The current AI landscape is a dizzying frenzy of constant evolution. We're long past the initial shock and awe of generative text, and new offerings from Big Tech are largely shaped by user demand. It's a feedback loop working on overdrive, and pushing the limits for what's possible. Google has developed a purpose-built AI assistant, NotebookLM, specifically designed for students, academics, and researchers seeking to accelerate progress with limited and personalized sources of information.

This research assistant is well-grounded by your PDFs, YouTube videos, and other uploaded files, so you can enjoy the generative chatbot experience for anything like learning how to brew coffee, making a mind map, understanding insurance paperwork, or even improving how you utilize software tools. On the free tier, it’s a powerful ally, and I've used NotebookLM in most of these scenarios, but I repeatedly ran into the same set of challenges that need highlighting. Hopefully, resolution is speedy considering the fast-paced development cycle, but these are my three biggest qualms with NotebookLM.

Adding new sources and then revisiting the conversation manually

Back to square one

Here’s a scenario I’ve lived through more times than I care to admit. I’m deep into a research session, wrapping my head around a complex topic. I uploaded a few key documents to NotebookLM, crafted the perfect, multi-layered prompt to synthesize the core arguments, and mash Enter. The response is good, but it’s missing a crucial perspective. I follow up with another prompt, desperate to finesse the insight from the language model, but am met with a frustrating answer that doesn't have the info I seek.

I can easily remedy this by adding additional sources to the notebook since Google supports up to 50 of them per notebook. However, the AI doesn't take cognizance of the new information immediately. This holds true even if you feed Google Docs files as sources and subsequently update their content. The outdated copy is what the AI references. My initial set of prompts ate into my daily quota, ran into a dissatisfactory response limited by sources, and now I must revisit the same prompts to elicit a potentially useful answer, digging further into my quota. Sure, the chances of running out of NotebookLM's free prompts are slim for most users, but when you're in the deep flow state of focus, this kind of clunky, manual intervention is a jarring interruption.

Google could easily remedy this using a single button to revisit the last prompt after new sources are added, or to select older prompts to revisit with said alternative sources. The need to add them often arises directly from an unsatisfactory answer. The tool forces extra steps and burns another query precisely when I'm correcting a deficiency in its output. I'm technically penalized for refining my research.

In a long and complex research session, these duplicate queries add up quickly. It's a small pain point that, over time, becomes a major annoyance, turning a seamless dialogue with my documents into a disjointed series of stop-start interrogations.

A chat with forgetful AI

History be damned

NotebookLM is positioned as a conversational tool, an AI you can chat with to explore your documents. Yet, it suffers from a baffling case of short-term memory loss that completely undermines this premise. I can spend an hour in a rich back-and-forth with the AI, building on previous questions, asking for clarifications, and slowly teasing out the nuanced connections within my sources. Then, if I accidentally refresh or close the browser tab, that entire conversation vanishes into the digital ether.

Reopening any Notebook wipes the slate clean unless you have the foresight to save the important or valuable responses as notes within the tool. This is a stark contrast to every other mainstream AI tool, including ones in Google's portfolio, such as Gemini, where chat history is a fundamental, expected feature. I found this limitation firsthand when I opened my coffee brewing notebook to revisit the correct proportions of coffee ground to water I'd seen in a prior unsaved response. Sadly, it wasn't there.

Sure, I could prompt the AI right then and receive the answer again in seconds, but Google's officially recommended placeholder solution is to just save important replies as Notes within your Notebook. I believe this is a clumsy misinterpretation of exploratory dialogue. It forces you to constantly break your flow to decide what's noteworthy, like an administrative task. Instead, I'd greatly appreciate a searchable chat history feature.

In its current state, this volatility makes NotebookLM resemble a typical query-response engine. It discourages the deep, iterative analysis it's supposed to excel at. Until then, it's monologues, and I'm doing the remembering.

The word count wall

Sources too good for their own good

When I started using NotebookLM, the usage limits seemed permissive even in the free tier — 50 sources, each with up to 500,000 words. This should suffice for anyone trying to summarize content or digest a few chapters of a book. You can load up a decent chunk of material and get to work without ever worrying about hitting a ceiling.

However, the moment you try to use NotebookLM for the heavy-duty tasks it seems perfectly designed for, you slam headfirst into these limits. Case in point, reviewing the legalese surrounding taxation on your income, I found out firsthand. God forbid you're a PhD scholar analyzing a few dissertations or an engineer working with a 1,200-page tech spec manual. These monolithic documents, where an AI assistant is most needed, are where NotebookLM throws in the towel.

The only workaround is to manually split your massive document into smaller, arbitrary chunks before uploading them. That's tedious and fundamentally undermines the AI's usefulness as a tool. Sure, I could cough up the big bucks for Google's Workspace-tier accounts and associated AI benefits, which include more liberal usage limits. I'd doubt a PhD scholar or the average insurance buyer looking for clarity has that kind of cash sitting around. For now, NotebookLM is my brilliant librarian who refuses to read any book longer than 500 pages.

There is still hope

For all my griping, I’m not about to abandon NotebookLM. It's an incredibly useful tool, and these frustrations stem from hopes of seeing it become truly exceptional. I don't see any of these as unfixable flaws, especially considering Google's rapid pace of AI development. A robust NotebookLM is likely on the roadmap already. Features like one-click prompt reruns, persistent chat history, and higher source limits feel like logical next steps for a paid tier.