Google has published a comprehensive 21-page policy paper titled 'A Pragmatic Approach to AI Governance in America', detailing the company’s vision for the future of artificial intelligence regulation. The tech giant maintains that training AI models on publicly available web data must remain protected under the principle of fair use. By framing the ingestion of such data as a transformative and non-expressive process—analogous to a student finding inspiration in an art gallery—Google aims to define the parameters for ongoing legal debates regarding copyright.
Key takeaways
- Google asserts that using web data for model training constitutes a transformative, non-expressive use protected by fair use.
- The company proposes that copyright enforcement should prioritise vetting AI outputs for plagiarism rather than regulating the input phase of development.
- Google recommends the creation of a 'Frontier AI regulatory organisation' (FARO) to standardise safety assessments.
- The proposals arrive amidst intense legal scrutiny from the music industry regarding the unauthorised use of creative works.
Defining fair use for AI training
Google argues that the scientific advancement of AI models should not be hampered by restricted access to information. In its paper, the company highlights its use of technical controls, such as 'Google-Extended', which allows website owners to opt out of having their content indexed for model development. While the firm maintains that training remains fair use, it is also exploring partnerships and value-exchange models to compensate rights holders, particularly those providing non-public, specialised content.
Shifting focus to model outputs
Central to the company's regulatory proposal is a fundamental pivot in how legislation addresses copyright infringement. Google contends that regulators should move away from micromanaging the development stage—the 'inputs'—and instead focus on whether specific 'outputs' from a generative model replicate protected intellectual property. This approach intends to address tangible harms, such as infringing digital replicas, without discouraging the innovation behind large language models. The paper also supports existing legislative efforts like the 'NO FAKES ACT' to safeguard individual likenesses and voices.
Proposals for new regulatory oversight
Beyond copyright, Google advocates for a structured approach to safety, suggesting the formation of an independent, industry-funded body known as a Frontier AI regulatory organisation. This body would be overseen by a federal agency and tasked with verifying audits and establishing safety standards for the most advanced models. The company frames this as a 'middle way'—a balanced regulatory framework intended to distinguish between widely deployed AI tools and frontier models while ensuring the industry maintains agility in an evolving legal landscape.