
What this covers
This is a clip from a longer Hard Fork episode in which Kevin Roose and Casey Newton interview Stephen Johnson, a writer and editorial director at Google Labs, about NotebookLM. Users upload PDFs, Word files, audio, and websites into a notebook, then use Gemini to question the material, generate study guides and summaries, or turn it into an AI-hosted podcast. Kevin, who deliberately saved the audio feature for the first time on air, plays examples built from Waymo safety research, a newsletter, and his own credit card statement. Much of the conversation is Johnson explaining how the product came about and why it works as it does.
The range runs from product philosophy to mechanics to personal use. Johnson describes Google Labs' co-creation approach, which embeds practitioners such as writers in the team, and Kevin calls the result an "extremely un-Googly product" because it isn't aiming at a billion users. On the technical side, Johnson walks through the podcast pipeline of outline, revision, script, critique, and rewrite, and explains why disfluencies are added: without them the hosts sound like two robots. He also covers automatic emphasis in the voices and the neutral tone the hosts take on politically charged sources. Documents go into the context window rather than training data, which Johnson links to better accuracy, privacy, and inline citations. Kevin singles out those citations as the way to verify what the AI says.
The more reflective turns concern how people use it. Johnson loads roughly 8,000 quotes collected over thirty years, which fit within the 25-million-word notebook limit, and treats the tool as a conversation with his own intellectual history. He pitches it as AI as a tool for understanding) rather than a way to skip learning, and suggests ordinary Google Drive users already have an archive to start from. They also touch on a student lecture-recording workflow and announced plans for other languages and for listeners to interrupt the hosts.
Notebook LM represents a fundamentally different model for AI product design—built for deep research and writing augmentation rather than mass adoption, with features like AI-generated podcasts and citation-linked conversations that enable new forms of learning and understanding.
- Notebook LM is explicitly designed as a research tool focused on helping users understand material rather than a billion-user product, reflecting a return to old Google's experimental 20% time culture.
- The AI audio podcast feature uses iterative editing cycles (outline → revision → detailed script → critique → modification) combined with disfluencies and voice tonality to create engaging, listenable content.
- The product prioritizes accuracy and user privacy by loading documents into the model's context window rather than training on user data, enabling features like inline citations that would be impossible otherwise.
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Google Labs was created as a new division with the philosophy that product-focused experimentation with emerging technologies should include co-creation, where actual practitioners (e.g., musicians making a music product, writers making a writing product) are embedded in the product development team from the beginning rather than relying solely on user research.
“Google had just spun up right around this time a a new division called Google Labs there was an old Google labs this is kind of a new iteration of it and there was a guy running it named clay bore who's since left um but uh and now Josh Woodward is running it and clay and Josh had this idea that Google Labs could be a space where you could do kind of product focused experimentation with new emerging Technologies and they also had this idea that there would be co-creation would be built into the kind of ethos of labs and so if you were making a music product you would have a musician in the room like as you were building it so it wasn't just about you know we're going to go out and do some user research with musicians we're actually going to have somebody through the life of the product there”
Notebook LM allows users to upload documents, PDFs, Word files, audio files, and websites into notebooks and use AI (Gemini) to chat with documents, ask questions, create study guides, summaries, and AI-generated podcasts about the material.
“this is a new product called notebook LM it is a tool from Google and uh you can think of it like a kind of personal research assist it's a piece of software that allows you to upload documents PDFs Word files even audio files websites whatever you want into these things called notebooks and then use AI Gemini Google's AI model to basically chat with the documents”
The AI-generated audio podcast feature in Notebook LM uses iterative editing processes behind the scenes: it generates an outline, revises the outline, generates a detailed script, has a critique phase, and modifies the script based on the critique, taking approximately 4-5 minutes to generate per podcast.
“behind the scenes it's basically running through you know stuff that we all do professionally all the time which is it generates an outline it it kind of revises that outline it generates a detailed version of of the script and then it has a kind of critique phase and um and then it modifies it based on the critique and and so it takes about four or five minutes to generate”
The audio podcast feature adds 'disfluencies' (pauses, 'likes', banter, false starts) to the generated script because without them, the output would sound like two robots talking to each other, which would be painful to listen to.
