Good morning, John. In the last few weeks, there's been a bunch of fantastic investigative reporting uncovering the reality that generative AI companies are pulling videos and transcripts from YouTube and using those creations owned by independent creators to train their AI models. 404 Media found that generative AI video unicorn Runway trained on thousands of videos without permission, including thousands of Sow and Crash Course videos. A while back, the CTO of OpenAI, when asked if YouTube videos were in their training data, made this face, which isn't really what you want to see. A Proof News investigation found the transcripts of 17,353 YouTube videos were provided to Apple, Salesforce, Nvidia, and Anthropic to train their models. Proof provided a little database that you can search, and indeed, Sow, Crash Course, Eons, Vlog Brothers, all in the training data. If you have a YouTuber that you like, they're probably in there. And the reaction to that from creators has not been great. It seems like a lot of creators do not want

their property used to train generative AI models. In fact, a substantial majority of creators I've talked to, whether artists, photographers, musicians, or YouTubers, say that they do not want their work used to train AI, or they would at least like the option to opt out. They all have different reasons for that, but that's how they feel. So, I'm working on this video, and it occurs to me I didn't actually say exactly what I'm talking about, which is generative AI. So, there's lots of different kinds of AI. The AI that's creating the transcript of this video right now, the AI that recommended this thumbnail to you that you then clicked on, and like, we're using this word to mean everything right now. So, I'm talking about a specific thing here, which is when models train on content and then they create more content. So, MidJourney trained on pictures, and it creates new pictures. Sora trains on videos, creates new videos. LLMs like ChatGPT or Gemini or Claude, they train on text, and then

they create text. Basically, they ingest information and they spit out new information or the same existing information but in a different way. And so sometimes in this video I say generative AI, sometimes I just say AI. I mean generative AI the whole time. Sometimes you'll see it written as gen AI, which is confusing because there's this other thing called artificial general intelligence, AGI, which is a vastly different thing that doesn't exist. That is not what this video is about. This video is about generative AI, AI models that ingest information and then spit out new things. Just wanted to be very specific about that. And now we're going back. The company that I founded from what I can tell actually has more videos than almost anyone else in this database of videos that was used by these massive companies to train their generative AI. Marz Brownlee has nine, Mr. Beast has 19, Mark Rober has 24, Minute Physics 85. Meanwhile, Vlog Brothers has 188, SAOW has over 300, and Crash Course has 982. Seems like whoever made this

database is kind of a fan of ours, in which case I'd like to say, uh, rain it in. The only channel I've seen that has more videos in this data set than us is KH Academy. Ted Ed has over 400, MIT has hundreds as well, making it fairly clear that educational content, in particular, is pretty valuable for training large language models. Another reason why I think it might make sense for some people to make different decisions about what they'd like to do with their content, just saying. Now, it's probably important to explain why people are upset about this. Why do creators feel like they're being ripped off? And there's basically two ways to imagine this, although there's, of course, a spectrum between them. The first perspective is that these computer programs are just learning the way that any human would, and you can't be mad when someone learns something from a YouTube video. The second perspective is that computer programs aren't people, and they don't learn the way that people do, and thus this is an entirely new way for

copyrighted content to be used, and these models, which are now billion-dollar products, would be nothing without the copyrighted content they were trained on, and that thus somehow are inside of them. And look, as a guy whose company owns a lot of the data inside of these data sets, if you're a lawyer and you think we have a case, Kelsey would love for you to email her at Kelsey@complex.com. Now, from where I sit, I'm definitely getting ripped off. Like, I know I'm getting ripped off because a bunch of big companies have signed licensing deals with AI companies so they can train on their data, so they're getting paid for their data to be in the model, and I'm not getting paid for my data to be in the model, and that seems like a bunch of bulls to me. Why not pay me just 'cause you didn't think you'd get caught? Yeah, I mean, I get that. I've also done stuff that I knew was wrong because I didn't think I was going to get caught. Like when I

