so this is a big deal. Apple, Nvidia, and Anthropic, and many other AI companies have been using content stolen from YouTubers like Mr. Beast, MKBHD, PewDiePie, and more. And it seems these big YouTubers are not happy about it. So I'm going to break it all down for you right now. So the story starts with an investigation published by Proof News, and the title is Apple, Nvidia, and Anthropic use thousands of swiped YouTube videos to train AI. So what does that actually mean? Let me actually take a step back and explain how these models get trained. So AI models require a ton of data, and the data comes in the form of text, usually. But companies like Apple, Nvidia, Anthropic, they don't actually have their own data. The exception to that would be Meta, which has Facebook, and all of the content you generate on Facebook is data that Meta owns and could be potentially used to train their models. And Twitter is the same thing. But for a lot of these huge companies, they don't actually have their own data.

so what do they do? They go out and they buy the data. They buy it from third-party companies. One of these companies is called Uther Ai, and they're essentially just data scraping companies. And Uther AI is the company behind a data set called the pile, an incredibly popular data set used by tons of different AI companies to train their models. So here's a little bit of information about the pile from Wikipedia: it is 886 GB of diverse open-source data set of English text created as a training data set for large language models. So it is open source, so anybody could grab the pile and use it to train their own models. But this data set is only supposed to have data that they were authorized to get, and that's where this story gets interesting. Now what Proof News was able to find is within that data set, the pile, they were able to find subtitles from tens of thousands of YouTube videos. So anytime that you're watching a YouTube video, you can hit the subtitle button and you're going to get

subtitles written on the screen. That is every word that is spoken in the YouTube video is essentially translated to text, and anybody can easily download this. So on any YouTube video, if you simply go down to the description and then right here click show transcript, you get the entire transcript. And of course, you can simply highlight and copy that and paste it anywhere you want. So as proof news started digging into the pile, they found transcripts from the most popular YouTubers on YouTube. And not only are the YouTubers not happy about that because they were never asked their permission to give over that data, but I bet Google is quite upset about it as well. And here's why. Google owns YouTube. Google is developing their own artificial intelligence, and many AI thought leaders have predicted that YouTube's content might be the most valuable data source on Earth. Not only do they have so much video and audio and transcriptions are uploaded every single minute, and all of that can be used to train future models. Now Google owns that data, but Apple, Nvidia, and every other AI

company that has been using the pile has been using that data without their permission. So here in the article, it says companies use the training data despite YouTube's rules against harvesting materials from the platform without permission. Our investigation found that subtitles from 173,000 YouTube videos siphoned from more than 48,000 channels were used by Silicon Valley heavyweights including anthropic Nvidia, apple, and Salesforce. The data set called YouTube subtitles so it's not like they even tried to hide it, they just scraped it and took it contains video transcripts from educational and online learning channels like KH Academy, MIT, and Harvard. The Wall Street Journal, NPR, and BBC also had their videos used to train AI, as did The Late Show with Steven colar last week with John Oliver and Jimmy Kimmel Live. They also found material from YouTube meast Stars including Mr. Beast, Marquez brownley also known as MKBHD, Jack sepai, and PewDiePie. They also have this Nifty tool where you can literally search for a Channel or a

video, and it's going to tell you exactly within the piles data set if that video or Creator is found. So if I simply type in Beast, we see Mr. Beast video right there and a bunch of other videos as well. And Marquez Brownley, MKBHD, has already commented on it on Twitter. Let's take a look at what he said. So he's retweeting the 9 to 5 Mac article about this finding. Apple has sourced data for their AI from several companies. One of them scraped tons of data transcripts from YouTube videos, including mine. Apple technically avoids fault here because they're not the ones scraping, but this is going to be an evolving problem for a long time. And completely agree with data sets this large. It's going to be extremely difficult to know whether every single point of data was gathered with permission from the author. And in fact, that seems like it might be an entire business in itself is to scan these data sets and find out if there is any copyrighted or permissionless data. And what's interesting is he points out

the legal side of it. Apple technically avoids fault because they're not the ones scraping. So although Apple and these other companies are benefiting from this data set, they're not actually at fault because they didn't do the scraping themselves. They simply bought it from a third-party company who did that scraping. And that is another AI. So you might be asking yourself, why is this a big deal? Well, first of all, as a content creator, of course, I only want my content used in the ways that I approve. MKBHD, Mr. Beast, they put a ton of time, effort, and money into each and every one of their videos. So they want to be able to say how that content is used. And beyond that, MKBHD actually says, Fun fact, I pay a service by the minute for more accurate transcriptions of my own videos, which I then upload to YouTube's back end. Some companies that scrape transcripts are stealing paid work in more than one way. So not only is he putting in the time and effort to create.