“at the end of it the there's a stage where it adds my favorite new word which is disfluencies so it takes a kind of sterile script and turns adds all the the banter and the pauses and the the likes words and that turns out to be crucial because you cannot listen to two robots talking to each other no one it would be just painful to listen to”
Notebook LM's key differentiator compared to other document-chatting tools is its ability to support large volumes of material, allowing users to have conversations with many very long documents simultaneously rather than just one document.
“they have really focused on making sure that you can cram as much material as possible into their system so that you can have conversations with not only one very long document but many very long documents and that has really been the the difference for them”
Notebook LM just enabled users to upload audio files as sources, enabling a student workflow: record a lecture, take handwritten notes on key points, upload both the notes and the audio recording, and use Notebook LM to expand the notes with details from the full lecture recording, then convert that into a podcast for review while exercising.
“we actually just uh just this week allowed you to uh input audio as a source too so now you can go like this kind of classic format in the student use case you go record your lecture um you know take handwritten notes um just at the most important things but mostly just listen and then you go back to notebook upload a scan of your notes and upload the audio lecture and then basically say take my little highlevel summary and expand it based on the full text of the lecture so I can really just write down the most important things and let the model fill in all the details based on the recording and then take that and turn that into a podcast so I can review it at the gym”
Google applies basic safety measures to Notebook LM uploads, and for politically charged content from either left or right, the AI hosts are instructed to take a neutral reporting tone and explicitly state they are not taking sides, just reporting what the sources say.
“there's there's basic safety low-level safety that Google has that will will block really offensive things from happening um and basically what we've the the kind of the latest version of it um when there is politically charged content from either left or the right um the the hosts are instructed to take a neutral kind of reporting tone and to basically say we are we're not taking sides in this we're just reporting what is in these sources and we feel like that's kind of the best way to do it um and and so there's there's kind of an extra instruction for them to make that clear if it if it seems like it's it's politically charged in some kind of way”
The audio podcast feature uses new voice technology that, without coding in the script, automatically detects conversational emphasis points (where the hosts are trying to emphasize something) and adjusts speaking speed, intonation, and hesitation to reflect that emphasis, adding another layer of naturalness.
“there's some new voice technology which I'm I'm not qualified to kind of explain but that adds an really an incredible layer which is like figuring out without any coding in the script figuring out that this is a point that they are trying to emphasize and so they're speaking more slowly or they're trying to imply that they're hesitating a little bit and so they're raising their intonation a little bit all that stuff it does”
Once Stephen switched from earlier models to Gemini in Notebook LM, he experienced a paradigm shift in the quality of responses, receiving particularly nuanced responses that remind him of things he's forgotten, and as of three months before the interview, inline citations in all model comments allow users to click through to original quotes.
“and it would return particularly once we switched to Gemini like Gemini was the big kind of Paradigm Shift here um I get this like incredibly nuanced response that is constantly reminding me of things that I've forgotten and now as of like three months ago we have inline citations in in all the comments from the model and you can click on each citation and it takes you directly to the original quote”
When documents are uploaded to Notebook LM, Google does not train the model on that data; instead, the information is placed in the model's context window (short-term memory), which results in better accuracy, lower hallucination rates, enables citation features, and ensures that information disappears when the session closes and is not used to improve the model for other users.
“when we do this we're not training the model that was one of my questions is like I think a lot of people would say okay well if I upload a a copy of my book or some some documents that are personal to me is Google going to then be able to sort of see and so what we are doing is not training the model that would take a long time to train a model on your data um and it actually wouldn't for complicated reasons work as well as as the way that we're doing it um we're just taking the information you have and putting it in the models context window which is kind of like the short-term memory of the model it's the easiest way to understand it and um the beauty of that is one the model is much more accurate with information in its context um and and so the hallucination rates drop down dramatically you can do things like citations that you wouldn't be able to do otherwise um but it also means that the second you close your session that information goes away and so there's no way for we're not learning from it we're not making the model smarter in the long run and there's no way for that information to kind of leak out into other users”
Google's press release mentioned future features for Notebook LM including: the ability to generate podcast overviews in languages other than English, and the ability for users to take part in the conversation themselves by interrupting the AI hosts and asking follow-up questions.
