Google has set out its preferred rules for AI in a 21-page paper, A Pragmatic Approach to AI Governance in America, published on June 25 by its president of global affairs, Kent Walker. On copyright its position is blunt: training on publicly available web data is a transformative, non-expressive use that should stay protected by fair use in the US and by text-and-data-mining exceptions elsewhere. The paper likens it to an art student learning by walking through a gallery.
The core of the argument is that the law should look at outputs, not inputs. Whether a particular image or text copies an existing work should be what matters, not how the model behind it was trained. Google says automated filters should not be making subjective calls about whether something is too similar to an earlier work, and that infringing material is better handled by ordinary notice-and-takedown.
The paper offers some ground to rights holders. It says developers should let site owners opt out of training through machine-readable signals such as its Google-Extended control, that Google is exploring partnership and payment models, and that it already pays for some specialised non-public content, including creative and educational material. More broadly it wants regulation driven by evidence of real harms and willing to accept some uncertainty, rather than rules on how models are built. It also backs the NO FAKES Act as a single national standard against unauthorised digital replicas of voices and likenesses.
For music the timing is pointed. The majors, through the RIAA, sued Suno and Udio in 2024; Udio has since switched from defending fair use to licensing deals with Universal, Warner, Merlin and Kobalt, while Sony keeps suing. Publishers are pressing two cases against Anthropic over lyrics, the second covering more than 20,000 songs, and industry groups have asked a court to reject its fair use defence. Google is itself a defendant, sued in March by independent artists who say Lyria 3 was trained on their YouTube uploads, a case it is trying to dismiss by pointing to YouTube's terms.
Outside copyright, the paper proposes an industry-funded, federally overseen body to set safety standards and verify audits for the most advanced models, and argues that anything illegal without AI remains illegal with it. It describes this as a middle path between heavy regulation and none. Google has argued the fair use point since at least a 2023 Copyright Office filing, where it also said courts rather than new laws should settle the question. The paper's talk of value exchange also echoes a run of music licensing deals, including the NMPA's industry-wide agreement with Udio in June.
Sources · 1