects, on the other hand, works very differently. You do have the option to upload your own sources to a project's knowledge base (PDFs, docs, text files, code, whatever you've got), and Claude will lean on that material when you're chatting inside the project.

However, the material you upload is simply the starting point. Claude still has access to its full general reasoning, its training data, and (if you've enabled it) live web search. So when you ask a question, Claude doesn't just look at what's in your project. Instead, it considers your sources alongside everything else it knows, and weaves them together into an answer.

The trade-off is really about what kind of help you need

It depends on the kind of work you're doing

I began using NotebookLM and leaning on it heavily because of the entire source-grounded aspect. With Claude Projects, I assumed I'd get a similar experience since my documents would still be given to the model as the foundation of every chat. But like i just mentioned, Claude uses your sources as one input among many, not as the only thing they're allowed to work with. That's a feature and a flaw, and which one it is depends entirely on what you're trying to do. Say I'm working on a research project, and I've got a folder on my laptop with fifteen papers I've gathered on a specific topic.

I want to interact with those documents like find connections across all of them, surface arguments I missed, turn them into a study guide or an audio overview I can listen to on a walk, ask questions and trust that the answers are coming from the material I actually vetted. In that scenario, Claude Projects isn't really going to help me the way I want it to. Claude will happily reach beyond my fifteen papers, pull in something from the wider web, lean on its own general knowledge, and weave all of that into the answer. That's useful in plenty of other contexts. But it's not what I need here. What I need here is a tool that respects the walls of my research, and NotebookLM is built exactly for that.

However, say I'm studying for a midterm and I have a course outline plus a few lecture slides. If my lecture slides don't cover all the content within my course outline, NotebookLM isn't really going to be able to help me all that much. I can generate quizzes, flashcards, slide decks, audio overviews, etc., or ask as many questions as I have about the existing material I've uploaded to a NotebookLM notebook. However, the moment I need information that isn't in my uploads, NotebookLM stops being useful for me. It'll politely tell me it can't find anything in the sources I've provided, and I'm back to opening a new tab and Googling whatever the lecture slides didn't cover.

Claude Projects handles this scenario completely differently. If I upload the course outline and the lecture slides to a project, Claude treats the outline as the actual structure of what I'm trying to learn. When I ask it about a topic on the outline that my slides don't cover, it goes and finds the material for me, using web search and its own general knowledge, while still keeping the course outline as the anchor for what's relevant.

That last part is the thing that makes it different from just using a regular Claude thread. A normal chat doesn't really know what I'm studying. It would search across everything, surface whatever felt most popular or recent, and leave me to figure out whether any of it actually matched my course. Claude inside a project knows exactly what I'm working on, because the outline is sitting right there as context. The web searches it runs are shaped by that outline, and so are the explanations it gives me.

I now prefer using both interchangeably

After spending real time inside both NotebookLM and Claude Projects, I've realized there's no point in sticking to just one. I've come to using both for different parts of the same project. NotebookLM is what I reach for when I've already gathered everything I need and I want to actually live inside that material. On the other hand, Claude Projects is what I reach for when my research isn't done yet and when I want to actually do something with what I've learned beyond just understanding it.