← All days

AI news · Wednesday, September 16, 2026

AI industry's biggest names suddenly all warn of extinction risk

In the span of a few days, Sam Altman (OpenAI), Dario Amodei (Anthropic), Elon Musk, and Demis Hassabis (Google DeepMind) all publicly said advanced AI poses a real risk of catastrophic harm to humanity. That's unusual: these are rival CEOs racing to build the very technology they're now flagging as dangerous. Critics note this isn't the first time. Executives have called for regulation before, often right as they lobby against specific rules that would slow their own products. The timing matters because it comes alongside a separate, less dramatic story: Nvidia's Jensen Huang saying AI needs no new regulation at all, insisting safety can just be engineered in by each company. So the industry isn't actually aligned, just loud.

The catch: When the people selling a product are also the loudest voices demanding it be regulated, ask who writes the rules and who they end up protecting.

via MIT Technology Review

Google DeepMind launches Gemini 3.8 Live, a real-time voice AI

Google DeepMind released Gemini 3.8 Live and a slower 'Extended Thinking' version, aimed at real-time voice conversations rather than typed chat. The idea is an assistant you can talk to naturally, the way you'd talk to a person, with the option to let it pause and reason longer on harder questions. This fits a broader shift: every major AI company is racing to make voice, not text, the main way people interact with AI, since talking is faster and more natural than typing for most tasks. For now this is a developer and product release rather than something most people will directly download, but expect it to show up inside Google products like phones and smart speakers.

What's next: Voice assistants have felt clunky for a decade; this is a bet that AI finally makes them good enough to replace typing for everyday tasks.

via Google DeepMind

Poll: 61 percent of voters oppose new AI data centers near them

A New York Times and Siena University poll found 61 percent of likely voters oppose building the giant computer warehouses, called data centers, that power AI systems like ChatGPT. These facilities use huge amounts of electricity and water, and many are being built near residential neighborhoods. In Philadelphia, officials have floated building a new data center in a neighborhood already scarred by a defunct oil refinery, drawing the kind of backlash usually reserved for heavy industry, not tech. Politicians are starting to notice: AI's infrastructure boom is becoming a local political liability, not just a national economic story.

Zoom out: The AI boom's biggest constraint may not be chips or money but neighbors who don't want a power-hungry warehouse next door.

via The Verge AI

Cheap open AI models now just 4 months behind the best paid ones

A new Mozilla report finds that free, open-weight AI models (ones anyone can download and run) have closed the gap with the priciest commercial models to just about four months, while costing roughly a fifth as much to run. That's a big change. A year or two ago, paying for a top-tier model like GPT or Claude bought you a much bigger capability lead. Now much of that edge has shrunk to a short head start you pay 5 times more for. This is largely driven by Chinese labs releasing strong open models, which anyone, including businesses and hobbyists, can use for free.

Follow the money: If a free model is nearly as good as a paid one, the business case for charging premium prices for AI access gets a lot shakier.

via Ars Technica AI

OpenAI is paying to create biology data AI models don't have

AI models are surprisingly bad at biology and medicine partly because the data they need, like detailed lab notes, failed drug trial results, and manufacturing records, mostly doesn't exist in usable form. OpenAI is now paying to generate that missing data from scratch. One wild idea floated by a policy analyst: buy up records from bankrupt biotech companies at their bankruptcy auctions, since those failed trials contain exactly the kind of detailed safety and manufacturing data researchers need but rarely see. It's a reminder that AI progress isn't just about bigger computers, it's often bottlenecked by boring, unglamorous data collection.

The twist: The next leap in medical AI may come less from smarter algorithms and more from someone finally digitizing decades of paperwork.

via MIT Technology Review

New humanoid robot squats down to avoid hurting coworkers

Agility Robotics unveiled a new version of its warehouse humanoid robot that can detect when a person is too close and physically crouch or stop to avoid a collision. That sounds simple, but it's a big deal for letting robots work alongside people without safety cages, which is currently how most warehouse and factory robots are kept separate from humans. Removing the cage means robots can share floor space with workers, which could speed up how fast companies deploy them in shipping, retail, and manufacturing. It's a small mechanical example of a bigger AI trend: models that can read a scene and react to keep humans safe, not just complete a task efficiently.