worked at Walmart. I used to fill up my water bottle with Sprite instead of water before bringing the carts. Set this is kind of like that, except instead of a Sprite, it's, you know, the property of 50,000 individual independent creators who all have fairly big audiences. And those audiences will absolutely go to bat for them when asked to. But let's cross that bridge when we come to it. I think there's maybe a more important thing to hit first. I think there's a question that has been bubbling around in the background of all of this that a lot of people are maybe a little bit afraid to ask. If it's bad for these companies who we do not have much of a relationship with to yoink our videos and then train generative AI with them, how would we feel if a company that we did have a relationship with was doing that? In short, the question is, does Google train its generative AI models on videos owned by YouTubers? To get to the bottom of this question, I would like to

show you two clips from a Bloomberg interview with Neil Mohan, the CEO of YouTube. Do you believe that YouTube was used to train Sora? If it was being used, would that be against your policy? It would be when a Creator uploads their hard work to our platform. They have certain expectations. One of those expectations is that the terms of service is going to be abided by. Our terms of service does allow for YouTube content, some YouTube content, like the title of a video or the channel name or the Creator's name, to be scraped because that's how you enable the open web. But it does not allow for things like transcripts or video bits to be downloaded, and that is a clear violation of our toss. And so those are the rules of the road in terms of content on our platform. If you're anything like me, two weeks ago, what you think Neil just said was creators expect that their content will not be used by companies to train generative AI. But if you look at the details, that's not what he said. He said

that creators expect YouTube's terms of service to not be violated, and YouTube's terms of service says nothing about AI training. And then later, the reporter from Bloomberg asks him straight up whether Google uses YouTube videos to help train Gemini, and here is what he says: And how does it work internally? Like, is Google using YouTube to train Gemini? Google uses, um, uh, YouTube, um, content really in accordance, again, back with those terms of service or individual contracts that we might have with creators or uploaders to our platform. Lots of creators have different sort of licensing contracts in terms of their content on our platform. Lots of rights holders do. Now, it's true that lots of different rights holders have lots of different deals with YouTube, but the people that you and I think of as YouTubers, rather than record labels or legacy media companies, they all have the exact same deal. Like, I don't have a special deal with YouTube. Marquez Brownlee doesn't have a special deal with YouTube. We're YouTubers. We all have the same deal, and that is simply

the YouTube terms of service. And when Neil says this, I kind of get the impression that YouTube has struck a deal with some creators to allow their content to be used to train AI, like they've clicked on a button that says it's okay to use my content. But there is no such button. However, there is a license. When you upload a video to YouTube, you agree to the YouTube terms of service, and included in that is a license. And here is what that license has to say about how YouTube can use your content: By providing content to the service, you grant to YouTube a worldwide, non-exclusive, royalty-free, sublicensable, and transferable license to use that content, including to reproduce, distribute, prepare derivative works, display, and perform it. Now, none of these words have the specific meaning of train AI with it, but as we will discover in this video, Google definitely sees this license as sufficient to include training generative AI. It's not as solid a legal foundation as if it actually contained language explicitly about training, but it appears to be solid enough for them. Is it actually legal to

train AI with content when you have a license that allows for derivative works. I don't know; no one knows. There are quite a number of lawsuits about that and related issues going on right now. These companies have been moving extremely fast and doing things that are very risky, and in many cases, have definitely got them in trouble. The New York Times lawsuit with OpenAI is amazing. Let me read to you the quote of what the lawsuit says. It says, The defendants, OpenAI, should be held responsible for billions of dollars in statutory and actual damages related to the quote unlawful copying and the use of the Times' uniquely valuable works. And it also calls for the companies to destroy any chatbot models and training data that use copyrighted material from the Times. So that's one of the lawsuits going on right now. At one point, you could type into OpenAI and say, like, Finish this article, and it would just show you the text of an article from behind the paywall. The text was just in there, and it was like, Here it is, but as

