NotebookLM launches feature to customize and guide audio overviews
blog.google
blog.google
https://github.com/ListenNotes/ai-generated-fake-podcasts/bl...
Google is taking a different approach this time, moving quickly. While NotebookLM is indeed a remarkable tool for personal productivity and learning, it also opens the door for spammers to mass-produce content that isn't meant for human consumption.
Amidst all the praise for this project, I’d like to offer a different perspective. I hope the NotebookLM team sees this and recognizes the seriousness of the spam issue, which will only grow if left unaddressed. If you know someone on the team, please bring this to their attention - Could you please provide a tool or some plain-English guidelines to help detect audio generated by NotebookLM? Is there a watermark or any other identifiable marker that can be used?
Just recently, a Hacker News post highlighted how nearly all Google image results for "baby peacock" are AI-generated: https://news.ycombinator.com/item?id=41767648
It won't be long before we see a similar trend with low-quality, AI-generated fake podcasts flooding the internet.
Maybe you have the problem backwards - we accidentally end up listening to non NotebookLM podcasts?
It was factually accurate, and presented the topic in a manner that was easy to digest and kept it interesting.
I didn't plan to but ended up listening to the whole thing, and I normally don't enjoy the podcast format.
For someone new to the topic, it'd be a pretty great intro compared to reading the official pages.
In terms of actual usefulness, it’s one of the AI tools that most impressed me.
The main issue is, of course, privacy.
I have tried to reproduce something similar using AnythingLLM and the low tier Llama models, but of course the experience is much worse, both in terms of results, response times and UI. If someone knows of a better local setup, I’m all ears.
I would consider a Workspace subscription if I could actually trust Google to make good on the commitment of not reading your stuff, which I’m finding hard to do…
I've always absolutely hated the focus on the individual humans and their personalities behind the podcast, and wished they'd be a better source of well-structured "information".
I never listened to a podcast I didn't get frustrated with, even at 2x speed. These NotebookLM podcasts have been exactly what I've always wished podcasts were.
I may have to re-task my paperclip making AI on generating podcasts...
On the plus side, the speech is almost perfect; so good, that I sincerely hope the voices themselves are never fully under user control.
With regards to the actual summary of the content I gave them, I would say they are grade B: only mostly correct, they're still inventing things I didn't say and missing things I did say.
That's not to say humans don't make mistakes, I still consider this objectively impressive, that is able to reach even this level was SciFi when I was a kid — but why waste time on a grade-B podcast when the AAA-tier costs you as a consumer a 30 second advert?
If a AAA-tier podcast on the subject you want to listen to a podcast about exists (and you know about it), then that's probably a better (and obvious) choice for your listening time.
However, if you want to listen to someone discuss or explain something and you don't know about a AAA-tier podcast, it's possible that a generated podcast is better than nothing.
On the other hand, it's also possible that the generated podcast will miss or hallucinate a key detail, and herein lies the dilemma. Is it better to listen to something that might get something wrong, or not to listen and perhaps someday to learn about the subject through some other form that is less likely to include mistakes?
If AI texts have a similar effect then AI = bad is quite correct.
Edit: If I'm understanding it correctly after some googling, supposedly the "name" of all podcasts generated by NotebookLM is "Deep Dive"? That's just confusing.
Its trained on too many shallow podcasts. Go compare any of NotebookLM podcast with an episode of Hardcore History. The latter goes into much more depth (even when you account of it being much longer).
If AI ends up destroying humanity, it isn’t going to be through weapons and death robots, but just by entertaining and placating us all to death.
I'm an avid podcast listener, but I already ignore 99.9% of podcasts out there. I'm not concerned that this is going to become 99.99%.
If these AI generated podcasts are all bad, I will just continue to ignore them. If some turn out to be good, it seems like a win to me.
If you're worried about an existential "what happens to the world if all media is machine generated", I guess I'm willing to hop on the ride and see what we find out.
(Also - ironically, one of the podcast out of those 10 that I listen to regularly - it's the Deep Dive on AI. A NotebookLM production! )
If people want to listen to AI generated podcasts, they can just make them themselves. They don't need publishing on a platform alongside human-made podcasts. If I was Apple, who ultimately control curation of podcasts, then I'd prevent them. After all, Apple Intelligence will soon do as good a job of making your custom podcast if that's what you want.
