AI News Roundup – State-level AI laws take effect with new year, xAI’s Grok generates sexualized depictions of minors, AI “world models” could revolutionize game development, 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.

• • New laws regulating AI in several U.S. states have taken effect with the new year, according to The Verge. The most high-profile are those in California, which include AI transparency requirements, whistleblower protections, restrictions on AI companions, and requirements that law enforcement agencies disclose their use of AI. This AI Roundup covered several of the California measures in further detail here. Texas is also implementing a law that prohibits the use of AI to cause harm, capture biometric information, or discriminate based on certain characteristics. Other states with AI regulations taking effect include Nevada, which passed a law that implements rules around the use of AI in election advertising, and New York, with safety and transparency requirements akin to California’s, though the latter will not take effect until March 19, 2026.
  • Reuters reports that Grok, the AI model created by Elon Musk’s xAI, is now generating sexualized images of women and minor children. Many users have reported that Grok, which is integrated into the social media platform X (formerly Twitter), has been used by X users to digitally “undress” pictures of women on the platform and put the people depicted into bikini swimsuits, and has even generated sexualized images of children. “Nudifying” computer programs have existed for many years, though the integration of the process into a large social media platform appears to have made the process much easier. Government officials have already begun to respond to these reports—French ministers have criticized X publicly and referred the company for prosecution, and India’s government sent a letter to X’s Indian subsidiary stating that the platform failed to prevent AI misuse. However, xAI owner Musk has dismissed concerns about Grok’s technology, responding to certain AI edits of bikini-clad people on X with a series of laughing emojis, while an earlier statement given to Reuters by xAI in response to reports of sexualized images of children stated that “Legacy Media Lies.”
  • Google DeepMind and other AI companies are developing a new series of “world models” that could find applications in the video game industry, according to the Financial Times. World models are AI models that recreate three-dimensional environments and have been applied in physical robotics tasks in the past. However, they also have the potential to revolutionize the development of video games. Genie, DeepMind’s world model, can create an entire virtual world from an image prompt, as this AI Roundup reported in May 2025. For instance, generative AI using world models as a foundation could allow gamers to create personalized interactive experiences, though some experts raised concerns about games becoming overwhelmed by low-quality AI “slop.” Indeed, the use of generative AI in game development has been controversial, with some users and critics pushing back against games such as ARC Raiders that use AI for game voices. However, generative AI is likely to continue to be used in game development as a potential cost-cutting measure.
  • Bloomberg reports on a new whitepaper from China’s DeepSeek AI detailing a new, more efficient training method for AI models. The paper, published this past week, describes Manifold-Constrained Hyper-Connections, or mHC, a way of structuring connections within AI models that reduces the computational (and thus energy) requirements involved in AI training without sacrificing model performance. DeepSeek, lacking access to the most advanced AI hardware due to U.S. export restrictions, has focused on making its model more efficient. The release of the company’s next flagship model, the R2, is expected in February. DeepSeek’s R1 reasoning model delivered a shock to the AI industry when it released in early 2025, offering comparable performance to leading U.S.-based models at a fraction of the development and training cost. It is unclear if mHC will be implemented in R2, though the company will likely continue to focus on efficiency unless export controls on advanced AI chips are lifted.
  • OpenAI is planning to release a new audio-focused AI model early this year, according to TechCrunch. The company has reportedly consolidated internal teams to focus on overhauling its audio-generation models, which will be designed to sound more natural and handle interruptions in conversations more fluidly. All of this comes as the company prepares to launch a personal device in 2027 to interact with the real world, and the company is reportedly planning a family of tech devices powered by ChatGPT, in partnership with former Apple hardware designer Jony Ive, as this AI Roundup reported in May 2025.