NotebookLM Is Overhyped
Google NotebookLM is being marketed as the greatest product. Repeated advertisement on YouTube forced me to try it. It is claimed to be the best tool for content creators. May be it is. But it fails on many fundamental aspects. The first and foremost is lack of editorial capacity. Next is privacy and censorship. Let me elaborate.
Political Fiction
Yesterday I accessed NotebookLM from the browser and uploaded three project notes from Obsidian vault folders. One was a fictional plot created from real life geopolitical events. It has characters like Donald Trump, Narendra Modi and Bejamin Netanyahu etc. I told the LLM that these are placeholders only to be replaced later.
It found the story and after analysing different notes found one complete story.
When I inquired if there are other plots in the notes, it woke up to ideas and replied that there are three more plots and listed those plot.
On asking how to use these alternative plots, it suggested three nice alternatives including a layered integration alternative it called ‘Russian Doll’ alternative of including multiple layers of conspiracy.
And the good news ended there. There were three questions as proposed prompt. One was to outline the story chapter by chapter to show proposed integration. A click on this option produced a buzz of cursor but no result for several minutes. After a while it answered that it can not answer the question. Then I asked it to give ideas as to where to integrate alternate plots? Same result. My guess is censorship was holding it back from assisting a project which was political in nature and had contemporary political leaders in it. Especially the President of USA.
I left it as it is. After two days I asked “How can you help me”. Voila, it turned out the structural design for the changes.
Now I am at loss why it denied earlier in a direct prompt. Helping user was more flexible option?
Privacy
For years I have seen advertisement posted by google matching my mailbox. Sometime it matches the mailbox of senders too. Years ago I saw advertisement of pre-owned car sales. I called up the person I was exchanging mails for several times and asked him if he was buying or selling car. He was surprised but admitted that he was doing it. This was first red-handed catch of google on privacy.
Ever since that incident, I keep track of what I do on google. Google’s surveillance is not something I accept willingly. DuckDuckGo is my default search engine. I have abandoned Gmail for anything personal. But Android and YouTube have no worthy alternatives. Google knows this. That captivity is the real privacy problem, not any single product. Yet I search those things on Google for which I want it to track me but mistakes too happen. The NotebookLM advertisement landed on me because I had been reading about articles on Obsidian and its plugins as I regularly use it for all my writing. The point is that Google is watching me and I am watching Google all the time.
The advertisements on YouTube persist even after manual request not to show a particular advertisement. I have stopped using gmail for any purpose other than to log on to website or google services. This lack of data has forced gmail to show very limited type of advertisements for years.
About the fiction project, it is on my hard disc synced with dropbox. Google did not know this before its upload. I am very careful of this fact. My Google cloud mostly have old files or junk. I am not using it for over 5 years due to same privacy concern.
The day NotebookLM was installed and fiction files were uploaded, next day YouTube was showing an advertisement of Sara, an app to write story board. Apparently NotebookLM passed on the information about my fiction writing ambition. This once again confirmed my long standing concern about privacy in using google products.
The fact that NotebookLM is also an online tool which process on remote server it can be compared with other AI even though google wants to market it as something different. In this comparison it is same. But with added concern of privacy.
Its AI Gemini frequently quotes my catch phase from one talk to another. It frequently call me a ‘practitioner of law’ in research I am making about some arms or drone without realising how stupid it looks and irritating. Therefore, all google products have serious privacy issue or should I say using a google product is like using a sieve to store private information.
Comparison
Any AI would have identified the entangled fiction files. Others may have done it in one instance but because NotebookLM has a default to create an output it did that first and ignored the complicated twists in story created in separate notes.
Large books.
It is claimed that NotebookLM can process upto 50 large files or even books in multiple formats like pdf and docs etc. I uploaded 19 files in which was a book of about 250 pages rest were small documents of under 9 pages. These were articles on Operation Sanskrit Mill which I have written about. NotebookLM took that idea and files to integrate an article. It had nothing from the book as it did not have a single passage about Sanskrit. The book was about colonial rule in India. NotebookLM couldn’t make a connection.
Another Project
This time I uploaded two articles on Bhirannha and Rakhigarhi archaeological discovery and a research paper. It nearly reiterated what was stated in there. When I asked “does this discovery demolish Aryan Invasion Theory?” It told me to verify independently. When I asked “do you know what Aryan Invasion Theory and it’s nuance?” This woke up the inner logic and then it understood the logic. Hindu civilization was either continuous here for more than 1500 years or it was established by migrants from abroad.
It kept retreating to disclaimers: “this is outside the provided sources,” “you may want to independently verify,” “I must note that the provided sources do not explicitly mention.” Every time you pushed toward a conclusion, it added a disclaimer. It understood the evidence but refused to own the inference. This is a design choice, not a knowledge gap. The model was trained to hedge rather than synthesize. It could assemble facts but could not build an argument. That is a fundamental limitation for any serious research or writing task.
At last I added the seminal work of K.N. Dikshit that is a paper from Abhidha Journal 2022. He is a former Joint Director General of the Archaeological Survey of India. His thesis is titled ‘Saraswati Valley Civilization.’
It made NotebookLM a little considerate and it wrote a short note on the subject suitable for homework in class 10.
Writing Skill
It has no drafting skill. It is surprising considering the Gemini does far better. When it did draft a compressed article from the upload about ‘Sarasvati Valley Civilization’ it had problems. Every sentence was a standalone declaration. There was no subordination, no causality built into the syntax, no sentence that earn its place by connecting to the one before it. “It was never imported by invading foreign armies” follows “The Sarasvati civilization grew continuously from native Indian roots” as if repetition is emphasis. It is not.
When instructed to keep sentences under 19 words and avoid em-dashes, the model followed the formatting rule mechanically but missed the purpose behind it. Short sentences are supposed to create rhythm and punch, not just brevity. What it produced reads like a listicle with the bullets removed. Each sentence lands with the same weight, which means nothing carries more weight than anything else.
NotebookLM rearranged the source sentences, thinned the vocabulary slightly, and called it writing. No new sentence earns its existence. No inference travels beyond what the source already stated explicitly.
That is plagiarism with a thesaurus. The model does not synthesise. It redistributes. The difference between a writer and a photocopier is that a writer transforms source material into an argument that could not have existed without the writer’s mind working on it. NotebookLM skips that step entirely. It gives back a compressed version of what it received, dressed in slightly different word order.
For a student submitting homework this may pass. The teacher sees structured prose covering the right topics. For a writer it is useless because the one thing writing requires, a mind that does something original with evidence, is precisely what the tool withholds.
The Project
Every AI model has a feature called ‘Project’. Chat GPT and Claude has it. In this feature user can upload files in a specially allocated window and ask the AI to perform analysis or whatever user may choose to decide. NotebookLM is a standalone concept of that ‘Project’ window in other AI and being marketed as a standalone model.
It has capacity to create in many mediums like audio and video, which I have not tried and I hope it will be fine with that but that relate to easy to use interface not the underlying feature as all, AI can perform those tasks.
Future
It may and hopefully may improve in future as the technology grows. It has interesting and easy to use interface. It has limited use today but in future it may evolve into something more powerful. But today it is not what it promises to be.
However, the lack of privacy of user data is and will remain serious issue.
Conclusion
NotebookLM is suitable for helping students cheat on their homework. It is not a serious tool for professional work. NotebookLM has three serious problems.
It filters political content without explanation. It feeds user data into Google’s surveillance machinery. It redistributes source material without transforming it, which is plagiarism by another name.
A student can use it to summarise. A professional cannot use it to think.