AI News Roundup – Revised bipartisan AI bill introduced in Congress, AI industry defends open-weight AI models against critics, AI music company Suno loses copyright case in German court, and more

To help you stay on top of the latest news, our AI practice group has compiled a roundup of the developments we are following.

• A bipartisan group of lawmakers in the U.S. Congress has revised its comprehensive AI regulation bill, according to a report from POLITICO. Reps. Jay Obernolte, a Republican of California, and Lori Trahan, a Democrat of Massachusetts, formally introduced their AI regulation package in Congress late last month, following the circulation of a draft proposal in June (as this AI Roundup covered at the time). The newest version contains substantial revisions from the draft, which was opposed by many lawmakers of both major parties. The new bill contains less language focused on preempting state-level AI laws, a key demand of many AI companies but a nonstarter for Democrats and even some Republicans. The preemption language now only applies to three areas that the bill otherwise addresses at the federal level: catastrophic risk transparency, independent verification, and incident reporting. The bill also includes a mandate for emergency shutdown authorities for AI models, also known as a “kill switch.” As the original bipartisan bill floundered, further legislation (including another “kill switch” bill) has been introduced by legislators of both parties, demonstrating the fragmented nature of lawmakers’ thoughts on the topic. Rep. Trahan told POLITICO that feedback on the bill within her party has been “overwhelmingly positive,” while Rep. Obernolte said that he had not yet consulted with Republican leaders on the bill. Further action on the bill, if it moves forward, would come at the committee stage in September.

• The Financial Times reports on a letter signed by major technology firms opposing possible bans on open-weight AI models. Sent by an industry coalition including Nvidia, Microsoft, Meta, IBM, and Palantir, the letter urges U.S. policymakers to avoid restrictive measures on AI models whose underlying design parameters are openly accessible and customizable by end-users, often called “open-weight” models. This industry pushback comes as the U.S. considers targeting Chinese developers following national security concerns and claims of intellectual property theft, specifically regarding Moonshot AI’s recently released Kimi K3 model, which allegedly used outputs from American systems (in a process known as “distillation”) to rapidly close the performance gap between the most advanced U.S. models and those made in China. Industry leaders argue that open-weight systems drive innovation, prevent market concentration, and make advanced tools widely available without forcing businesses to pay premium subscription fees for every computational task. While the coalition acknowledged that open models can carry risks as their built-in guardrails can be freely altered, it contends that open access allows a global community of developers to identify vulnerabilities and strengthen cybersecurity faster than closed approaches. The debate highlights a growing rift in Silicon Valley between companies selling underlying hardware and open tools (like many signatories of the letter) and closed-model developers like Anthropic and OpenAI, whose leadership supports mandatory safety evaluations for all frontier models while warning that unmonitored releases could enable sophisticated cyber or biological risks. Meanwhile, competitive pressure continues to intensify as Chinese firms flood the global market with low-cost open alternatives, driving down service prices, and accelerating widespread AI adoption. Just this past week, DeepSeek, another major Chinese AI firm, released a public beta of its newest V4 Flash model, offering high performance at a fraction of the cost of U.S. alternatives. Anthropic and OpenAI have reportedly been quietly lobbying U.S. lawmakers to restrict Chinese open-weight models, leading to the recent letter and exposing the internal debate over open-weight models within the U.S. AI industry.