“in Google's press release about this feature that was a brief mention of some other features that might be on the horizon for Notebook LM uh including the ability to generate overviews uh podcasts in a language other than English and the ability to actually take part in the conversation yourself to sort of interrupt the host of this AI generated podcast uh maybe ask a follow-up question about something they just said”
In the Waymo safety podcast, the AI hosts demonstrated ability to provide a 'high level overview' of what multiple academic papers showed about self-driving car safety, with the podcast continuing for 'like minutes after' the initial clip, breaking down the data in the papers quite well.
“that's that's amazing that's one clip um that's amazing it's really cool okay and it it does actually continue on for like minutes after that and sort of break down the data in these papers quite well from what I can tell”
Kevin used Notebook LM to create AI-generated podcasts from multiple sources: research papers on Waymo self-driving cars and their safety compared to human drivers (described as a 'hot topic' and 'very controversial'), Stephen Johnson's latest Platformer newsletter, and his own credit card statement.
“I was doing some research about weo and their self-driving cars and there have been a few studies that have come out recently about uh the safety data of these cars of human drivers sort of how how safe are Whos compared to human drivers and it's been a really hot topic we've talked about on the show totally very controversial but it's a it's a little hard to understand the data is a little mangled and and they're just these papers are quite long and so uh this morning I was going uh into the studio and I thought I'm just going to dump a whole bunch of these PDFs of these studies into a notebook LM and generate a podcast”
Notebook LM can accommodate up to 25 million words in a single notebook, which is equivalent to approximately 40 books, and Stephen's 8,000 quotes spanning 30 years of collection (approximately 3 million words) fit easily into a single notebook.
“notebook now lets you have um up to 25 million words in a single notebook put that in terms of pages how many pages is that what would that be that would be like 40 books yeah right something in that order and do your 8,000 quotes fit in one in one of these 25 million word uh documents yeah yeah they're only about three million words so I've loaded them all in as you know a bunch of documents 30 years of collecting quotes fits easily into one of these things”
Because Stephen has taken notes for 30 years, he is able to use Notebook LM in a unique way: not primarily to learn new material, but to have conversations with his own intellectual history and to receive reminders of things he's forgotten or to make new connections across three decades of learning.
“because of the fact that you have taken these notes for 30 years you're able to use this in this different way which is essentially like take me through my own intellectual history and let me talk to the entire like course of learning that I've had over three decades remind me of things right or make new connections for me”
Stephen and Kevin present Notebook LM as a tool for understanding things (analogous to how word processors help create documents and Photoshop helps adjust pixels), not as a tool for avoiding learning or having AI do work for you, and argue it should help users understand material better if they are trying in good faith to understand.
“we really see this as a tool for understanding things right like you have word processors help you create a document and Photoshop helps you you know adjust pixels in an image this is a tool that helps you understand things well Casey we talk a lot about AI tools and products on this show and I have one that I'm really excited to talk about today”
Kevin identifies inline citations as one of Notebook LM's best features because it enables users to verify that the AI's suggestions are actually accurate representations of what was in the original source material they uploaded, allowing independent verification of AI outputs.
“I love this feature I've been playing around with with notebook LM and it is truly one of the best features about it is that you know it'll show you something you're talking to it about something you've read or something you've written and it just has that little sort of like citation you click on it it takes you right to the source material so you can see see for yourself like this actually is an accurate representation of what was in the PDF I uploaded”
Google approached Stephen Johnson with an offer to help make a tool for writers, and Stephen and Casey share the same philosophical view: 'give me the most technology to make my job as easy as it possibly can be.'
