Two US copyright rulings in three days. What they mean for your generated content
Both went largely in favour of the AI companies, and both drew the same distinction. For an SME the practical takeaway is narrower than the headlines.
On 23 June 2025 a federal judge in California held that training a language model on copyrighted books was fair use, describing it as exceedingly transformative — while finding that downloading millions of pirated books for a permanent internal library was not. Two days later, a different judge in the same district reached a broadly similar conclusion on training in a separate case against another company.
These are first-instance US decisions. They will be appealed, they do not bind anyone in Switzerland, and European courts may land elsewhere entirely. With that said, they are the first substantive rulings on a question that has hung over commercial AI use since 2022.
The distinction both drew
Training on lawfully obtained material was treated as transformative. Acquiring the material unlawfully was treated as a separate wrong, not excused by what was done with it afterwards.
That split is the useful part. It suggests the legal question is moving from « is training on copyrighted work permitted » toward « how was the training material obtained » — which is exactly the kind of thing the documentation arriving on 2 August is meant to describe.
What this does not settle for you
Three things, and they are the three that actually affect a Swiss SME.
Who owns what a model produces for you. Not addressed. Under Swiss law, protection generally requires human creation, so a purely machine-generated image or text may have no copyright protection at all — meaning a competitor could reuse your generated marketing material.
Whether an output infringes something. Also not addressed. These cases were about inputs. If a generated image closely resembles a protected work, that is a separate question with a separate answer.
Anything about Swiss or EU law. The EU route runs through the text-and-data-mining provisions and the AI Act’s copyright policy requirements, not through American fair use.
The practical position, unchanged
Use generated material where you would have used stock. Do not use it where similarity to an existing work would be a problem. Keep the record of what was generated and when, as recommended on 25 April — that field is what lets you answer a question two years from now.
And when your suppliers publish their copyright policies in August, read the one behind the tool you use for customer-facing material. That is now a document that exists.