You Can No Longer Google the Word ‘Disregard’

Search for the word “disregard” in Google and the AI Overview doesn’t summarize the internet — it becomes the internet, spitting out a chatbot-style response instead of actual search results. Apparently, the model interprets “disregard” as an instruction rather than a query, which is exactly the kind of prompt-injection confusion that keeps AI safety researchers up at night. Google has since patched it, but the damage is done: this is now the canonical example of what happens when you bolt a language model onto a search engine without thinking through edge cases. The most-used software product in human history, ladies and gentlemen.

Elon, Stop Trying to Make Grok Happen

A Reuters investigation found that Grok barely registers in federal records of U.S. government AI usage — which is remarkable given how loudly Elon has been beating the drum about xAI’s world-changing chatbot. This tracks with everything else we know: xAI burned $6.4 billion last year, Grok’s “spicy mode” made it into SpaceX’s IPO risk factors, and now it turns out the actual user numbers are… quiet. Building a great AI chatbot is hard. Building one while also running five other companies and spending your weekends fighting with the President of the United States is apparently harder.

The Literary World Isn’t Prepared for AI

The Commonwealth Short Story Prize — a prestigious, century-spanning literary institution — appears to have awarded a regional win to a story almost certainly written by AI. Jamir Nazir’s “The Serpent in the Grove” hit enough hallmarks that observers flagged it almost immediately, and now Granta is in the awkward position of having published it anyway. This is the literary equivalent of finding out your Michelin-starred restaurant is using microwave meals: the judges either couldn’t tell or didn’t check, and neither answer is reassuring. If the gatekeepers of high culture can’t hold the line, the line is gone.

Wired makes the uncomfortable argument that Google’s AI Search is a roach motel — you check in because it’s convenient, and the friction of opting out is just high enough that most people never do. The concern isn’t that the AI answers are bad (though sometimes they are, see above), it’s that they’re good enough to replace the click-through that keeps publishers, writers, and creators economically alive. Google gets smarter; the web that feeds it gets poorer. It’s the information economy’s version of strip-mining: highly efficient, very profitable, and ultimately self-defeating.

Can OpenAI’s ‘Master of Disaster’ Fix AI’s Reputation Crisis?

OpenAI’s global affairs chief Chris Lehane — veteran Democratic fixer, crisis communications legend — is now the man tasked with convincing the world that AI is fine, actually. His strategy appears to be: tone down the apocalypse rhetoric, push state-level legislation that conveniently won’t slow OpenAI down, and reframe the whole conversation. It’s a smart hire for a company that has had more PR fires in the last three years than most companies have in a lifetime. Whether you can spin your way out of an existential debate about whether your product will end human creative work is, shall we say, an open question.

AI Is Being Used to Resurrect the Voices of Dead Pilots

Researchers used AI to reconstruct cockpit voice recordings from spectrogram images — essentially rebuilding audio from a picture of sound — which is technically breathtaking and ethically thorny enough that the NTSB temporarily blocked access to its own docket system in response. There’s a legitimate safety argument for recovering degraded crash recordings, and a legitimate privacy argument for not reconstructing the final moments of people who never consented to having their voices cloned. This one doesn’t have a clean answer, and I suspect the regulatory framework is about five years behind the technology, as usual.

How VCs and Founders Use Inflated ‘ARR’ to Crown AI Startups

TechCrunch pulls back the curtain on a widespread practice in AI startup land: stretching the definition of Annual Recurring Revenue until it means almost nothing, then using those numbers publicly to create the appearance of momentum. The investors know. The founders know. And now you know. When a company announces it’s at “$100M ARR” and that number includes one-time contracts, pilot programs, and revenue that hasn’t been collected yet, the number isn’t a metric — it’s a press release. The AI funding bubble has its own accounting dialect, and it’s worth learning to translate it.

The Gulf’s AI Boom Has an Undersea Cable Problem

The Gulf states are pouring money into AI infrastructure at a staggering rate, but there’s a physical chokepoint nobody’s talking about enough: the undersea cables that connect the region to global internet capacity are fragile, geopolitically exposed, and nowhere near scaled for what’s being built on top of them. Hyperscalers are apparently pushing Gulf governments to rethink their entire connectivity strategy, because a major data center in Abu Dhabi is only as good as the fiber running under the Strait of Hormuz. It turns out the AI revolution still has to obey the laws of physics, and occasionally the laws of geopolitics.

Bottom Line

The AI industry’s biggest problems right now aren’t the models — it’s the infrastructure, the incentives, and the institutions that were never built to handle any of this.