the content. He pays another company to write the transcripts, so he's uploading those transcripts, and then they're being stolen. So, yeah, he's justifiably upset. And this isn't the first time that we've found stolen data from within AI training data sets. That was the entire basis of the New York Times suing OpenAI. They were able to replicate large portions of some of their articles by simply prompting ChatGPT. Now, I believe that lawsuit is still ongoing. And of course, if there are any updates, I will let you know. And beyond that, OpenAI has been on a data acquisition spree lately. They partnered with Reddit to get their data. They've partnered with other news publications to get their data. Here is a blog post from June 27th: Strategic Content Partnership with Time on the OpenAI blog. So, they are working with Time Magazine. Here's another partnership announcement with The Atlantic, Vox Media, News Corp. So, again, they are making sure that they're getting permission from everybody they possibly can. And this is an arms race. This is an arms race to get

the most data to train future models, and OpenAI seems to be making all the right deals. But if you followed my Channel at all, you know that OpenAI has been developing a product, a text-to-video generation product called Sora. And there's speculation that one of the reasons that they've not actually released it to the public yet is because it's actually trained on stolen data. Thank you to On Demand for sponsoring this portion of the video. On Demand is the FastTrack to deploying real-world AI applications. Let me tell you about it. First, they have Bring Your Own Model. So with On Demand, you can easily deploy your own custom models directly from Hugging Face, like Llama 3. This allows you to integrate the models that you know and love easily into your applications. They also offer Bring Your Own Inference. So if you have models that are hosted on an external service like Amazon SageMaker, you can easily integrate those into your applications as well, allowing you to use your preferred inference service. They also have a plug-in marketplace where you can

explore a vast array of AI plugins to either enhance your projects or to share your own with the community. You can create, distribute, and monetize your own plugins through on-demand, fostering collaboration and innovation. Next, I want to talk about Playground. This is your testing ground where you can configure, debug, and see live outputs from your plugins and models. It is an invaluable tool for fine-tuning your AI applications before deploying them to production. And can export your generative AI applications into any programming language. And finally, automations combine AI agents and plugins to manage complex workflows effortlessly, which means you can automate repetitive tasks so you can focus on what matters most, innovating and developing AI applications. So check out On-Demand. They have a suite of powerful features for AI development. If you're a developer, a researcher, an AI enthusiast like myself, On-Demand has something for you. Thanks again to On-Demand. Now back to the video. Video: Mid-Journey is another example of an AI company accused of stealing data. Here's a perfect example of it. This is from

Reed Southern, he says, I overlaid the AI Joker image from mid Journey V6 with the film frame. I think this is pretty damning, and as you can see, it's almost identical. It's clear that they stole the frame of that movie to train their data. And here's another few examples. These are frames from a Marvel movie, and as you can see on the left side, this is the actual frame from the film, and on the right side, this is something that has been generated with mid Journey. So yes, you can see it is basically identical. But taking a step back, what does this mean within the context of fair use? If we think back to the letter that OpenAI published as a counter to the New York Times lawsuit, they basically said, What's the difference between a human reading a bunch of text articles that are free and publicly available on the internet as YouTube subtitles are learning from them and then being able to answer questions about that topic? That's essentially how large language models work. They are reading the text, and they are learning.

to answer the questions now. I'm torn on this. I definitely think there's some fair use argument here, but I'm a content creator also, and I would also be quite upset if I knew AI models were being trained on my videos without my permission. They are basically taking creators' hard work and extracting value from customers based on that hard work without actually creating any new ideas on top of them. And there's that famous meme from CTO of Open AI, Ilya Sutskever, where she gives this awkward-looking face after being asked if Sora is trained on YouTube videos. Here in the article, they reference that Open AI executives have repeatedly declined to publicly answer questions about whether it used YouTube videos to train its AI product Sora, which creates videos from text prompts. Earlier this year, a reporter with The Wall Street Journal put the question to Ilya Sutskever, Open AI's Chief Technology Officer. I'm not actually sure about that. Now, here's another angle in which to think about it. So right here it says YouTube subtitles, which is the data set, not the actual

subtitles, which was published in 2020, also contains subtitles from more than 12,000 videos that have since been deleted from YouTube. In at least one case, the Creator deleted their entire online presence, yet that work has been incorporated into an unknown number of AI models. So, if you're a Creator and you no longer want your work out there, you simply delete your channel. But what do you do with all that data that has been used to train AI models? And this is probably pretty similar to problems that Google went through in the early days of search. Essentially, they are scraping content, they're indexing that content, and then they give you little previews of the content. And over time, Google's previews have gotten bigger and better and thus given consumers less reason to actually click through that blue link to the source website and instead just get the content straight from Google. Now, AI models take that to a whole another level and taking it to legal action, there have been a number of lawsuits from musicians, from authors, from other artists in general. Here it says books.

three another pile data set. And last year published a piece in the Atlantic reporting his finding that more than 180,000 books, including those written by Margaret Atwood, Michael Pollan, Zadie Smith, had been lifted. And they have since sued AI companies. Now in response, defendants such as Meta, OpenAI, and Bloomberg have argued that their actions constitute fair use. So this likely is going to make it up to the highest levels of the US courts. So all of this is super interesting to think about, and I'm going to be following it closely. I will update you as it progresses. If you enjoyed this video, please consider giving a like and subscribe. And I'll see you in the next one.

Apple Stealing Data From YouTubers?!?

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **