If you've been following XDA's AI coverage, something you surely know is that we love NotebookLM. We've been reporting on the tool since its early Google Labs days, and we're always looking for new (and often slightly unhinged) ways to push the tool to new limits. A lot of the NotebookLM articles you'll find on the site have been written by me, and I've usually only had good things to say about the tool.

While you'll definitely find the occasional disappointment, those have always been the exception. Besides, it's only natural to have some gripes about a tool you fundamentally believe in every now and then. However, I feel like the NotebookLM I've championed from day one is disappearing more and more with every update.

It feels like NotebookLM is slipping away from itself

It's not you, NotebookLM, it's... actually, it is you

When Google first began rolling out Project Tailwind with its new name, i.e. NotebookLM, the company explained that the whole thing was built around a single, very human problem: it's hard to go from information to insight. We're all drowning in sources, and the actual work is synthesizing facts and ideas across them by making relevant connections. NotebookLM was Google's answer to this problem. The way it achieved this was grounding. Instead of being yet another chatbot that knew a little about everything, NotebookLM let you "ground" the language model in your own notes and sources. This way, you created a personalized AI versed only in the material you cared about.

That was the key difference Google itself drew between NotebookLM and a traditional AI chatbot, and it's what left many (including me) impressed with the tool. Rather than the model giving you information from the web and its wider training data, it was restricted to what you handed it, and nothing else. Beyond keeping things tidy, this was also Google's answer to the trust problem that plagues every other chatbot, even today. Since the model could only draw from your sources, the risk of it confidently making something up dropped dramatically. On top of that, every response came with citations pointing back to exactly where the AI model picked up the information from. This way, you were never left guessing where an answer came from. If NotebookLM told you something, you could click straight through and verify it against the original.

So, grounding wasn't one feature among many. It was the thing that made the tool what it was, and separated NotebookLM from the rest of the AI tools. Audio Overviews, Video Overviews, Mind Maps, Slide Decks, all of it came later. NotebookLM began with a simple interface where you could upload a document and interact with it directly via text-based chat. Everything that came after was built on top of that foundation. For a long while, it was built around it, and never at its expense. Unfortunately, it feels like that's been changing a bit lately.

It started with Discover Sources

The feature I praised is the one I should've side-eyed

Now, I'm in no way claiming that NotebookLM woke up one morning and abandoned everything that made it good. The shift has been gradual, and I've admittedly cheered some parts of it as it happened. If I had to point to the first crack, it'd be the Discover Sources feature the company launched in April 2025. Up until that feature, you always had to populate a notebook yourself. You had to feed it the PDFs, Google Docs, YouTube links, URLs, pasted text, and so on. You could simply do nothing with the notebook until you did since the tool had absolutely nothing to work with. The Discover Sources feature changed that. Instead of manually hunting for material yourself, you could now just describe what you were after and let NotebookLM search the web for you. This feature is powered by Gemini (obviously), and it curates up to 10 recommended sources based on your topic. You can decide whether or not you want to import the sources it finds into your notebook, and exactly which ones make the cut. Nothing gets pulled in without your say-so.

While I cheered this on when it launched, I mentioned in my news article announcing the feature that part of me worries that NotebookLM could start hallucinating information. Looking back, that worry was pointing at something bigger than I realized at the time.

What you need to understand here is that NotebookLM's whole reliability pitch rests on a quiet version of "garbage in, garbage out." If the model could only ever draw from the sources you fed it, then the quality of what came out was bounded by the quality of what you put in — and crucially, you were the one in control of that input. You knew where every source came from because you chose every source. The model couldn't surprise you with something dodgy because it had nothing to work with except your pile.

Discover Sources changed that. Yes, you still ticked the boxes for which results to import, so on paper you kept control. But the finding, which is the part that decides what even makes it into your field of view, was now being done by the same AI that NotebookLM had defined itself against. You were curating from a shortlist someone else assembled. That's a subtle but real inversion of the original premise: the tool that was built to be limited to what you brought it had started bringing things to you.

After Discover Sources, features like Deep Research followed. Deep Research took the idea behind Discover Sources and pushed it considerably further. Instead of just shortlisting sources for you, it takes your question, creates a research plan, and browses hundreds of websites on your behalf, refining its search as it learns, before generating an organized, source-grounded report in a few minutes. In other words, the tool that once waited for you to hand it everything could now go off, do the reading itself, and come back with a finished briefing.

Again, NotebookLM's grounding instinct didn't vanish entirely here either. The report comes with its sources attached, and you can add both the report and those sources directly into your notebook, so you can still trace where everything came from. However, it was also clearly the point where NotebookLM went from a tool that retrieves answers from information you provide to one that fetches information for you and decides what's worth your attention in the first place. If you think about it, the entire pitch of NotebookLM essentially flipped.

The original pitch was that you were the one deciding what NotebookLM could see, and the model would only work with that. Deep Research changed the order of operations. Now the model does the searching, picks out what it thinks matters, and hands you a finished report. You still review the sources and choose what to keep, but you're reviewing a selection the AI made, not building the selection yourself.

NotebookLM's latest update takes this the furthest it's gone

It can write code now, because of course it can

Features that Google added to NotebookLM months ago certainly aren't what drove me to write this article. If that was the case, this article would have been published months back. The June 2026 update, which Google described as the biggest upgrade to NotebookLM yet, is what finally tipped me over. On June 8, Google made an announcement that essentially moved NotebookLM from a pure document-analysis tool toward an autonomous research agent that can write code, browse the web, and produce downloadable reports. It now runs on Gemini 3.5 and Antigravity, and each notebook comes with a secure cloud computer that can write and execute code for deeper analysis, along with access to more than 100 curated software skills.

And remember how I mentioned my reservations about the Discover Sources feature NotebookLM added? Unfortunately, this update doesn't just keep that idea around, it builds the entire experience on top of it. You no longer need to start with a pile of documents at all. You can now begin with a single question or a vague idea, and NotebookLM will go find, suggest, and organize the sources for you, building out the notebook before you've added a single thing yourself. You're still the one who ticks the box on what gets included and what gets cited, so the control isn't gone. But unfortunately, the default has officially flipped.

I miss the original NotebookLM

Speaking of this update, while NotebookLM's ability to write and run code is certainly impressive, i don't really understand why it needs to do. It feels like Google is trying to cram every possible feature into the same product, whether or not it actually belongs there. NotebookLM was a research and note-taking tool. I came to it to understand my sources, not to run Python on them! The original NotebookLM was good precisely because it did one thing and did it well. It knew what it was. Every update since has added something genuinely capable, but the more it can do, the less it feels like it has a center.