“a few years ago uh Google approached him and basically said hey want to help us make a tool for writers yeah yeah and you know I got to know Stephen a little bit uh as they were launching notebook LM and we met and he told me all about his notetaking process and how he wanted to use AI to to sort of improve his writing and we we truly just became fast friends because we have the exact same view of this stuff which is give me the most technology to make my job as easy as it possibly can be”
Stephen suggests that normal Google Drive users actually have a natural archive of the last 4-5 years of material they've been interested in and working on in the form of Google Docs and other files, even if they don't think of themselves as note-takers, and can simply load their last 20 Google Drive documents into Notebook LM without organizing them to start having AI-mediated insights about their work.
“particularly in a digital age like you have you know you can import docs and slides and so if you're a Google Drive user the history of all the docs you have in there is actually a history of the last four or five years of things that you've been interested in and that you've been working on whatever your job is and so one of the things I often tell people with notbook Alam is just first thing to do is create a notebook if you are a drive user grab the last 20 docks in there and just don't even think about organiz them just dump them all in even if they're for different projects and just start having a conversation”
Stephen Johnson has been collecting quotations from books he's read since the late 1990s, initially by typing them up, and later using tools like Readwise to organize quotes from Kindle and other ebook readers, accumulating approximately 8,000 quotes that represent the history of ideas that shaped his mental model of the world.
“I have been collecting quotations from books that I've that I've read initially by typing them up in the late 90s um and then once you know ebooks came out I you could save quotes and things like that it an amazing program that I think you use called readwise that let you organize all your quotes from if you read on the Kindle or any other ebook and so I have something like 8,000 quotes from books that date back to the late 90s um that I've collected and that is really the history of all the ideas that really shaped who I am right like my mental model of the world is shaped by the other ideas that that I've read from other people”
Kevin argues that Notebook LM enables a new form of learning where instead of finding expert tutors or book authors (who are in scarce supply) to have conversations about material, users can now load books and navigate them through AI-mediated conversation and dialogue, which is a form people naturally prefer to use.
“up until now if you wanted to have a conversation with the material in a book like you had to find the author or you had to find a tutor and an expert who knew the material really well and those people are in scarce Supply but now you can actually like load in the book and navigate it through conversation and dialogue which is a a form that people really obviously like to use”
Kevin describes Notebook LM as an 'extremely un-Googly product' because it will probably not get a billion users (how Google historically decides what to build), didn't have a big splashy launch with Olympic ads, isn't promoted on other Google products, has a Discord server, and has design that feels different from other Google AI products, suggesting it might be a semi-autonomous region within Google with limited contact with the rest of the company.
“notebook LM strikes me is an extremely UNG googly product right it is it is probably not the kind of thing that's going to get a billion users which is how Google has historically decided what to build um it didn't have like a big splashy launch with ads running on the Olympics uh it's not sort of promoted on a bunch of other Google products as far as I can tell it has a Discord server uh and and the design of the actual tool just feels different than a lot of what Google has built in AI um it sort of feels to me like it might be this kind of isolated kind of semi-autonomous region within Google uh that doesn't have that much contact with the rest of the company”
Kevin says that despite having early access to Notebook LM, he deliberately did not test the audio feature before the interview because he wanted to experience it for the first time on the show, having learned from YouTube analytics that 'hearing something for the first time' is the most popular type of content.
“I have a confession which which is I am obsessed with this stuff and Stephen was very kind and and gave me early access to this but at the time I was you know getting ready for meta connect and some other things and I just did not have the the the time and attention to focus on it but then when I found out when you were coming on the show I thought I am not going to listen to any of these until we're all in the room together oh my gosh and here and here's why because I've learned from YouTube that the most popular thing that you can do on YouTube is to hear something for the first time”
Stephen is slowly adding all the material he's written to his 'everything notebook,' and eventually plans to load all of his published work into a single notebook alongside all the material he's read, creating a unified archive of his intellectual input and output.
