A square filled with tourists , a waiter deftly navigating between tables, a bike whizzing by, or even a journalist meticulously arranging a set—all of these scenes can now be conjured up effortlessly by generative AI in mere seconds. The results are not only astonishing but also provoke a critical question that few have contemplated until recently: How have these AIs learned to so accurately mimic the world ? According to an insightful report from The Atlantic, a significant part of the answer lies in millions of videos harvested from platforms like YouTube without explicit consent.
The rapid ascent of generative AI has unfolded so swiftly that many intriguing questions remain unanswered. In just a couple of years, we have transitioned from rudimentary experiments to sophisticated models capable of generating videos that are often indistinguishable from actual recordings. While the spotlight has generally been on these capabilities, the matter of transparency has been escalating in significance. For instance, OpenAI has indicated that their Sora model is trained using “publicly available” data, yet they haven’t clarified what exactly that data entails.
A Massive Workout that Points to YouTube
The investigation in The Atlantic offers a revealing glimpse into the background of this phenomenon. It unveils the staggering reality that over 15 million videos have been amassed to train AI models, with a significant portion sourced from YouTube without proper authorization. Various initiatives have been cited that employed data sets created by several companies, designed to enhance the capability of video-generating AI . Alarmingly, this process occurred without the knowledge of the original creators of the content.
What’s even more jarring is the nature of the material being exploited. These aren’t merely random videos or amateur clips; rather, they include informative content and high-quality productions. Investigators found that thousands of clips originated from reputable channels belonging to esteemed publications such as The New York Times , BBC , The Guardian , The Washington Post , and Al Jazeera . This reveals a substantial volume of journalistic content being siphoned off to fuel AI systems without any prior agreements with those who own the rights.
Runway , one of the pioneers in generative video, has emerged as a notable example in this review of data sets. According to the documents referenced, their models have been trained using clips categorized by type and context: interviews, explanatory content, graphical pieces, cooking tutorials , and more. The rationale behind this is straightforward: for AI to replicate real-life situations and audiovisual narratives, it requires actual references, encompassing everything from gestures to editing styles .
<span>Fragments of a video generated with the Runway tool</span>In addition to Runway, the research highlights the data sets utilized in the laboratories of major tech companies like Meta and ByteDance , confirming that a similar dynamic occurred: vast quantities of online videos were gathered and shared among research teams to enhance audiovisual capabilities.
YouTube’s official viewpoint is starkly clear. Its regulations explicitly prohibit downloading videos for the purpose of training AI models, a stance underscored repeatedly by its CEO, Neal Mohan . He emphasizes that creators rightfully expect their content to be used strictly within the guidelines of the service. The influx of millions of videos into AI databases has brought this legal framework into sharp focus, amplifying pressure on platforms engaged in the development of generative models.
The response from the media landscape has been a two-pronged approach. On one hand, companies like Vox Media and Prisa have entered into agreements to license their content to artificial intelligence platforms, seeking a defined framework and financial remuneration. Conversely, some media organizations have aggressively defended their rights: The New York Times has initiated legal action against OpenAI and Microsoft for the unauthorized use of their materials, stressing that this includes video content as well.
The legal landscape is murky. Existing legislation was not designed to manage models processing millions of videos concurrently, and courts are just beginning to define the parameters. Some experts assert that publishing content publicly does not equate to relinquishing training rights, while AI companies contend that indexing and utilizing public material is integral to technological progress. This unresolved tension continues to create a complex balancing act between media outlets and developers.
<img alt="If you thought the AI bubble was worrying, it's because we hadn't entered its next phase: debt" width="375" height="142" src="https://i.blogs.es/8782aa/ia1200x900/375_142.jpeg"/>As this discourse evolves, it transcends mere technological implications, delving into the realms of ethics and legality. The practice of training AI models with internet-available material is not new, yet the pressing question remains: where do we draw the line? Companies promise transparency and agreements, while media outlets seek assurances, and creators demand control. The forthcoming phase will be both technological and political, as the manner in which artificial intelligence is nurtured will ultimately dictate who stands to gain.
Images | Xataka with Gemini 2.5
In Xataka | All the big AIs have ignored copyright laws. The amazing thing is that there are still no consequences