AI News Roundup – Trump administration plans restrictions on DeepSeek and Nvidia, OpenAI releases latest AI models, UAE to use AI in lawmaking, 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.

• • The New York Times reports that the Trump administration is planning restrictions on the Chinese AI startup DeepSeek, as well as Nvidia, the world’s leading supplier of AI-focused processors. In particular, the government is considering measures to block DeepSeek from purchasing U.S. technology and to restrict the ability of Americans to access DeepSeek’s products. Earlier this past week, the government had also further restricted the sales of some of Nvidia’s most advanced chips to China, marking the first major limits that the second Trump administration has placed on the technology. After the Biden administration imposed export restrictions on Nvidia in 2022, the company modified its H100 chip so as to not be covered by the restrictions. The result, the H20, is exclusively sold in China. The measures come amid rising geopolitical tensions between the U.S. and China, marked by an escalating trade war and fears in the U.S. of losing AI supremacy to their rival. 
• • OpenAI has released the next generation of its flagship AI models. In a series of blog posts, the company described its latest reasoning models, OpenAI o3 and o4-mini, as well as GPT-4.1, the next iteration of GPT-4o, its “omni” multipurpose large language model. The company billed the new reasoning models as “the smartest models we’ve released to date,” and claims it presents a new state of the art for several benchmarks focused on coding, math, and science. The reasoning models also now have the ability to integrate images into their chain of thought (the “thinking” action of the model), allowing the models to use visual reasoning and manipulate images as part of their thought process. The company claims GPT-4.1 outperforms GPT-4o across the board, and supports longer context windows (the amount of text tokens the model “remembers” while in operation). OpenAI o3 and o4-mini are now available to paid ChatGPT users, while the GPT-4.1 family is available to application developers through OpenAI’s API. 
• • The Financial Times reports that the United Arab Emirates is aiming to use AI to help write and revise legislation. While details on the plan have not been released to the public, state media in the UAE has called the plan “AI-driven regulation.” Sheikh Mohammad bin Rashid Al Maktoum, the ruler of Dubai and Vice President of the UAE, said that the “new legislative system, powered by artificial intelligence, will change how we create laws, making the process faster and more precise.” Sheikh Mohammad also said that AI will be used to “regularly suggest updates to our legislation,” and the government stated that lawmaking could be sped up by 70%. Several AI experts expressed concerns with the plans, noting that AI models often have reliability issues and can generate falsehoods and errors through hallucinations. The UAE has invested heavily into AI technologies in recent years as it seeks to diversify its economy away from petroleum, directing state-backed investment funds to contribute to AI infrastructure projects globally, and has even proposed attracting AI companies through creating a “special economic zone” that would eliminate copyright laws, freeing them to train AI models without running afoul of intellectual property laws. 
• • Bloomberg reports that Apple is planning to indirectly use user data as part of the training process for its AI models. Apple generally trains its AI models using synthetic data meant to mimic real inputs but without containing sensitive information. However, the synthetic data does not always accurately reflect the real world, which can limit the abilities of AI systems that are trained on such data. To remedy this, Apple has developed a new component in its AI training pipeline that takes synthetic data and compares it to a sample of emails from a user’s email app, and thus determines which items in the synthetic dataset are most in line with real items. This information is then used to improve the text-generation capabilities of the company’s Apple Intelligence AI platform. Apple’s AI efforts have accelerated in recent months as the company attempts to catch up to rivals Google and Meta in the space and amid internal turmoil in the company’s AI team. Further improvements to Apple Intelligence are expected this summer, though AI-powered upgrades to Apple’s voice assistant Siri are not expected until 2026. 
• • Google’s DeepMind AI subsidiary has developed an AI model focused on analyzing the sounds used by dolphins for communication. According to the company, DeepMind partnered with the Wild Dolphin Project (WDP), a Florida research nonprofit, to build a variant of its open-source Gemma AI model that was trained on WDP’s database of sounds sourced from the Atlantic spotted dolphins that WDP studies. The resulting model, DolphinGemma, can process sequences of dolphin sounds to identify patterns and structure in the audio and ultimately predict the next sounds in the sequence. WDP will begin deploying the model, which is optimized to run on Google’s Pixel phones that WDP also uses, this season in search of identifying larger patterns within dolphin communication that could serve as a foundation for future interaction between dolphins and humans.