AI News Roundup – Efforts to regulate AI companions grow, Italy passes AI regulation bill, DeepMind and OpenAI models win coding competition, 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 MIT Technology Review reports on growing efforts to regulate AI “companions.” AI has been used by many people for emotional support, but particular attention has been focused on the use of the technology by young people. Recent lawsuits have placed particular attention on AI companion behavior in models from Character.ai and OpenAI that is alleged to have contributed to the suicides of two teenagers. The U.S. Senate recently held a hearing on the matter, and California’s legislature recently passed a bill that would require stricter rules for AI companies who have underage users. The Federal Trade Commission (FTC) also announced an investigation into several major AI companies, including Character.ai and OpenAI regarding their AI companions. Further action on this serious matter at all levels of government is thus expected in the coming weeks. 
• • The Guardian reports on a new AI regulation law approved this past week in Italy. The legislation, the first in the EU to follow the bloc’s AI Act, is aimed to promote “human-centric, transparent and safe AI use.” The bill provides for 1- to 5-year prison sentences for spreading harmful AI-generated content, as well as harsher penalties for using AI to commit other crimes. Other provisions would require children under 14 years of age to obtain parental permission to use AI, and to allow data mining only on non-copyrighted content. Italian Prime Minister Giorgia Meloni recently spoke of AI as “the greatest revolution of our time,” but also that it required “a framework of ethical rules that focus on people and their rights and needs.” 
• • The Financial Times reports on the success of AI models at the International Collegiate Programming Contest World Finals this past week. Models from both Google DeepMind and OpenAI performed at a “gold-medal level” during the prestigious event, known as the “coding Olympics.” AI companies have often used competitions such as this to show off the capabilities of their AI models — both DeepMind and OpenAI have also achieved gold at the International Mathematical Olympiad this past summer. In response to the latest results, Quoc Le, vice president of DeepMind, said that “this is a historic moment towards [artificial general intelligence],” the type of AI that surpasses human capabilities, adding that “solving math and computer competitive coding is a key step to understanding how our intelligence works.” 
• • The South China Morning Post reports on a new paper detailing the internal workings of DeepSeek’s R1 model. The paper, published in Nature this past week, describes the reinforcement learning process used by DeepSeek to bring out reasoning abilities in its R1 AI model, which made great waves in the AI industry earlier this year by claiming comparable performance to competitor models from U.S. companies at a much smaller cost. Part of this efficiency, according to the company, came from the reinforcement learning process, which allowed them to bypass a supervised fine-tuning stage of model training that most other models undergo. While R1 did still involve some human supervision, it had much less than other models, leading the developers to conclude that the model’s prowess “serves as a reminder of the potential of [reinforcement learning] to unlock higher levels of capabilities in LLMs, paving the way for more autonomous and adaptive models in the future.” 
• • The New York Times reports on new AI-enabled smart glasses unveiled by Meta this past week. At its annual developer conference, the company unveiled three new types of glasses, developed in partnership with eyeglass brands Ray-Ban and Oakley. The Meta Ray-Ban Display glasses include a small screen that can display apps and play music, as well as an AI-based voice assistant that can “see” through a glasses-mounted camera. The glasses experienced some issues during a live demonstration during the event, as it failed to provide a recipe and make a phone call as instructed. Meta CEO Mark Zuckerberg has billed the glasses as “able to see what you see, hear what you hear, and then go off and think about it,” and as a continuing part of the company’s “metaverse” virtual world. The Display model will be released on September 30.