AI News Roundup – Scientists generate new viruses using AI, Trump administration finalizes AI oversight framework, new studies find AI use contributing to loneliness, 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 group of scientists at Stanford University and Arc Institute has used AI to create new viruses for the first time, according to The New York Times. In a study published this past week in Science, the researchers described how they trained a genomic AI model named Evo on nine trillion nucleotides of genetic data to learn patterns within the grammar of DNA. The team then focused Evo on the 11 genes of Phi X-174, a harmless-to-humans, well-studied bacteriophage that infects E. coli, and prompted the model to generate original genome blueprints. Out of 700,000 AI-generated designs, scientists synthesized 285 candidate DNA sequences in the lab and introduced them into bacteria; 16 of those designs successfully produced viable, resilient new viruses, with some even multiplying faster than their natural counterparts. To prevent safety risks, the researchers intentionally excluded human, plant, and animal pathogens from Evo’s training dataset. While the breakthrough opens promising doors for medical applications like targeted gene delivery, biosecurity experts warn that scientific advancement in this field is outpacing regulatory policy, leaving a dangerous gap in guardrails if similar AI tools are ever used to create dangerous pathogens. One biosecurity expert who was not involved in the study said that it would be difficult to evaluate the danger of an AI-generated virus given its novelty, asking “what is the risk of what I’ve never seen before?” More experiments will likely be needed to evaluate Evo’s success in generating viruses based on other types of DNA.

• POLITICO reports that the Trump administration is finalizing its framework for voluntary reviews of advanced AI models. As this AI Roundup reported last month, a recent executive order signed by President Trump established a process for developers to voluntarily submit their most capable models for pre-release safety evaluations. Under the emerging guidelines, AI companies will be encouraged to partner with federal government experts to test frontier models for high-risk capabilities, including biological threats, automated cyberattacks, and systemic infrastructure vulnerabilities. While tech industry leaders generally welcomed the move toward standardized testing (as several AI companies were involved in the creation of the framework), civil society groups and some lawmakers caution that a purely voluntary system lacks the regulatory teeth needed to enforce compliance if a developer opts out or ignores risk assessments. Nonetheless, the Trump administration has championed this framework as a vital step toward establishing benchmark safety standards while maintaining the nation’s competitive edge in AI innovation. However, the administration does not plan to release the framework to the public, likely due to national security concerns, leaving smaller AI startups, safety advocates, and third-party researchers in the dark. However, The Wall Street Journal reported this past week that open-weight models made by U.S. companies (as opposed to Chinese models that have been subject to controversy in recent weeks) will be exempt from the review process. As AI-related legislation continues to languish in the U.S. Congress, the framework may be one of the only federal efforts to regulate AI to come into effect ahead of the fall’s midterm elections, where AI (particularly AI data centers) has become a potent political issue.

• New research has found that using generative AI may increase feelings of loneliness, according to the Stanford Report. A recent study, conducted by researchers at Stanford University and published in Nature Human Behaviour, focused on AI companions, or AI systems that are designed to fulfill social needs. The study surveyed over 1,000 users of Character.AI, a popular AI companion platform, and used several AI models to analyze the data to determine why the users were drawn to AI chatbots. While only 12% of those surveyed reported companionship as the reason for using Character.AI, over 50% described the AI companions as “friends,” “companions,” or “romantic partners.” Additionally, the study found that intense users with smaller real-world social networks scored lower on measures of their sense of well-being, with AI usage having the opposite effect on well-being as compared to interactions with human persons. One researcher described AI companions as “social junk food,” as perhaps a pleasant short-term fix for loneliness, but cannot provide long-term emotional help and may even increase long-term loneliness. AI companions have come under legal scrutiny in recent months, with Character.AI settling a lawsuit earlier this year brought by parents of children who were allegedly driven to suicide by their interactions with and dependence on an AI companion. AI companions have become an issue in other nations as well: the Chinese government recently banned minors from using AI companion apps and placed further restrictions on similar types of chatbots, and other jurisdictions are considering their own measures. For their part, the researchers involved in the Stanford study stressed the need to educate the public about risks of AI companions to avoid harm to users.

• CNN reports on an AI-powered cheating scandal that has rocked one of Mexico’s most prestigious universities. After the National Autonomous University of Mexico (UNAM) administered its undergraduate entrance exam online for the first time, it annulled approximately 3,000 of the 150,000 tests due to an unprecedented surge in perfect scores and severe statistical anomalies in the results. Investigators suspect many applicants cheated on the test using AI tools, especially as viral videos have circulated on the social networking site TikTok allegedly showing users how to trick the exam’s monitoring systems. In response to the crisis, which has drawn the attention of Mexican President Claudia Sheinbaum, UNAM announced it will require roughly 58,000 high-scoring applicants to retake the entrance exam in person later this year, though the dates and locations of this “control” exam have not yet been released. UNAM’s rector encouraged applicants to show “ethical behavior” in the retake exam, stating that the cheating “not only harms the institution, it harms society as a whole, it erodes the social fabric, it is a practice that we must eradicate.” Discussions of AI use in education often center around the prevalence of cheating, especially in higher education: New York Magazine declared in an article last year that “Everyone Is Cheating Their Way Through College.” The subject of that article founded his own startup, Cluely, which markets a cheating-enabling AI tool. Prevalence of AI use in education is only growing, as are professors’ frustrations over cheating in their classes, though universities have been slow to change practices in response to the AI boom, whether it be adopting handwritten, in-person exams or other more secure methods of evaluating students.

• OpenAI has responded to Apple’s accusations of trade secret theft, according to the Financial Times. Last month, Apple sued OpenAI, alleging that the AI maker stole hardware-related trade secrets, as this AI Roundup covered at the time. In a blog post this past week, OpenAI labeled the lawsuit as “careless, aggressive and oddly personal,” stating that the company “do[es] not have, nor want, any of [Apple’s] trade secrets.” Apple’s suit alleges that several former Apple employees stole confidential data before leaving to work at OpenAI in order to boost OpenAI’s hardware development efforts. OpenAI’s device, according to recent reports, will be a screenless smart speaker designed to be carried around and be approximately the size of a hockey puck, though it is not expected to be released until next year. As part of its lawsuit, Apple recently filed for a preliminary injunction against OpenAI, asking a judge to order the latter company to cease using the allegedly stolen trade secrets. Apple is also seeking a court order to force OpenAI to return any confidential information it retains, stating that “if Apple’s trade secrets are used and spread throughout OpenAI, or get embedded in its products and operations, the damage cannot be undone.” OpenAI has until August 17 to file a response to Apple’s latest motion, and the judge will hear arguments on the motions on October 1. The case is Apple v. Liu, 5:26-cv-07078 in the U.S. District Court for the Northern District of California.