As an exception, consider Infowars. Now imagine someone 10 times smarter, maybe even with no monetary goals.
NotebookLM seems wonderful for digesting various content in an alternative way. It’s not a “fake podcast” either.
Nobody is saying that the audio output should or should not be published somewhere. That’s a user decision for both publishing and subscribing.
Indexes and discovery on the internet is where you advocate policing instead of nit picking a useful tool.
Like other data, provenance suddenly matters a lot. From my POV, that's good. Not all data sources are created equal, and this is putting it into stark enough relief it might actually change the landscape. (In case it isn't obvious, I strongly believe most of the Internet was garbage well before LLMs. We just called it "SEO". Still garbage)
Yet more "but humans also".
I’d rather let the end user know if it was made by humans or not, and let the marker decide. If people love listening to such content, let it be. But hiding how it was made, feels a bit disingenuous.
Or like we should go back to carts because cars are noisy and not only that but might collide with pedestrians and not only that, might even collide among each other.
Instead of containing the tools and curtailing the progress (email and cars) we should probably try to contain and curtail abusers. Very hard to do, I know but the right thing to do.
You're assuming that everybody shares your opinion of cars being progress and/or progress being good - you're assuming too much.
But consider it from Googles perspective and this is why I think they won’t care, serving snippets and caches of articles had rights holders attacking them, serving thumbnails of images has rights holders attacking them, serving tiny bits of songs in the background of video had rights holders attacking them.
Serving AI doesn’t, I don’t think the current management at Google will care if Google shows fake baby peacocks as long as it can serve them without being bothered by rights holders, same way a Gemini summery can launder article information.
What's new? Every novel class of genAI product has brought a tidal wave of slop, spam and/or scams to the medium it generates. If anyone working on a product like this doesn't anticipate it being used to mass produce vapid white-noise "content" on an industrial scale then they haven't been paying attention.
What I’m aiming for is to ensure that the NotebookLM team is aware of the impact and actively considering it. Hopefully, they are already working on tools or mechanisms to address the problem—ideally before their colleagues at YouTube and Google Search come asking for help to fight NotebookLM-generated spams :)
It's certainly easier for the creators of genAI to build detection tools than for outsiders to do so. AI audio detection is a hard problem - https://www.npr.org/2024/04/05/1241446778/deepfake-audio-det...
What is the impact? Have any of them attracted an audience of any meaningful size? If a month from now there are 1.3 million generated podcasts, what do you anticipate the fallout to be?
You can even see it in the case of e-mail spam, where while technically not platforms, it helped Gmail and Hotmail to become so big that they can basically impose their own rules on the protocol.
Is this a rhetorical question? Because the answer for podcast indexing and search services is surely pretty obvious.
I just don't understand how people can pretend this isn't happening just because they find each new twist of a technology fascinating.
It's not obvious at all. You see a problem, I see opportunities. The person you responded to mentioned Reddit. Well, the glut of garbage on the internet makes Reddit more valuable.
The "please, god, find me something a real person actually said about this real thing" isn't actually a problem very many people have. Despite all the complaining here, most people are happy with the results they get when they search Google or Bing or Facebook or Reddit.
So like every technology, there are positives and negatives. It isn't clear to me where this all nets out but I'm leaning slightly positive.
And if you don't find this technology fascinating, then I'm not sure what to say.
If given the choice between a dime and a fiver, is the fiver somehow worth more than if I'd offered you the choice between a fiver and a tenner instead?
The problem with this is it's not feasible long-term, or even medium-term - as soon as a watermarking system is implemented, a watermark-removal system will be created.
(Happy to be proven wrong)
Unfortunately Kagi image search is polluted with the same crap. I'd started to trust it recently but not so sure now.
[Edit] This is true even when you specifically use the 'exclude AI' filter
Of course, the commenter never says this, living in the world today, where the writing he likes would never be published by the New York Times like it is on Twitter, the TV he likes would never be offered for free like it is on YouTube, and the music he likes would never been offered for pennies on Spotify. Some meaningful creators will lose from every remedy you could think of, where Google "something somethings" AI. Maybe the root problem is generalizing.
Whether that is saying how high quality NotebookLM is or how low quality NPRs podcast are is an exercise for the reader.
The only reason “Stuff you should know” is better is because of the random off topic discussions they go into and that’s not a complaint about SYSK.
The 1,300+ shows are just the ones recently removed from Listen Notes.
Give it a few days, and I’m sure the number will double, quadruple, and continue to grow. :(
If you care a lot about podcast quality, go and make your own podcast service with better discovery. Once you realize the antagonist was collaborative filtering, made possible by non-negative matrix factorization dating from the year 2000, and not AI, you will at least have learned something from the comment, instead of just feeling better. And then, how do you propose to curate by hand, and why would someone choose your curation over the New Yorker's? And maybe those very purists, trying to make everything sentimental, accusing everyone of slop and spam - well, why do so many creators thrive and ignore the New Yorker's opinion about them entirely? Perhaps curation is not only not scalable, but also wrong. Difficult questions for listeners and podcast authors alike.
There was a vast improvement in quality from giving it a prompt when generating the overview. The generic un-prompted overview was for entirely the wrong audience, in our case users of our infrastructure rather than the developers. When instructing it to generate an overview for the SRE team and what they should focus on it was far better.
Was it useful for our in-depth analysis, no. Would I listen to one based on the last 100 postmortems for a new team I joined, absolutely. As an overview it was ideal, pulling out common themes from a lot of data and getting some of the vibe right too.
Google of course is the birthplace of attention is all you need.
[1] https://www.wsj.com/tech/ai/noam-shazeer-google-ai-deal-d360...
Improvements over NotebookLM:
(1) You can start with just a subject, and you don't need a full document to begin (though you can do that too![1])
(2) The podcast generates much faster
(3) The podcasts are interactive -- you can ask the hosts to change direction mid podcast, and they will do so.
(4) (Coming soon) You'll be able to make a Spotify-style Queue of Podcast topics, which you can add to as you encounter new ideas.
The primary tradeoff is that the voices / personalities are somewhat less engaging than NotebookLM at this time, though this will be dramatically improved over the coming months.
This is all in addition to the core value proposition, which is roughly "AI Generated Duolingo for Any Subject".
It's early days, but I'd love for you all to check it out and give me feedback :)
[1] Documents are currently heavily length-limited but this will be improved shortly
I'm also playing with other voice models -- built an awesome "voice actor simulator" with OpenAI Realtime Voice -- but it's expensive. Considering asking users to pass in an OpenAI API key for advanced voice? Or maybe just passing the per-token cost along to the user in their subscription.
audience=technical, duration=long, tone=professional & engaging
On a web forum I am admin on, a user opened a DM a week ago titled "Google Notebook LM", someone else had shared a generated podcast thing that summarised the view of the forum on a particular subject, and it called out the usernames of someone who had strong opinions.
In response, another user ran with this and asked for a podcast to be generated summarising everything that was said by the user, their political views and all their hot takes.
Erm... uh-oh.
The use of real identity, the use of the same username across multiple sites, now makes it trivial for things like "take this Github username, find what sites the same username exists on, make a narrative of everything they've ever said, find the best and worst of what they've ever said"... which is terrifying.
I've said to the user the same old line we always repeat, "anything placed on the internet is effectively public forever", but only now are the consequences of this really being seen.
The forums I run allow username changes, encourage anonymity as much as possible, but we're at a point where multiple online identities, one for every site, interest, employer, etc... is probably the best way to go.
I notice on HN that there are many accounts that seem to register just to comment on particular stories and nothing more, and the comments are constructive and well thought out, and now I wonder whether some are just ahead of the curve on this — obscuring the totality of their identity from future employers, or anyone else who might use their words against them.
It feels like our lightweight choices in the past will start to have significant consequences in the present or future, and it's only a failure of imagination that is delaying a change in user behaviour.
