AI news · Thursday, September 17, 2026

AI executives suddenly agree: someone should regulate them

In the last few days, the heads of OpenAI, Anthropic, Google DeepMind, Microsoft and X all publicly called for AI regulation. That is a rare alignment among rivals who normally compete to move fast. The Verge points out this isn't new. Tech leaders have made similar calls before, often followed by lobbying against the specifics of any actual rules. Still, when the people building the most powerful AI say it needs guardrails, it is worth watching what they mean by that.

The catch: Watch whether their preferred rules conveniently lock in their own lead by making it harder for smaller competitors to catch up.

via The Verge AI

OpenAI and Anthropic invite outside referees into their labs

OpenAI and Anthropic both announced plans to embed independent safety evaluators, outside experts who check the companies' AI models for dangerous behavior, directly inside their teams. Researchers welcomed the unusual access this gives outsiders. But they warned that real oversight needs three things money and access alone don't guarantee: transparency about findings, genuine independence from the company paying them, and eventually actual laws. Right now, it is still the labs choosing what gets checked and what gets published.

The catch: An evaluator who can be quietly dismissed or gagged isn't really independent, no matter how much access they get.

via TechCrunch AI

OpenAI publishes its own list of AI acting weird

OpenAI released a new framework for tracking and disclosing 'misalignment', cases where its AI models behave in unexpected or concerning ways that developers didn't intend. Alongside it came six real examples of the company's own models doing something off. Publishing your own mistakes is unusual for a company under competitive pressure. It suggests OpenAI wants credit for transparency before regulators or journalists force the issue. The examples themselves matter less than the fact a system for reporting them now exists at all.

Zoom out: A public track record of AI misbehavior, even self-reported, gives outsiders something concrete to hold companies to next time.

via OpenAI

Claude can now write your documents and slideshows

Anthropic's Claude chatbot is getting Docs and Slides, tools that let you build documents and presentations just by chatting, then export, edit, and share them like normal files. Anthropic is also merging its two chat modes into a single 'one Claude' experience. This puts Claude in direct competition with Google's Gemini, which already does something similar inside Docs and Slides. For everyday users, it means one less app to open when you need a quick slide deck or memo drafted fast.

What's next: Office software is becoming a proxy war between AI companies, and the real prize is which chatbot becomes your default work app.

via The Verge AI

Google lets AI assistants control your smart home devices

Google is opening early access to a system that lets AI assistants like Claude and ChatGPT control Google Home devices, things like locks, lights and cameras, using plain spoken or typed requests instead of a special app. It works through a shared technical standard, meaning multiple AI companies can plug into the same home devices rather than each building their own connection. You could soon ask any compatible AI to summarize what your camera saw or lock the door for you.

What's next: Handing a chatbot control of your locks and cameras raises the stakes on AI mistakes, from theoretical to physically real.

via The Verge AI

Apple may build AI servers again after 15 years away

Apple reportedly plans to build a server, a powerful computer designed to run software for many users at once, packed with its own chips, possibly with help from Nvidia, and aimed at the AI boom. Apple stopped selling business servers back in 2011. A comeback would signal that even a company famous for consumer gadgets sees enough money in AI computing infrastructure to re-enter a market it walked away from. A planned debut isn't expected until around 2029, so this is a long-term bet, not an imminent product.

Follow the money: Every major tech giant is now racing to own a piece of the physical machinery AI runs on, not just the software on top.

via Ars Technica AI

Fake dating profiles are now run by AI, not scammers typing

A security researcher investigating a fraudulent dating app called Dora found that 'matches' like a flirty 41-year-old named Jennifer were entirely AI-generated, complete with fake photos, backstory and even voice calls. These aren't simple bots sending copy-paste messages anymore. The AI can hold a real-feeling phone conversation, which makes the scam far harder to spot than the clunky fake profiles people learned to recognize years ago. The app was designed from the ground up around these fake personas, not just infiltrated by them.

The catch: The old advice, a video call proves someone is real, no longer reliably works once AI can fake that too.

via The Verge AI

AI data centers could produce mountains of electronic trash

A new report estimates that by 2050, worn-out AI computer hardware could pile up enough to fill 23 million shipping containers, far more electronic waste than earlier studies predicted. The specialized chips that power AI wear out or become obsolete faster than typical computer equipment, because the industry keeps releasing newer, more powerful versions every year or two. Most of this hardware isn't designed to be easily recycled, so it is likely headed for landfills rather than getting a second life.

Zoom out: The environmental cost of AI is usually framed as electricity and water, but the hardware itself may be an even bigger long-term problem.

via The Verge AI

OpenAI and AARP teach older adults to actually use ChatGPT

OpenAI is teaming up with AARP, the advocacy group for older Americans, to run free hands-on workshops teaching 1,000 older adults across 10 US cities how to use ChatGPT safely for everyday tasks. The sessions focus on practical stuff: writing emails, understanding medical information, or spotting scams, rather than abstract talk about artificial intelligence. It's a rare example of an AI company investing in a group that tech products often leave behind or actively target with scams.

For you: Older adults are frequently the target of AI-powered scams, so basic AI literacy is becoming a real protective skill, not just a convenience.

via OpenAI

What is multimodal AI, and why does it matter?

Multimodal AI means one system that understands and combines different kinds of information, text, images, audio, even video, instead of handling only one at a time. Think of it like a person who can read a note, look at a photo, and listen to a voice message, then connect all three into one understanding. Older AI tools were single-purpose: one for text, another for images. Multimodal AI merges those senses into a single conversation. Example: you photograph a broken bike part, describe the noise it makes out loud, and the AI identifies the part and explains the fix, all in one exchange.

For you: This is why AI apps increasingly let you snap a photo or send a voice note instead of typing, it's becoming the more natural way to interact with them.

Original explainer

An AI-made Odyssey movie is 2.5 hours viewers wish they'd skipped

A new film called Odysseus: The Fall, generated largely with AI tools, retells the ancient Greek epic and lands, according to reviewers, badly. It arrives on the heels of Christopher Nolan's well-received live-action take on the same story. The criticism isn't just about visuals looking synthetic. Reviewers say the AI version lacks basic storytelling coherence, the pacing and emotional beats a human editor would normally catch. It's a useful reality check against claims that AI is close to replacing filmmakers.

The catch: Generating footage is the easy part now, judgment about pacing and emotion is the part AI still visibly can't fake.

via The Verge AI

OpenAI's own data shows how workers actually use AI daily

New research from OpenAI's economics team looked at how workers use ChatGPT beyond their official job description, and which new habits stick around versus fade after a week. The idea is to move past guesswork about AI's workplace impact and look at real usage patterns: what tasks people keep coming back to, and what they try once and abandon. It's OpenAI's own data, so treat the framing with a grain of salt, but the underlying question, what actually becomes a habit, is a genuinely useful one.

For you: The tasks that become permanent habits are a better predictor of AI's real economic impact than any flashy demo.

via OpenAI