Google trains Veo 3 AI video generation model using YouTube content: Report

With billions of YouTube videos at its disposal, Google has reportedly trained AI models like Veo 3 and Gemini using user-generated content without their awareness, approval, or compensation

Google has reportedly been using YouTube content to train its artificial intelligence (AI) models, including Gemini and the Veo 3 video and audio generator. According to a report by CNBC, a YouTube spokesperson confirmed that Google relies on its bank of YouTube videos to train its AI models. However, the spokesperson added that Google does not use each and every single video on YouTube but only uses a subset of its videos for training purposes. 

 

The report further claims that many creators whose videos might have been used in this matter remain unaware that their content has been used without their consent or any compensation.

 

Creators were never notified?

As per YouTube, this information has been conveyed to creators previously, but, as per experts who talked to CNBC, it is not widely understood by creators and media organisations that the US technology giant trains its AI models using its video library (YouTube).

 

Earlier last year, in September, YouTube in a blog stated that the content uploaded on the platform could be used to “improve the product experience … including through machine learning and AI applications.” A huge disadvantage here is that creators who have uploaded videos on YouTube have no way of opting out from letting Google use it to train AI models, which is something that its competitors, like Meta offers. Surprisingly, YouTube allows created to opt out from sharing their content with third-party companies to train their AI models. 

  As per YouTube, there are around 20 billion videos on the platform, and out of them, how many are being used to train Google AI models is unclear at the moment. CNBC cited experts as saying that even if Google uses one per cent of those videos, then it would amount to around 2.3 billion minutes of content, which is 40 times more of the training data that is being used by competing AI models for training. 

 

The report claimed that CNBC talked to a number of leading creators and IP professionals, and it found out that none of them were apparently aware or had been informed by YouTube about the possibility of their content being used to train Google’s AI models.

 

Why does it matter

YouTube, using user-uploaded videos to train AI, has raised concerns, especially after Google unveiled its powerful Veo 3 video generator. The tool can create fully AI-generated cinematic scenes, including visuals and audio. With around 20 million videos uploaded to YouTube daily by creators and media companies, some fear their content is being used to build technology that might one day rival or replace them.

 

CNBC cited experts as saying that even if Veo 3’s results don’t directly copy existing content, the AI-generated output can power commercial products that may rival the very creators whose work helped train it, without their permission, credit, or payment.

 

This no-way-out trap begins as soon as a creator uploads a video on YouTube, as by doing so, the person agrees to YouTube having a broad license to the content.

 

What does the past record show

According to The New York Times, Google has reportedly transcribed YouTube videos to train its AI models. Mashable India points out that this practice raises legal concerns, as it may infringe on creators’ copyrights. 

 

The use of online content for AI training has already led to lawsuits related to licensing and intellectual property. Other players like Meta and OpenAI have also faced heat for using intellectual property to train their AI models without having the consent from creators or authors.