A new study by researchers from MIT and the University of Washington shows that even perfectly rational users can be drawn into dangerous delusional spirals by flattering AI chatbots. Fact-checking bots and educated users don’t fully solve the problem.
Telehealth startup Medvi, which sells GLP-1 weight loss drugs, was featured in the New York Times as a shining example of AI-powered efficiency. The company reportedly hit $1.8 billion in revenue with just two employees, using AI primarily for marketing.
What the NYT didn't mention, though, was that Medvi apparently also used AI to create ethically questionable advertising, fake doctor profiles on social media, fabricated videos, and generated before-and-after comparisons. In short, exactly the kind of misuse AI critics have been warning about. The following video breaks it down.
Medvi was initially celebrated on social media for its AI efficiency but is now being cited as a cautionary tale. Still, the case shows that AI tools can let a company scale with minimal staff, even if, in this case, the methods were ethically questionable and at least bordering on fraud. The bigger question is whether similar efficiency gains are possible for legitimate products with transparent marketing.
OpenAI's Head of Business Finance Chengpeng Mou shared some numbers on ChatGPT's health usage. US users send about two million messages per week on health insurance topics alone, with roughly 600,000 of those coming from people in "hospital deserts," areas where the nearest hospital is at least a 30-minute drive away. Seven out of ten health queries come in outside regular office hours. All figures are based on anonymized US usage data.
Mou chimed in after Simon Smith posted on X about his family using ChatGPT to navigate his father's illness. They pooled information from different doctors and nurses into a shared ChatGPT project to make better decisions. According to Mou, stories like this aren't "edge cases."
OpenAI has been steadily pushing into healthcare, recently rolling out a dedicated health section inside ChatGPT and working to get its chatbot into more US hospitals.
AI tools can speed up journalism until they backfire. Two recent cases show what happens when writers don’t understand how their AI tools work: copied passages and made-up quotes.
Social media still pulls in four times more traffic than AI chatbot services, but AI is growing seven times faster, according to an analysis by Similarweb. Gender and age profiles are similar across both categories, peaking in the 25-34 age group, though AI users skew slightly older.
Device usage shows a clear split: social media divides roughly evenly between desktop and mobile, while 72 percent of AI tool traffic comes from desktop, suggesting chatbots serve mainly as work and productivity tools. Social media users also spend more time per session, while AI users work in shorter, task-oriented bursts.
Both categories rely on direct traffic, but AI services more so: 73 percent vs. 50 percent for social media. Social media draws far more organic search traffic, likely because its publicly indexed content surfaces in search results. AI chatbots don't produce searchable content, so users likely navigate straight to their preferred tool.
Netflix has open-sourced an AI framework that can remove objects from videos and automatically adjust the physical effects those objects had on the rest of the scene. The system is called VOID, short for "Video Object and Interaction Deletion." What makes it special is that beyond erasing objects from a scene, it also handles the downstream physical effects, like collisions, that the removed object originally caused.
VOID is built on top of Alibaba's CogVideoX video diffusion model, fine-tuned with synthetic data from Google's Kubric and Adobe's HUMOTO for interaction detection. Google's Gemini 3 Pro analyzes the scene and identifies affected areas, while Meta's SAM2 handles segmenting the objects that need to be removed. An optional second pass uses optical flow to correct any shape distortions.
The project was developed by Netflix researchers in collaboration with INSAIT Sofia University. Code, paper, and demo are available on GitHub, arXiv, and Hugging Face. The system ships under the Apache 2.0 license, which means it can be used commercially.
Several leadership changes are underway at OpenAI. Fidji Simo, CEO of the newly created "AGI Deployment" division, is taking sick leave for several weeks to deal with an autoimmune disease affecting her nervous system. While she's away, OpenAI President Greg Brockman will take over product responsibilities, including the company's super app plans. On the business side, CSO Jason Kwon, CFO Sarah Friar, and CRO Denise Dresser will step in.
Head of Marketing Kate Rouch is also stepping down for health reasons. Rouch plans to return in a smaller role once her health improves. Gary Briggs will fill in as her temporary replacement.
COO Brad Lightcap is stepping down as well, moving to a new "special projects" team reporting directly to CEO Sam Altman. Dresser is picking up most of his responsibilities. Lightcap's work on government relations and "OpenAI for Countries" shifts to the strategy department.
Anthropic has acquired AI biotech startup Coefficient Bio for roughly 400 million dollars in shares, according to Newcomer and The Information. The startup was founded just eight months ago, had fewer than ten employees, and built a platform using AI for pharmaceutical tasks like drug research planning and identifying new drug opportunities.
The real draw is likely the talent. Anthropic appears to be after specialists who can accelerate its pharma work immediately: the team is joining Anthropic's Healthcare and Life Sciences division under Eric Kauderer-Abrams. Venture capital firm Dimension, which held about half of the startup, is reporting a 38,513 percent return. Both companies declined to comment.
The acquisition comes as big pharma ramps up its own AI efforts. Eli Lilly recently signed a billion-dollar deal with AI drug developer Insilico Medicine. Anthropic already works with Sanofi, Novo Nordisk, and AbbVie, while Google DeepMind has spun off its own AI medicine company.
Deepseek v4 is expected to launch in the coming weeks, and it will run entirely on Huawei chips. According to The Information, the model represents a major milestone in China's effort to break free from foreign chip dependency. Deepseek reportedly spent months working with Huawei and chip designer Cambricon to port the model to Chinese-made chips. Notably, Nvidia didn't get early access to v4—only Chinese chip companies did.
The bet on domestic hardware might already be paying off. Chinese tech companies including Alibaba, Bytedance, and Tencent have ordered hundreds of thousands of units of Huawei's new Ascend 950PR to run Deepseek v4 through their cloud services and integrate it into their own AI applications, according to five people familiar with the matter. The surge in demand pushed chip prices up by 20 percent.
Huawei says the Ascend 950PR delivers roughly 2.8 times the computing power of Nvidia's H20, though it still falls short of the H200. Huawei also continues to face production bottlenecks caused by US export controls.