> I notice on HN that there are many accounts that seem to register just to comment on particular stories and nothing more, and the comments are constructive and well thought out, and now I wonder whether some are just ahead of the curve on this — obscuring the totality of their identity from future employers, or anyone else who might use their words against them.
Throw aways are very common here for that purpose! On my end I'm becoming more interested in how to safeguard users -- anonymize them -- and also how to make easy to _generate_ throwaways without opening the door to spam (e.g. generate from a valid account, but then detach it). HN likely gets around this by being niche; I think the somewhat unattractive site design helps there.
Yeah the light cone of online activity seems to only grow with little diminishing, which seems unnatural and counter to the type of environment we evolved for. GDPR and the right to be forgotten seemed funny in my youth, now I see it as wisdom ahead of its time.
I'm still getting the tooling right so that the videos will get made in a better and more consistent schedule.
Something like this:
The two podcast hosts have very different levels of knowledge on the topic. The first host is the expert on the topic and explains the subject and the details to the second host. The second host has very little existing knowledge about the subject but will react to the information and ask follow up questions.
https://github.com/souzatharsis/podcastfy
Developer's twitter: @souzatharsis
https://trends.google.com/trends/explore?geo=US&q=NotebookLM...
Google never has problems with first the first few millions for consumer-launched tools. They have problems with the first few millions of net profit almost 100% of the time and shut it down a few years later.
But I do agree this is a good play for Google, it plays to their strengths.
Good point. I couldn't come up with a well-enough known competing tool to compare against.
However like you I cottoned on early that one could put a "Podcast Production Notes.txt" in each one of my Notebooks that allowed me to really put some horsepower behind the generated audio :D
https://github.com/etewiah/awesome-notebooklm
Will be interesting to see what new ideas NotebookLM leads to. I feel this is how custom GPTs should have been launched. OpenAI is on the backfoot here.
Really. "Using"? (as in an email from an org owned domain logged in to notebooklm page?..)
Just posted some ISWC, MobileHCI and UbiComp papers, UIST is up next.
Google has the attention span and product focus of a crack addled flea. I’m afraid the entire project will be killed.
NotebookLM is a great product. I just started using it this week to ingest artifacts for a new project and get an overview.
People building on top of this will likely want to know what the Open Source / non doomed version will be!
It’s a work in progress, and I literally just added the first foundation for note taking next to the RAG chat last night(not working yet) but it’s getting there. The plan is to build the functionality out and then migrate the codebase to an API backend so that people can make their own front-ends for their own needs. Also, it doesn’t have podcast creation or TTS, yet. They are planned, but it can transcribe and ingest a podcast for RAG chat currently.
<Person1> "Welcome to Podcastfy! <break time="0.5s" /> Today, we're summarizing an interesting text about [topic from input text]. Let's dive in!"</Person1>
Avoid using statements such as "Today, we're summarizing a fascinating conversation about ...". ]
[PauseInsertion: Avoid using breaks (<break> tag) but if included they should not go over 0.1 seconds]
... ...
tl;dr, not much.
You are both pelicans who work as data
journalist at a pelican news service.
Discuss this from the perspective of
pelican data journalists, being sure
to inject as many pelican related
anecdotes as possible
You can listen to the 7m40s resulting MP4 here: https://simonwillison.net/2024/Oct/17/notebooklm-pelicans/
Example snippets:
You ever find yourself wading through
mountains of data trying to pluck out
the juicy bits? It's like hunting for
a single shrimp in a whole kelp forest,
am I right?
The future of data journalism is
looking brighter than a school of
silversides reflecting the morning sun.
Until next time, keep those wings
spread, those eyes sharp, and those
minds open. There's a whole ocean
of data out there just waiting to be
explored.Hah, just found this on Reddit: https://www.reddit.com/r/notebooklm/comments/1g64iyi/holy_sh... - many f-bombs.
It’s amazing how it doesn’t occur to OpenAI and others that the “safety guardrails” really dilute the output. And it’s a display of conservatism. Bizarre to think that in America AI couldn’t swear.
Until today, where Google allowed the AI to behave as the lord intended.
It seems to me these alignment questions are a conscious trade-off between giving users what they ask for and brand safety knowing that every spicy output will immediately find its way to twitter/reddit/etc.