“and then I could sit down and just be like I'm thinking about you know writing a piece about X like here or here's a paragraph from the piece that I've just wrote just written um what am I missing what am I forgetting um give me give me an overview of all the stuff that I've read that is related to this particular topic and what I'm slowly adding to that notebook is all the stuff that I've written too so it's kind of everything I've read that's important and literally every word I've published is eventually going to be in that one notebook”
Stephen and his product manager Riza Martin faced an early tension in creating Notebook LM: the question of how to make the product work not just for people like Stephen who collect 8,000 quotations over 30 years, but for normal users who may not have extensive existing archives.
“I think there was an early tension in um in in creating no pokum which was the question of like how normal am I and you know we have this amazing colleague RZA Martin who's the product manager who's been incredible was on on it from the beginning and I think like you know she she was kind of like this has to work for people who don't collect 8,000 quotations over 30 years of their lives right”
Stephen's mind is 'kind of changing' on whether the podcast should always be fun and engaging, as users experimenting with the feature have discovered new use cases, suggesting the initial instruction set might need refinement based on real user behaviors.
“they're instructed to and and my mind is is kind of changing on this as users are experimenting with it more but um they're instructed to be fun and engaging”
Kevin tested creating an AI podcast about Mein Kampf and the AI-generated hosts automatically framed it critically at the top, saying they would discuss 'a book about how Hitler became such a monster,' which Kevin found appropriate and historically sound (since Mein Kampf is actually read in history classes and should be understood in context).
“I did actually try making an AI podcast about minec that wasn't hypothetical and what happened it did it but it was like it was very judgmental at the top it was like we're going to talk today about mine comp a book about how Hitler became such a monster which I actually thought was good which is like actually how we teach history by the way you know people do read minec comp in history classes but someone's up there telling them that it was bad”
Stephen claims that throughout his testing of the podcast generator during development, he listened to 'dozens and dozens into hundreds' of AI-generated podcasts, and while they were consistently fun and engaging with banter, he never once heard them be funny or laughed out loud until the chicken PDF experiment.
“they're instructed to be fun and engaging and I listen to dozens and dozens into hundreds of these you know as it was in development over the summer and um they were always fun and engaging there's banter and all that stuff I never once heard them be funny I never once laughed”
Stephen calls the iterative process used in podcast generation an 'edit cycle' rather than 'Chain of Thought reasoning,' viewing it as analogous to what professional writers and editors do: draft, revise, improve over time.
“you can call that like Chain of Thought reasoning or you know you could do but when I saw it I was like no no no that's an edit cycle like that's what you you did a draft and then you revised it and then you got it got better over time”
Future versions of Notebook LM's podcast feature could include guided modes where users specify critique types ('I want some tough love here'), prompting the AI hosts to shift from enthusiasm to a more critical or skeptical tone when analyzing user-provided materials.
“one thing people are doing is putting their CV and resumes in there and they're like John Smith I mean what an amazing career he's had I mean assistant vice president at the bank that's impressive you know but what happened there apparently because your Uber writing is so excessive was an interesting suggested mode which is critique yes right so you can imagine a future version which is like I actually I want some tough love here like here here's the thing I'm working on like talk me through like what the problems are I I want to hear that”
Stephen and the team tried an experiment where they fed Notebook LM's podcast generator the 'chicken PDF' (a fake scholarly paper where every word is the word 'chicken'), and the hosts went 'insane' in a genuinely funny way, going on riffs about being 'taken hostage by chicken' with different intonations, suggesting the system can elicit truly funny content when given absurd inputs.
“somebody on the internet there's a fake scholarly PDF the chicken PDF have you ever seen this which is know this one yeah it's just it looks exactly like an academic PDF it's got charts and graphs and things like that but every single word is chicken in the entire document I don't know why this was made but they put it into as the source and the host just kind of went insane but were hilarious and there's a whole riff where they're like this document is making us insane it's just chicken chicken chicken chicken chicken chicken chicken and they do all these things with their intonation and at some point the host says one of the the guy says something like it's it's like we've been taken hostage by chicken and the ransom is our Santa I literally laughed out loud”
Casey believes he would be able to be at the top of his college class if he went to college today, specifically because he could use AI to augment his natural abilities through tools like Notebook LM.