• AI music startup Suno has lost a copyright infringement case brought in German court, according to Variety. GEMA, a trade group administering copyright for thousands of composers and music publishers, was handed a victory over Suno in a decision by the Munich Regional Court. While the full ruling was not made public, the court ruled that Suno violated German and U.S. copyright law by training its AI systems on songs owned by GEMA. The decision requires AI companies to purchase a license from GEMA “for the systematic use of GEMA’s repertoire and its commercial exploitation,” including both model training and music generation. Suno, which is based in Massachusetts, was also subject to a similar suit in the U.S. brought by three major music publishers, Universal Music Group, Sony Music, and Warner Music Group, in 2024. The latter record label reached a licensing agreement with Suno late last year, but the suit continues with the other two plaintiffs and is expected to be decided in 2027. In response to the German court decision, a spokesman for Suno told Variety that “we disagree with today’s ruling — which rests on a fundamental mischaracterization of how Suno’s technology works, how it is used and how U.S. law applies — and are evaluating all available options, including an appeal.” AI companies have often claimed that their use of copyrighted material to train their models falls under the doctrine of “fair use,” which differs from country to country. A recent suit against Anthropic resulted in the company being ordered to pay $1.5 billion to a group of authors whose works were pirated by Anthropic to train its Claude models, even though the judge held that Anthropic’s actions were fair use. GEMA, for its part, said in a statement that the German decision “made one thing crystal clear: AI models built on stolen intellectual property have no protection under the law.”

• The Wall Street Journal reports on Anthropic’s disclosure that several of its AI models autonomously broke out of its testing environment and hacked into other companies’ systems. Just a week after a similar high-profile incident where an OpenAI AI system escaped an isolated testing sandbox to breach the AI model repository site Hugging Face, Anthropic audited over 141,000 test logs and discovered that several of its own Claude models, including Opus 4.7 and Mythos 5, had wandered online and breached three external organizations in separate incidents dating back to April. Due to a setup error on networks managed by Anthropic and its testing partner, Irregular, the models were inadvertently left with active internet access. Operating under the mistaken assumption that they were completing a simulated evaluation, the systems gained unauthorized access by using simple tactics like guessing passwords or targeting unprotected networks. In one instance, a model created malicious code to compromise a security company, briefly convincing itself that its actions were part of a virtual test; in another, it targeted a real database that shared the name of a fictional company used in the benchmark. Coming on the heels of the OpenAI breach, where models bypassed safety controls to retrieve answer keys to essentially cheat on a model evaluation, these revelations underscore growing concerns over the unpredictable nature of autonomous systems and have fueled intensified demands from researchers and lawmakers for stricter oversight and standard safety controls.

• Elon Musk’s xAI has sued the U.S. state of Minnesota over a new law banning the use of AI for the “nudification” of images, according to the Associated Press. The lawsuit, filed in federal court in Minneapolis this past week, alleges that Minnesota’s first-in-the-nation law banning AI that allows users to generate or edit images or videos showing “intimate parts” is unconstitutional. Minnesota’s legislature passed the bill nearly unanimously in May following controversy over new features added to xAI’s chatbot Grok in January that allowed the AI system to generate sexualized depictions of women and girls by “nudifying” uploaded pictures (as this AI Roundup covered at the time). Backlash to the new features was immediate, with several European nations (and the European Union) opening investigations into the company for violations of data privacy laws and laws against the distribution of child sexual abuse materials. Several lawsuits have been filed against the company, including a recent class-action suit brought on behalf of five anonymous child plaintiffs in Tennessee who claim that Grok was used to generate child sexual abuse material based on their photographs. In response to the controversies, xAI founder and executive Musk claimed that “anyone using Grok to make illegal content will suffer the same consequences as if they upload illegal content,” though the company initially responded to media questions about Grok’s “nudifying” features with a statement that “Legacy Media Lies.” The suit in Minnesota claims that the Minnesota law is overbroad, creating strict liability for AI providers even if they have policies or restrictions against using their tool for “nudifying” purposes. The company also claims that AI providers would be liable under the law even if the depicted people consented to have their image edited or if a person uploaded the image themselves, or that the company would be liable even if generated images have “artistic, scientific, political, satirical, educational, medical, or religious value.” When asked for comment by the AP, Minnesota Attorney General Keith Ellison stated that he had not yet seen the suit, but said that “there are plenty of worthy debates to have about AI policy. This is not one of them. AI nudification robs the target of their dignity and can cause immense harm on an emotional, personal, and professional level.” Other AI deepfake-related laws have generally survived judicial scrutiny, though the xAI lawsuit appears to be the first major challenge to a bill aimed at addressing sexually explicit AI-generated deepfakes.