On Thursday (June 25), Google’s President of Global Affairs Kent Walker published a 21-page policy paper titled A Pragmatic Approach to AI Governance in America, laying out the company’s preferred framework for US AI regulation. The document covers a wide range of topics, from frontier model safety to data center energy infrastructure, but its copyright section carries the most direct implications for the music industry. Google argues that training AI models on publicly available web data should “remain protected” under fair use in the United States, and that copyright enforcement should focus on AI outputs, not inputs. That position is not new for Google. It is, however, now formally articulated in a policy document submitted to US policymakers at the precise moment that the music industry is fighting multiple copyright battles over the same question in federal courts across the country.
The core argument Google is making is grounded in the art student analogy its paper deploys: “Using publicly available web data for training models is a transformative, non-expressive use, like an art student taking inspiration from walking through a gallery, that should remain protected under fair use in the US and text-and-data-mining exceptions abroad.” The analogy is doing significant work here. An art student who walks through a gallery does not copy the paintings. Google’s AI models ingest the audio files, transcripts, and metadata that make up publicly available music in order to generate music that competes in the same market as the originals. Whether that process is more like inspiration or copying is the central question in Suno’s pending fair use defense in Massachusetts and in Udio’s ongoing Sony lawsuit in New York, both of which are heading toward their most consequential rulings in summer 2026.
On enforcement, Google argues the focus “should again be on outputs: whether a specific image or piece of text actually copies an existing work, regardless of how it was created.” Under that framework, the question of whether a label or publisher consented to their catalog being ingested into a training dataset is irrelevant. What matters is whether the AI-generated song that emerges from that training infringes a specific copyrighted work. Google says technical filters should not “automate subjective decisions like whether something is ‘too similar’ to a prior work,” and that infringing material is best handled through standard notice-and-takedown systems. That is precisely the framework the music industry has been fighting against: a post-hoc whack-a-mole system that puts the burden of identifying infringement on rights holders rather than requiring upfront licensing.
