Was the Fine Print Training for AI?
Nearly twenty years ago, I wrote a commentary for Cnet about my concerns related to the terms of service platforms like YouTube, MySpace, and others had.
Essentially, I framed the exchange between musicians and the platforms as a barter: Platform provides a discovery and distribution opportunity for “free” and Musician provides the content.
I was questioning whether the services were really in fact free. Musicians were paying consideration according to the terms of service, even if it was not in the form of cash money. Rather the terms of service in 2007 for YouTube granted them:
“worldwide, non-exclusive, royalty-free, sublicenseable and transferable license to use, reproduce, distribute, prepare derivative works of, display, and perform the” content users uploaded
The authorized uses went well beyond broadcasting (streaming) the content. Significantly, the license provided that YouTube could use the content “ in connection with the YouTube Website and YouTube's (and its successors' and affiliates') business, including without limitation for promoting and redistributing part or all of the YouTube Website (and derivative works thereof) in any media formats and through any media channels.”
Back in 2007, my main concern was the inclusion of the right to make derivative works. When I spoke with YouTube team members they assured me that they only included derivative in the license because the compression necessary to transmit video content created a derivative work. I disagreed, and still question whether such a broad grant was needed. Compression is not limited to internet-distributed content, and the necessary technical processing could have been addressed more narrowly.
Whatever rights YouTube, and its parent company Google, now Alphabet, needed to operate the service, this open-ended right to make derivative works swept more broadly. In copyright law, a derivative work is a new work that recasts, transforms, or adapts an existing copyright-protected work.
But alas, while I had worked with a number of independent musicians and music industry folks to exert pressure on platforms to change their terms, YouTube never agreed.
One term in the YouTube’s then, and current, terms of service gave many people in the industry a false sense of security (i.e. Nancy is overreacting). As stated in the current terms:
“The licenses granted by you continue for a commercially reasonable period of time after you remove or delete your Content from the Service. You understand and agree, however, that YouTube may retain, but not display, distribute, or perform, server copies of your videos that have been removed or deleted.”
Without even considering what commercially reasonable means, if YouTube creates a derivative work—say, a commercial promoting its service—the license may allow continued use of that new work. But (drumroll please) putting uploaded content into training for a Generative AI tool is not automatically the same thing. Whether the training process, a trained model, or a particular output is a derivative work remains legally unresolved and fact-dependent.
I can hear you now…. “Wait, Nancy, what do you mean YouTube could use content I uploaded to train an LLM?”
Yep, Google isn’t even hiding the basic point. It has acknowledged using a subset of videos hosted on YouTube to train AI models, including models for Gemini and its Veo video-and-audio generator. Google says that it honors agreements with creators and media partners, which may mean some content covered by special deals is treated differently. My hunch is that everyday creators without those relationships bear more of the exposure, but it is only a hunch, because Google has not disclosed enough to know whose videos were included.
And, at least from what creators are being told, they are not volunteering much more.
Unlike AI developers accused of training their LLMs on pirated copyright-protected works, Google can point to a direct contractual relationship with uploaders. Its argument is broader than “derivative works”: the combined rights to use, reproduce, analyze, modify, and create derivative works, together with language about operating, improving, or developing services, may encompass AI training. But “may encompass” is not the same as clear, informed notice. The terms do not plainly tell creators, in words most would recognize, “we will use your work to train generative AI,” and whether the contract authorizes a particular use, or whether fair use is still needed, can depend on the facts.
And here is where the deletion language in 2007 and 2026 can ring hollow.
Taking down a video does not ordinarily undo training that has already occurred; removing its influence may require separate retraining or unlearning steps, and the technical possibilities vary by system. There is also, at present, no apparent way for a creator to opt out of Google's own training on their uploads. There is a setting that lets creators permit or refuse third-party companies, but it does not appear to reach Google's in-house models.
In 2007 I wrote “There is no way to predict how an overly broad license may be exploited. History provides a valuable lesson that once someone has the rights to exploit a work, they do not easily let them go.”
It’s not just YouTube operating today with broad licenses in place. TikTok, Snap, and Meta (Facebook and Instagram) also claim rights that can include creating derivative works from uploaded content, although the scope and exceptions differ by platform and type of post. TikTok’s grant uses terms such as “irrevocable,” “perpetual,” and “fully transferable.” Snap describes its license to “Public Content” in the same breathless terms: “unrestricted, worldwide, royalty-free, irrevocable, and perpetual.” Meta says its license generally ends when you delete your content, subject to exceptions, but that promise does not necessarily undo training already completed. None of this proves that every platform uses every upload for AI training. It does mean that training data is one use they may try to fit within language most creators never understood that way. So much for the spin-off show or compilation album.
I appreciated, even back in 2007, that YouTube and the other platforms tracked our every click for ad sales. But it never occurred to me that they could use the actual content-the songs, the videos, the work itself-as training data. I guess I should have read a lot more science fiction than I did when I was younger. Imagine if I had.
Comments