Google is using YouTube videos to train Gemini, VEO 3 AI models

Google It uses a vast library of YouTube videos to train artificial intelligence models such as Gemini, VEO 3 video and audio generators. CNBC learned.

According to people not allowed to publicly talk about the issue, tech companies are turning to catalogs of 20 billion YouTube videos to train these new age AI tools. Google confirmed to CNBC that it relies on a YouTube video vault to train AI models, but told it the company only uses a subset of the videos for training and respects specific contracts with creators and media companies.

“We have always used YouTube content to improve our products, and this has not changed with the advent of AI,” a YouTube spokesperson said in a statement. “We also recognize the need for guardrails, which is why we invested in strong protections that enable creators to protect their images and portraits in the age of AI.

The use of YouTube videos like this could lead to an intellectual property crisis for creators and media companies, experts said.

YouTube said it had previously shared this information, but an expert who spoke with CNBC said that creators and media organizations don't understand that Google is using video libraries to train AI models.

YouTube doesn't say about the number of the 2 billion videos on the platform, or the number of videos used for AI training. However, given the size of the platform, training for just 1% of the catalogue amounts to 2.3 billion minutes of content. Experts say this is more than 40 times the training data used in competing AI models.

In a blog post published in September, the company shared that it could be used to improve product experiences such as “including machine learning and AI applications.” Users who upload content to the service have no way of opting out of Google's training with videos.

“It's plausible to have data from a lot of creators who have spent a lot of time and energy on it, and from a lot of creators who are thinking of putting it in these videos,” said Luke Arrigoni, CEO of the company that works to protect creators' digital identities. “The VEO 3 models help these creators create synthetic versions, poor fax machines. That's not necessarily fair to them.”

CNBC spoke with several major creators and IP professionals, but no one was notified by YouTube that YouTube could use content to train Google's AI models.

The revelation that YouTube is training with users' videos is worth noting after Google announced in May its VEO 3, one of the most advanced AI video generators on the market. As Google has been announced, they have shown movie-level video sequences, including scenes of an old man on a boat or another person showing an animal like Pixar talking to each other. The entire scene, both visual and audio, was completely generated.

According to YouTube, an average of 20 million videos are uploaded to the platform by independent creators by almost every major media company. Many creators say they are concerned now. They say they may be helping to unconsciously train the system.

“Telling people what videos they trained and how much they trained doesn't hurt their competitive advantage at all,” Arrigoni said. “The only thing that really impacts is their relationship with their creators.”

Even if the final output of Veo 3 does not replicate existing work directly, the generated content will fuel and burn commercial tools that can compete with creators who have enabled training data without credit, consent or reward, experts said.

When uploading videos to the platform, users agree that YouTube has a wide range of licenses to the content.

“By providing content to the Services, we grant YouTube a non-exclusive, royalty-free, sublimable and transferable license to use that content worldwide.

“More and more creators have discovered fake versions that run through the platform. New tools like VEO3 only accelerate trends,” says Dan Neely, who helps individuals protect against misuse and promote secure licensing of certified content.

Neely's companies are both individuals and businesses challenging AI platforms to generate content that allegedly infringes on clients' intellectual property. Neely said YouTube has the right to use this content, but many content creators who post to the platform are unaware that the video is being used to train the AI ​​software that generates the video.

Vermillio uses a proprietary tool called Trace ID to assess whether the video generated by AI significantly overlaps with human-generated videos. Trace ID assigns scores on a scale of zero to 100. A score above 10 for videos using audio is considered meaningful, Neely said.

In one example cited by Neely, the video of YouTube creator Brodie Moss closely matched the content generated by Veo 3. TraceID was due to the original video with an audio score of over 90 on audio alone.

Some creators told CNBC they are welcoming the opportunity to use VEO 3, even if they are trained with content.

“I'm trying to treat it as a friendly competition. They treat it more than they are enemies,” said Sam Beres, a creator with 10 million subscribers on YouTube. “It's inevitable, so I'm trying to do things actively, but it's kind of exciting and inevitable.”

Google includes compensation provisions for generative AI products, including VEO. This means that if users face more copyright challenges than AI-generated content, Google will be liable and cover the associated costs.

YouTube announced a partnership with a creative artist agency in December, developing access to top talents to identify and manage AI-generated content featuring portraits. YouTube also has a tool that asks the creator to remove the video if they believe that they will abuse their likeness.

However, Arrigoni said the tool is unreliable for clients.

YouTube also allows creators to opt out of third-party training from certain AI companies. Amazon, apple and nvidiahowever, users cannot stop Google from training its own model.

Walt Disney Company And Universal filed a joint lawsuit last Wednesday against AI image generator Midjourney, alleging the first lawsuit of its kind from Hollywood: copyright infringement.

“The people who are losing are artists, creators, teenagers whose lives are covered in their lives,” says R-Mo. Sen. Josh Hawley, of the group, spoke at a Senate hearing in May about the use of AI to recreate human portraits. “We need to either return to individuals' enforceable rights and their property in their property, or this will never stop.”

Disclosure: Universal is part of NBCuniversal, the parent company of CNBC.

clock: Google Buyouts highlights Tech cost savings in AI Capex Boom