These things are working their way through the courts, and we don't know exactly what is and isn't illegal. There's one thing that I want anybody who works at YouTube or who is a YouTuber to take away from this video. When the license I have with YouTube was written, the use case of training a large language model did not exist. This is a new use. It's not like anything else, and to say that the existing license that I have with YouTube covers this brand-new use that no one could have anticipated when I signed on to it does not feel great. I have personally uploaded more than 3,000 videos to YouTube. My company has over 9,000 videos on YouTube. We did not upload them thinking that this was part of the license we had with YouTube. Right now, the AI training data situation is an absolute mess, and it's only getting messier, and I am going to talk about it. But first, let's just connect the dots here. Neil says that lots of creators have different licensing contracts, and that any

training that may be happening, though he doesn't say it definitely is, but if it were happening, it would be in concert with the terms of service, and that terms of service explicitly allows, as we have seen, for Google to prepare derivative works of content owned by YouTubers. So, I don't know why he doesn't say it, but it seems pretty clear that when he says some portion of the YouTube Corpus may be being used to train Gemini, the portion he means is the part owned by independent creators who are operating under the standard YouTube terms of service. I think it'd be a little weird for me to speculate about why exactly YouTube is being so obtuse when it comes to this stuff, so instead, I just asked Gemini. They might be concerned about using copyrighted content like YouTube videos because that could raise legal and ethical issues, and also admitting to using user-generated content without explicit permission could lead to negative publicity. But of course, Gemini May display incorrect info now for additional context around all of this, specifically the part where YouTube

believes it has the right to train generative AI given the derivative Works Clause of the terms of service. I have one short thing to say, one medium thing to say, and one very long thing to say. The short thing is that you might be thinking, Hank, if you have all these questions, why didn't you just ask people at YouTube to answer your questions? And here's what I'll say to that: I reached out to some of my contacts at YouTube to ask whether Google trains Gemini with the property of YouTubers, and they told me that YouTube had no further comment beyond Neil's interview with Bloomberg. The medium-like thing is that Google has already spent quite a lot of time training AI on YouTube videos in ways that I am actually in favor of. For example, training the models that generate subtitles and translations for videos. Now, this is the kind of use that makes sense to me under conditions of derivative use. It uses my videos to train systems, and it also uses my videos to create translations of my videos, and this provides me with value and

the rest of the world with value, and it's the sort of thing that if I were asked whether or not I wanted to opt out of it, I would not opt out. I think it's good, but we can say definitively that the creation of those translation and transcription AI models didn't require anything to be added to YouTube's terms of service to specifically allow that. So, I think it's fair to say that the existing, according to YouTube's lawyers, allows them to train AI models. So, when Neil says that there are some creators who have licensing contracts that allow YouTube to train AI, I think the creators in question are the folks whose license is specified by the standard YouTube terms of service, AKA YouTubers. Now, the long thing I think is the best case, but it is long, and it requires me to explain to you the entire mess of the current state of data collection for AI training data. But I think it makes a pretty good case that Google is using YouTube videos to train AI. So, we're going to go through it. One thing you

probably already know is that the kinds of AI that generate cool stuff, whether that's a recipe for cornbread or an image of a cornbread or a movie of a cornbread. They only work because they are trained on an enormous amount of data. They learn, again in quotation marks, what images and videos and language are like by running just massive amounts of images and videos and language through them in order to get better, which they are racing extremely fast to try and do. They need more data. The thing OpenAI did to create ChatGPT, which kicked off this whole explosion, was that they trained on everything, like everything. When they were trying to figure out what it was okay to train their model on, they said everything. Here is the CEO of Microsoft AI talking about what is acceptable to use as training data. I think that with respect to content that is already on the open web, the social contract of that content since the '90s has been that it is fair use. Anyone can copy it, recreate with it, reproduce with it. That has been free air, if you like.

this is that's wrong. Not not like ethically wrong. It is ethically wrong, but it's also just incorrect. It is legally incorrect. Any lawyer would know that that's incorrect. I should explain that the open web is a specific thing. It's different from things that are on the internet, like obviously your emails are not on the open web, and neither is stuff that's like behind a paywall or stuff that explicitly has systems that tell search engine crawlers and other bots that they should not be indexed. They're basically those systems are saying we are not part of the open web. But regardless, it is a deep, almost willfully intentional misunderstanding of copyright law that publishing something on the open web makes it free for anyone to do whatever they want with. Like Reddit is part of the open web or was. We'll get there, but I cannot just like grab stories from r/n sleep and publish a spooky storybook. Those stories are the property of not Reddit, but of the people who wrote them. And it doesn't sound to me like he misspoke.

here it sounds like he was trying to make a specific argument. It's just an insane one. The argument being, if search engines can reach your property, then it becomes everyone's property. And there are miles of legal precedent saying that that just is not the case. Also, just FYI, he said all of this a month ago. This is all recent. But then, to continue building context, he continues, There's a separate category where a website or a publisher or a news organization had explicitly said, 'Do not scrape or crawl me for any other reason than indexing me so that other people can find that content.' That's a gray area, and I think that's going to work its way through the courts. And I would like to point out that OpenAI, which is in part owned by Microsoft but also kind of considered a competitor by Microsoft, built ChatGPT not just by gobbling up stuff from the open web. They did it by gobbling up, again, everything, including stuff from this gray area, books and YouTube transcripts and stuff behind paywalls. They would not have

been able to make this powerful a model without using data from everywhere, which I would kind of be okay with if it were for research, but then they went ahead and just launched a paid product and started taking investments and making billions of dollars. But now OpenAI is saying that stuff that isn't on the open web but is on the web is a legal gray area, and now they are doing deals with companies that have data like that to clean up their data set, because this is actually something of a potential crisis for this new trillion-dollar AI industry. There's only so much training data in the world, and it seems like more is very much better, and already companies like Google and OpenAI have gathered basically all of it, and now they're actually losing some of it to questions about whether it's actually okay for them to train on that data. Now, there is some hope that the models could maybe create data that they could then train themselves on, and you know, I'm not an expert, but that sounds silly. They need

data to make these extraordinarily powerful, mind-blowing inventions. So, there is a battle brewing about what content is and isn't okay to use in these models. And to explain this to you, and I know this is already taking a long time, I need to tell you about robots.txt. There are computer programs that spend all of their time crawling through the internet, reading web pages, and storing data they find there. This was the genius innovation that basically made the web possible. It made search engines possible, computer programs following links around and grabbing information from the websites they discover. That might be the title of a page or the whole text of the page or just some hidden metadata. And pretty much every website has a public file that you can find by just typing SL robots.txt after the.com. Here is YouTube's. I'll give you a little tour of it. It starts off with a Flight of the Conchords reference. It is the distant future, the year 2000. Because you could put whatever you want in a robots.txt file, and someone at YouTube is still

allowing fun, but then after that, there's some instructions for robots down here. You have a little asterisk, and that means everyone, and there's a list of pages that the robots are not allowed to go to. Now for clarity, the robots could go to any of those places, but if they do, they will be violating YouTube's rules, and possibly, depending on what they do with the data they gather, they could be sued. But any robot is allowed and even encouraged to go to tons of places on YouTube: video pages, channel pages, category pages, all of the places YouTube wants folks to scrape data for, for search engines and other internet products, because the open web is good for everyone, and without people allowing this data collection, search engines wouldn't exist, and the internet would suck. There's a separate group, Google Media Partners, and their robots can go anywhere on YouTube. So if you are defined by YouTube as a media partner, you have free reign. And I only mentioned that because robots.txt can differentiate between what different robots are allowed to do. So if you are

In a legal agreement with YouTube as a media partner, you have free reign. All Media Partners disallow nothing, basically allow everything. robots.txt was the original way websites told other companies what parts of their site it was okay to scrape. It has since then gotten significantly more complicated because sometimes you want some stuff in an area of a website to be able to be scraped, like titles and thumbnails, but not other things like transcripts and videos. And that is what Neil was talking about in his clip, so let's listen to that again. Our terms of service does allow for YouTube content, some YouTube content like the title of a video or the channel name or the Creator's name, to be scraped because that's how you enable the open web, but it does not allow for things like transcripts or video bits to be downloaded, and that is a clear violation of our toss. What's worrying him and what is going wrong is that companies are using these robots to scrape data from YouTube that YouTube says they are not allowed to scrape.

they're just computer programs. They're not actual robots. That's against YouTube's terms of service and possibly against the law. And you think maybe this is the time when we're going to bring it back around to YouTube 'cause I'm talking about YouTube again, but no, no. I have to put a chill into your spine by showing you something wild. Here was Reddit's robots.txt file a month ago. It's got a bunch of fun little cute nerdy references buried in here because someone at Reddit is still allowing fun, but also they disallow some specific web crawlers, and then they disallow web crawlers from scraping some specific parts of Reddit. But almost all of Reddit is open to any search engine that wants to index Reddit. And why wouldn't they do that? Search engines are a huge part of Reddit's traffic. Well, here is Reddit's robots.txt today, skipping over all the stuff at the top. User agent: everyone. Disallow: all. No robots allowed. What the hell is going on? Open AI all of a sudden after launching Chat GPT ended up being

worth like $86 billion, and anthropic Google Apple meta Nvidia all started desperately wanting to build their version of a powerful LLM, doing the same thing OpenAI did by using all of this training data they could pull from all over the Internet. And if you're watching the Olympics, you're seeing Google's ads. There is something here; no one's quite sure yet what it is, but it is something they are touting it like crazy. Now, it all might come tumbling down at any moment because of copyright issues or just because there's not much there actually there, but there is something that they think is there. And like the CEO of Microsoft AI said, if it's part of the open web, meaning it's okay for a web scraper to scrape it, then that is fair game to put in the training data. Reddit saw the reading on the wall and was like, Yeah, absolutely not. Like if the legal relationship is going to be if it's okay to scrape you for search, it'll be okay to scrape you for training data, and

summarization then it's not okay to scrape us for anything user agent. Everyone disallow everything, but they will happily do deals, and they have done deals both with OpenAI and with Google, saying that if you pay me, you can train on our data, and we'll even send that data straight to you, so you don't even need to use web scrapers, and that way we can keep our robots.txt file a virtual wall of note. But here is where we come back to YouTubers. Did you forget we were talking about YouTube? Yeah, this has an implication on the YouTube situation because the information that Reddit is selling is not Reddit's. Reddit is selling access to the door, and once you're inside, you can have whatever's in there, but none of that stuff was created by Reddit or owned by Reddit. It was created by Reddit's users, and those users own those creations, but Reddit sold the ability to train AI on those creations. What? How? How is that possible? Now again, no one has any idea whether it's okay to

train AI with creations from people who did not want to be in the training data, but Reddit does have a license with the people who create content and post it on Reddit, just like YouTube does. And here is what that license says: When your content is created with or submitted to the services, you grant us a worldwide royalty-free, perpetual, irrevocable, non-exclusive, transferable, and sub-licensable license to use, copy, modify, adapt, prepare derivative works of, distribute, store, perform, and display your content. This sounds a lot like the license YouTube has with YouTubers, specifically that derivative works phrase. And the license that Reddit has is sublicensable and transferable, meaning they can sell the rights they have to redditors' work to other licensors, like they just did with Google. So, if the derivative works clause is good enough for Google's lawyers to say they can train on copyrighted content from redditors, it's good enough to train on copyrighted content from YouTubers. I told you it was long, but that's my case.

you see how it all connects together. That's the case that I have built. They are definitely doing it with some YouTube content. They say it's the YouTube content. They have a license agreement that allows for that use. And they say that this particular license agreement that they have with every YouTuber allows for this youth. They believe it's fine to do with Redditors' copyrighted content. I think they think it's fine to do with YouTubers' copyrighted content. And I would like to say I think that they are wrong. I do not know whether it's legal. That's going to take a long time to figure out. I just don't think it's right. I've been uploading videos to YouTube for over 18 years. I love it. There are so many good people who work there. And everybody ever met seems to understand that they are an engine of a creative revolution. And they take that responsibility super seriously. Like far more seriously than any other major player in the space. I like Neil. I think he's a fantastic

leader. I'm glad to have him at the helm, and I want this to be clear: I am not and cannot be coming at this story as a journalist. I'm a YouTuber who loves being a YouTuber and who has a very long and productive relationship with YouTube. I have a huge bias here; like, I'm biased in favor of YouTubers, but also I'm biased in favor of YouTube. I have a lot of trust for a lot of people on that team, and I do not want to sour my relationship with them. It could cause serious economic harm to me. Probably it won't; like, honestly, probably it won't; like, it seems like YouTube is good at siloing these things off from each other, but I'm worried about it, right? Like, of course, even if it's irrational, I'm worried about it. So I can't be unbiased in this situation. I'm not here to try and blow things up. I'm not here to try and spin a story and get a bunch of views. I would not be doing this if I did.

not think it was important. No one knows whether training for llms is legally allowed by derivative works, but I do know that when I implicitly agreed to that license by uploading videos, I did not think I was agreeing to have my content used to train llms. I could not have. So take out legality and just understand I did not know I was agreeing to that. And having talked to them, I know lots of YouTubers do not want to have their content used to train llms. And I don't know whether they're right or wrong. I don't know what I would choose to do. There are certainly some training situations I would be super in favor of and some that would absolutely infuriate me. I could just trust that YouTube won't use their training data in a way that would be bad for me. And maybe they wouldn't, you know? But like, we have different incentives. You got to understand that. So like, it's not that I don't trust you as people, it's that I don't trust that we are always going to

have the same incentives. X recently allowed Twitter users to opt out of having their data used to train their LLMs. The box is checked automatically, but you can uncheck it. And honestly, that's all I'm asking for here. You can absolutely opt me in, but give me a chance to opt out. Because I cannot opt out of using YouTube. I've built my entire company here. We have 70 employees. My community is here. My heart is here. I live here. If you think you don't like it, go somewhere else. That's not the situation I'm in. I don't know. I could spend a lot of time explaining that to you, but please trust me, that's not the situation I'm in. I don't know how every YouTuber feels. Absolutely, I'm not trying to represent everyone here. I'm still trying to figure out how I feel. This is a weird situation, but every YouTuber should be given the chance to opt out. We did not agree to a license that allows for using our property in

this way, but YouTube is, as far as I can tell, acting as if we did in 2026. The EU's AI law is going to require you to disclose your training data. So, in a mere 2 years, we're going to know. Would you rather us find out then, or would you rather give us a chance to clean up your data set so that you don't have to have any? This legal confusion around it, and also give us the opportunity to agree or not to a use that is new and that we did not know we were agreeing to. If you are a YouTube Creator and you think we should be able to opt out of having our content used to train generative AI, leave a comment saying that you would like creators to be able to opt out. Also, there's a link in the description to a survey that you can fill out, either anonymously or not, but I am going to be sharing that data with YouTube to help them understand how YouTubers feel about this. So, know that when you're deciding.

whether or not you'd like to be anonymous. All right, it is 3 weeks after I recorded this video. It's been a whole project to make this shout-out to Milo for helping me edit it and doing a bunch of the graphics. John, I haven't actually talked to you much about this video, which feels weird. I mean, we don't usually talk to each other before we publish a video, but this one's very long and messy. In those three weeks, a bunch of stuff has happened that I want to put into the video, but it doesn't really change the situation that we are in. But I did realize one thing after filming that I wanted to put in people's heads as a final point. If YouTube were not owned by Google, YouTube would obviously not be allowing Google to train on videos from YouTubers. In that world, Google or OpenAI or Anthropic or whoever would certainly be trying to pay for access to that tremendous trove of data 'cause it's super valuable. If one of those companies just

took it as many of them did. YouTube would move to try and stop them like YouTube is doing right now with all of these companies except for Google. And if YouTube got paid for that data, you'd better believe YouTubers would expect them to share that revenue. But in the world that we are in, the one where Google owns YouTube, Google doesn't have to pay apparently. And no one else can pay because OpenAI and Apple and Anthropic and all these other companies are Google's competitors. And so Google, of course, would not sell them access to YouTube's data at any price. And I just wanted to say all that, and you can take what you will from it. And also, John, because this video touches on a lot of things that I feel like are important and like deeper than I usually have time to spend thinking about, I just want to say how proud and amazed I am of what you and I and all of the people at Complexly and also all of my other YouTuber friends have done on this platform. And I am really amazed by

what YouTube has enabled and how much more sustainable it feels than other similar platforms. I think the biggest reason for that is because the people who create the content here are not treated like people who use a website. We are treated like business partners. That almost always feels like the situation to me, but in this particular situation, that is not how it feels. Again, if you're a YouTuber, there's a survey in the description. John, I'll see you on Tuesday.

Is Google Training AI on YouTube Videos?

Is Google Training AI on YouTube Videos?

스크립트