Google Says the Upload Was the Licence.

Most AI training cases turn on fair use. This one may not get that far. In Kogon v. Google, a group of independent musicians say Google copied their YouTube recordings to train its Lyria music model. Google's motion to dismiss, filed on 8 June 2026, does not argue that training was fair. It argues that the musicians already granted the licence — by uploading.

What the musicians allege

The plaintiffs are working independent artists — among them a New York singer-songwriter, a Los Angeles composer, an Atlanta producer, a folk-rock father-and-son duo and the members of a Chicago band. They filed in the Northern District of Illinois on 6 March 2026.

Their complaint alleges that Google copied millions of copyrighted recordings, compositions and lyrics — at least 44 million clips and 280,000 hours of music — and used them to train Lyria 3. The claims run under the Copyright Act, the DMCA and the Lanham Act, with further claims under Illinois consumer-protection and biometric-privacy law.

The defence that isn't fair use

The striking thing about Google's response is what it declines to rely on. The industry's standard answer to a training claim is that training is transformative and therefore fair. Google leads instead with contract.

Its first heading is that the plaintiffs' express licence to YouTube defeats their claims. The argument is that in choosing to distribute their songs on YouTube, the musicians accepted terms of service that already authorised the conduct they are now suing over. If that is right, the fair-use question never has to be reached, because there was nothing to excuse.

The word doing the work

Everything hangs on a definition. Google's motion points out that each plaintiff granted YouTube and its "Affiliates" a "worldwide, non-exclusive, royalty-free, sublicensable and transferable license" to use uploaded content — including to "reproduce, distribute, [and] prepare derivative works".

Then it closes the loop: the terms of service, Google says, expressly confirm that "Affiliates" are the other companies within Alphabet Inc., including Google itself. On that reading the licence was never confined to the video platform. It extended, by its own definition, across the corporate group that owns it — and "prepare derivative works" is broad enough to reach a model trained on the audio.

Terms from 2008

Google's filing attaches historical versions of the terms as exhibits, reaching back to 2008 and 2012, and argues that the licence language has been substantially consistent across them despite differences in phrasing.

It also relies on the clause by which users agree to be bound by later modifications, which is how terms accepted years ago are said to carry the current wording forward. The practical claim is that an artist who uploaded a demo in 2009 has already licensed conduct that did not exist as a technology at the time.

That is precisely where the plaintiffs push back. Their counsel has said publicly that Google has not shown a single plaintiff ever agreed to its terms of service, and that no version of those terms has ever said a word about AI training.

The second argument, held in reserve

Google does not rest on the licence alone. It argues in the alternative that the plaintiffs have not plausibly connected their own recordings to the training data at all, describing their case as resting on an unsupported hypothesis that Google trained on their specific works.

It attacks the other counts separately — that the complaint fails to allege copyright management information was present, or the state of mind the DMCA provision requires, and that the biometric claims do not allege Google isolated or extracted voiceprints or that the data could identify anyone.

This layering matters. Even if the licence theory fails, the traceability problem is the one that has proved hardest for individual artists in AI cases: proving that your specific recording, rather than music like yours, went into the machine.

What a motion to dismiss can and cannot do

At this stage the court generally takes well-pleaded facts as true and asks whether the complaint states a claim. It does not weigh evidence or decide who is right about the technology.

Google's description of the licence is advocacy, not a finding. The court will have to interpret the text of the terms, the relationship between YouTube and other Alphabet companies, and any limits imposed by context, notice or applicable law — including whether a general grant to "use" content can be read to cover training a generative model that competes with the uploader.

Why the answer matters beyond this case

If the argument succeeds, platform upload terms become the central evidence in AI-training disputes, and the negotiated dataset licence becomes unnecessary for anything a platform already holds. Every musician who has ever posted a track would need to read a decade of terms of service to know what they have given away.

If it fails, the ruling will start to mark the outer edge of broad platform language — the point at which "use in connection with our business" stops absorbing whatever the business later becomes.

Either way the outcome will rest on the court's written analysis, not on the strength of Google's filing. Nothing has been decided yet.