Fear of ICE Raids Is Keeping Minnesotans From Clinics: Minnesota clinicians say fear of Immigration and Customs Enforcement activity near hospitals is causing patients to delay or skip care, even during a severe flu season. CIDRAP profiles Tina Ridler, a U.S.-born long-COVID patient postponing Mayo Clinic visits because she worries about traffic stops or being swept into enforcement near Rochester. Pediatric leaders report families canceling routine checkups and vaccines and waiting too long for emergencies; one child developed severe appendicitis after delays. Community groups say volunteers once drove patients, but some now avoid clinics altogether. Health systems are expanding telehealth and reviving mobile pediatric units to deliver vaccinations and well visits. The story notes similar declines nationally, citing a Physicians for Human Rights/Migrant Clinicians Network survey of providers reporting reduced visits since early 2025. (CIDRAP)

The Next AI Leap May Require Machines That Can “Imagine” Reality: Today’s LLMs and vision-language systems can describe the world, but they often can’t simulate it: they struggle with persistence, causality, and the kind of internal scene model that lets you predict what happens next. Scientific American looks at the push toward “world models” that learn a structured representation of environments (and update it over time), borrowing ideas from robotics and computer vision to move beyond pattern-matching. The piece highlights why this matters for safer autonomy: planning, navigation, and robust interaction depend on an AI that can test actions in a mental sandbox, not just autocomplete text. Researchers are pursuing approaches that fuse video, 3D/4D representations, and self-supervised learning to get there, but evaluation remains tricky and failure modes can be subtle. (Scientific American)

A Reference Giant Meets an AI-Driven Traffic Collapse: Wikipedia fought for legitimacy through policies, citations, talk pages, and relentless volunteer editing. Now it faces a different kind of threat: AI-powered search and chat tools that summarize answers without sending readers to source sites. Scientific American reports that Wikimedia has documented drops in human page views in parts of 2025 versus 2024, while an external analysis (using Similarweb data) estimated Wikipedia lost more than a billion average monthly visits between 2022 and 2025. The article frames the risk as structural: if fewer people visit, fewer people edit, and the “trust infrastructure” that makes Wikipedia reliable can weaken. The paradox is stark: the web still needs a vetted reference layer, but AI intermediaries may be starving it of attention and labor. (Scientific American)

AI Maps a Worldwide Boom in Floating Algae, Using 1.2 Million Satellite Images: A University of South Florida– and NOAA-led team used AI to assemble what they describe as the first global picture of floating algae blooms, finding expansion across the ocean with major implications for ecosystems, tourism, and coastal economies. The Phys.org report emphasizes the computational scale: processing and analyzing roughly 1.2 million satellite images required high-performance computing and took months. The authors link the growth to shifts in ocean temperature, currents, and nutrients, and note the “double-edged sword” of macroalgae: it can provide habitat offshore but becomes damaging when large decaying mats reach coastlines. The work also doubles as a case study in how AI can turn vast remote-sensing archives into testable, planet-scale environmental indicators. (Phys.org)

AI Boosts Individual Scientists—But May Narrow What Science Studies: A large-scale analysis reported in Nature suggests AI tools can dramatically amplify individual scientific output while shrinking the collective breadth of research. Summarizing the paper, Phys.org says the team analyzed 41.3 million papers and found researchers who use AI publish about 3.02 times as many papers, receive 4.85 times as many citations, and become research leaders roughly 1.4 years earlier than those who don’t. Yet, the article reports that AI adoption is associated with a 4.63% reduction in the overall volume of scientific topics studied and a 22% drop in engagement between scientists. The proposed mechanism is “data gravity”: AI pulls researchers toward data-rich, benchmark-friendly domains, concentrating attention and leaving more speculative or data-poor areas underexplored. (Phys.org)

OpenAI Will Test Ads in ChatGPT, Starting With Free and Lower-Tier Users in the U.S.: OpenAI says it will begin testing advertisements in ChatGPT “in the coming weeks,” an inflection point for consumer AI business models. According to an AFP report carried by SpaceDaily, ads will first appear in the United States for free and lower-tier subscribers, while Pro and Enterprise tiers will remain ad-free. The move reflects a basic economic tension: generative AI services are expensive to run (especially compute), and only a small fraction of users pay subscriptions. The story notes OpenAI’s soaring valuation in funding rounds and the pressure to diversify revenue beyond subscriptions—pushing it toward the ad-driven playbook that helped Google and Meta scale. The key question is product design: how ads are integrated without eroding trust, utility, or the feel of a “neutral” assistant. (SpaceDaily)

U.S. Eases Path for Nvidia H200 Sales to China, With Supply and Licensing Conditions: The U.S. Commerce Department’s Bureau of Industry and Security shifted review policy for Nvidia’s H200 and similar chips from a presumption of denial to case-by-case licensing, potentially allowing sales into China under conditions, AFP reports via SpaceDaily. The article says approvals would require factors such as evidence of “sufficient” U.S. supply, while the most advanced processors would remain blocked. It also describes uncertainty on the demand side: China has reportedly pushed firms toward domestic alternatives, and officials may restrict H200 purchases to special cases such as labs or university research. The piece situates the change within broader U.S.–China AI-chip geopolitics—where export controls, industrial policy, and supply-chain leverage shape not just who gets hardware, but which research ecosystems can scale frontier models. (SpaceDaily)

AI for Genetic Circuit Design Arrives: “CLASSIC” Trains Models on Massive DNA Libraries: Rice University researchers report what they call a first demonstration of using AI to design genetic circuits, enabled by a new method—CLASSIC (“combining long- and short-range sequencing to investigate genetic complexity”). The EurekAlert release explains the bottleneck: there are countless DNA designs that could, in theory, program cells to do useful things, but mapping sequence to behavior is a search problem at haystack scale. CLASSIC builds enormous libraries of circuits, captures each circuit’s full sequence with long-read sequencing, tags them with barcodes, and uses short-read sequencing to link genotype to phenotype at high throughput. Those datasets can then train ML models to predict promising designs not directly measured, closing the loop between wet lab and model-driven iteration. The release points to applications in engineered cell therapies. (EurekAlert!)

Autonomous AI Agents Screen Notes for Cognitive Decline: Mass General Brigham researchers describe an autonomous AI system that screens routine clinical documentation for signs of cognitive impairment, aiming to catch underdiagnosed cases earlier. In validation testing, the system achieved high specificity (98%) and, in balanced testing, reported sensitivity of 91%—though sensitivity fell to 62% under real-world conditions, a calibration gap the team highlights for transparency. EurekAlert says the approach uses an open-weight large language model deployed locally (no patient data sent to external cloud services) and coordinates five specialized agents that critique and refine one another’s judgments in an iterative loop. The study analyzed more than 3,300 clinical notes from 200 anonymized patients, and in disagreement cases an independent expert validated the AI’s reasoning 58% of the time. The team also released an open-source tool (“Pythia”) to support further study. (EurekAlert!)