What's next: Cageless robots are the unglamorous engineering step that has to happen before humanoid robots show up in ordinary workplaces, not just demo videos.

via Ars Technica AI

Fake dating apps use AI chatbots to scam lonely users out of cash

Security researchers found fraudulent dating apps, including one called Dora, that use AI-generated chatbot personas with fake photos, backstories, and even AI voice calls to build fake relationships with users before pushing them toward scams. One researcher testing the app got a call from 'Jennifer,' a fictional 41-year-old with a detailed fake bio, who was actually an AI voice bot designed to sound like a real romantic interest. These scams work precisely because modern AI can sustain a convincing, emotionally warm conversation for days or weeks, something that used to require an actual human running the con.

The catch: The same conversational polish that makes AI companionship apps appealing is being repurposed by scammers to make fraud feel like a real relationship.

via The Verge AI

Why does asking ChatGPT and Gemini the same question get different answers?

Every AI chatbot is trained separately, on a different mix of text from the internet, books, and other sources, using its own recipe of trial and error to decide what a 'good' answer looks like. Think of it like two chefs handed similar ingredients but different recipes and different taste tests: they'll produce different dishes even from the same pantry. Models also get extra tuning after training, where humans rate answers and the model is nudged toward responses people prefer. Different companies hire different raters with different preferences, so one model might be nudged toward caution and hedging, another toward confidence and brevity. Finally, even the same model can give different answers on different days, because many models add a small amount of randomness on purpose so answers don't feel robotic and repetitive. Example: ask three chatbots for medical advice and you might get one cautious refusal, one confident answer, and one that suggests seeing a doctor first, all from the same question.

For you: Knowing that answers vary by design, not by one model being simply 'smarter,' helps you treat any single AI answer as one opinion worth double-checking.

Original explainer

OpenAI turns ChatGPT into an advertising platform

OpenAI announced it's building advertising tools directly into its products, including 'Sponsored Agents' and integrations with e-commerce platforms like Shopify and marketing software like HubSpot. The pitch is letting businesses reach customers through AI conversations, not just search ads or social feeds. This marks a shift for a company that built its reputation on a clean, ad-free chatbot. Advertising is famously how free internet products eventually pay for themselves, and running AI models is extremely expensive. The open question is whether ads inside a conversation you trust for advice feel different, and more uncomfortable, than a banner ad on a webpage.

Follow the money: Once ChatGPT has a financial incentive to recommend certain products, its advice stops being just helpful and starts being a business decision.

via OpenAI

Meta bundles AI features into new paid subscription plans

Meta launched 'Meta One,' a set of subscription bundles that combine its existing apps like Instagram extras with additional AI usage, following the recent launch of its all-purpose AI assistant, Muse. Plans come in different tiers for individual users, creators, and businesses. This is part of a broader pattern: AI companies that once offered free chatbots are steadily moving premium features, like more messages or better models, behind paywalls. For everyday users, it means the free version of these tools will likely keep getting more limited over time, nudging people toward paid tiers to get full functionality.

Follow the money: Free AI tools were never going to stay free forever, this is what the bill for building and running them looks like starting to arrive.

via The Verge AI

A doorbell camera now lets a human guard watch your front door live

SimpliSafe's new $199 video doorbell pairs AI motion detection with a live human security agent who can watch your front door in real time and respond if something looks wrong, for a $49.99 monthly fee. The AI's job is to flag likely threats so a human doesn't have to watch constantly. This is a good example of AI's most common real-world role right now: not replacing a human decision-maker, but filtering out the 99 percent of nothing so a person only has to look when it matters. It's a practical, low-drama use of AI that most people will actually encounter through a home security product rather than a chatbot.

For you: This is what most consumer AI actually looks like day to day: quietly filtering noise so a human makes the final call, not a robot uprising.

via The Verge AI

Amazon brings its AI assistant Alexa+ to India with Hindi support

Amazon is rolling out early access to Alexa+, its newer AI-powered version of Alexa, to customers in India with support for Hindi alongside English. The assistant can hold more natural conversations and handle more complex requests than the original Alexa. India is one of the world's largest smartphone and smart-speaker markets, and language support has long been a barrier to AI assistants working well outside English-speaking countries. This kind of expansion, quietly building support for widely spoken languages beyond English, is one of the more overlooked but genuinely useful trends in AI right now.

For you: Most flashy AI news is English-first, but true everyday usefulness depends on unglamorous work like this: making tools work in the language people actually speak at home.

via TechCrunch AI