Google has published a new policy report calling for a balanced approach to AI governance, rejecting what it describes as a false choice between over-regulation and no regulation. The company argues for a pragmatic, evidence-based middle way that addresses both frontier AI and widely deployed applications.
Training Data and Fair Use
The report asserts that using publicly available web data to train AI models is a transformative, non-expressive use, comparable to an art student drawing inspiration from a gallery visit. It maintains that such use should remain protected under fair use in the U.S. and text-and-data-mining exceptions in other jurisdictions.
Google also urges AI developers to “recognize rights holders’ desires to find beneficial pathways for creative professionals and knowledge workers.” It points to its own agreements “in which we are paying for access to and delivery of diverse types of specialized, non-public content” as examples of partnerships that are “most likely to be sustainable for developers and meaningful for the ecosystem.”
Outputs, Not Inputs
On enforcement, the report stresses that copyright concerns should focus on outputs: whether a specific image or text actually copies an existing work, regardless of how it was created. It states:
“Technical safeguards can help prevent models from generating outputs that reproduce works they were trained on. However, to protect the open space needed for new forms of creative expression, it is important that filters do not try to automate subjective decisions like whether something is ‘too similar’ to a prior work.”
Google recommends relying on established notice-and-action frameworks with standard reporting and takedown mechanisms to remove infringing content.
Scale of AI-Generated Content
The music industry has voiced doubts about the adequacy of such mechanisms given the rapid growth of AI-generated material. According to data from Deezer, 75,000 fully AI-generated tracks are now uploaded to streaming services every day, up from 10,000 daily at the beginning of 2025. Rightsholders question whether traditional notice-and-takedown processes can cope with this volume.