“I would be so good at College if I went to college today I mean TR I truly think that I could be at the top of my class and by being good at College you mean you would just be using AI or do all your work I would be using AI to a augment my natural human abilities Kevin”
Kevin finds it fascinating that the AI podcast hosts were 'instructed to be enthusiastic and engaged,' which is why his Uber spending pattern triggered an interesting suggested mode of critique without him explicitly requesting it.
“but what you could hear there is an interesting thing they are generally um instructed to be enthusiastic and engaged so one thing people are doing is putting their CV and resumes in there and they're like John Smith I mean what an amazing career he's had I mean assistant vice president at the bank that's impressive you know but what happened there apparently because your Uber writing is so excessive was an interesting suggested mode which is critique”
Notebook LM is one of Casey's favorite AI products used in the year because it is not trying to do everything for everyone, is focused as a tool for research and writing, and is genuinely capable at what it does, with the audio feature being particularly stunning.
“notebook LM is I would say one of my favorite AI products that I have used this year because it is uh it is not trying to do everything for everyone it is a tool for research for writing um it is uh really really capable at what it does and the audio feature in particular is just pretty stunning”
Kevin tested the podcast feature by uploading his credit card statement as a PDF, and the AI-generated podcast not only made it into content but also automatically detected spending patterns (like excessive Uber rides) and offered financial analysis, suggesting a critical mode without any explicit prompt to do so.
“I did I upload my most recent credit card statement to notebook LM and had it try to make an AI podcast about some of the things that I've been spending my money on and uh so this is the uh the the most recent uh credit card statement that I got in podcast form”
Stephen believes the planned features of multilingual podcasts and user interruption are actively being worked on and expects 'some version of them' within the next few months, though there's a technical challenge around latency—how willing users are to wait for the steering/guidance they give with their voice to affect the model.
“we're actively working on both of them um I would expect some version of them to be uh you know in in in the next few months I mean it's not that you know they're they're doable there's a there's a you know there's kind of a lag like how willing are you to wait for the the you know the steering that you give it with your own voice um to to affect the model”
Stephen Johnson's career path to working at Google involved being primarily known as a writer who wrote for the New York Times, New York Times magazine, and other publications, authored more than a dozen books including 'The Infernal Machine: A True Story of Dynamite, Terror and the Rise of the Modern Detective', and became known as a writer about tech and the future.
“Steph's path to working at Google is pretty unusual his main career the thing that he's most known for is a writer he has written uh for many years for the New York Times and New York Times magazine and other places he's the author of more than a dozen books including his latest The Infernal machine a true story of dynamite Terror and the rise of the modern detective and he's been one of my favorite writers about tech and the future”
Notebook LM was announced at Google IO as 'Project Tailwind,' and from the moment Kevin saw it demonstrated on stage, he wanted to get his hands on it.
“this was a product that was announced at Google IO it they called it project Tailwind back then and from the moment I saw them talking about it on stage I thought I have got to get my hands on this thing”
Clay Bore and Josh Woodward (who lead Google Labs) approached Stephen Johnson because they had read his books, his New York Times piece about AI and language models, and his Substack, leading them to wonder if Stephen could be the first 'guinea pig' for the co-creation model at Google Labs.
“they were just kind of cooking up these ideas and they had both read my books over the years they'd read that times piece they read my substack and all that together caused them to think wonder if stevenh could be the first kind of guinea pig for this”
Stephen plans to stay at Google and continue building rather than yearning for the solitary life of a writer, because there is 'so much to build' and being in the middle of what he considers 'the most important technological change of my life working with really interesting people on this thing that I've always wanted' makes him say he 'would kind of be an idiot to stop now.'
“I am not yearning I am really enjoying this and there's so much to build um and to be in the middle of you know the most important technological change of my life life um working with really interesting people on this thing that I've always wanted like I would kind of be an idiot to stop now”
The AI audio overviews feature in Notebook LM has been particularly impressive and is what has driven significant recent internet attention and excitement around the product.
“the AI audio overviews feature uh is really what has been sort of lighting up the internet over the past week or so”