AI's legal and technical dependencies became clearer. A copyright settlement separated model training from source acquisition, Sony identified 30,117 recordings in a new Udio complaint, and Google reportedly explored hardwiring parts of Gemini into a future chip. Diffusion research supplied a narrower mechanism for novelty without resolving memorization.
1. A judge gives final approval to Anthropic's $1.5 billion copyright settlement
A federal judge approved Anthropic's $1.5 billion settlement with authors who said the company pirated books used to build Claude. More than 91% of covered authors and publishers claimed a share, while some rights holders opted out and retained separate cases. The agreement is the largest known recovery in a US copyright case.
The settlement addresses acquisition, not a general rule against AI training. The court had found model training to be fair use but held that storing more than 7 million pirated books in a central library could infringe independently. Final approval closes this class action without turning that district-court fair-use analysis into binding appellate precedent.
Sources: Reuters on final approval of the Anthropic settlement · TechCrunch on the acquisition and training distinction
2. Sony brings 30,117 more recordings into a separate lawsuit against Udio
Sony Music filed a new lawsuit alleging that Udio copied 30,117 recordings to train its AI music models. Sony identified the works after obtaining training information in discovery and applying audio fingerprinting. A judge had refused to add the catalog to an earlier case involving 333 works, prompting Sony to pursue the additional recordings separately.
The complaint seeks an injunction and statutory damages of up to $150,000 per infringed work, but its allegations have not been adjudicated. Udio has argued that training is fair use. Universal and Warner have since licensed their catalogs to Udio, leaving Sony's case to test whether later licensing deals affect liability for data acquired before permission was granted.
Sources: The Verge on Sony's new Udio complaint and recording list · Engadget on the separate lawsuit and requested damages
3. Google reportedly explores a Gemini-specific server chip for 2028
Google is reportedly designing a server chip, informally called Frozen v2, that would hardwire parts of Gemini to improve inference efficiency. The Information's anonymous sources put a possible deployment in 2028 and estimated six to ten times more tokens per unit of power than Google's current AI chips. Engineers were still deciding what model information to embed.
Google did not confirm the project or its performance estimate, saying that not every experiment reaches production. A model-specific design could lower energy and capacity costs, but it also couples hardware to model architecture years before deployment. Frozen v2 remains a research program, not capacity that customers can include in present infrastructure plans.
Sources: Reuters on the reported Frozen v2 project · TechCrunch on Google's response and tentative timeline
4. Score smoothing offers a mechanism for diffusion models' novelty
Google researchers argue that generative AI diffusion models produce novel samples because neural networks learn a smoothed approximation of the score function used during denoising. Their ICLR 2026 work links explicit or implicit regularization to interpolation: trajectories still move toward the hidden data manifold but are less likely to collapse onto individual training examples along tangential directions.
The paper supplies a mathematical mechanism for novelty without showing that deployed image models never memorize. Google describes an initial treatment with simplified settings and open questions about complex data distributions, architectures, and conditional generation. The testable connection is between training regularization, output diversity, and memorization risk.
Sources: Google Research on score smoothing in diffusion models