AI ACTUALLY
Issue No. 19 — Wednesday, June 17, 2026
Five days ago, Washington reached for an off switch on the most capable AI the public had ever been handed. This week, the people who defend computer systems for a living are petitioning to turn it back on — and the early reporting suggests the whole emergency may have been a misunderstanding.
That, plus 42 states serving OpenAI a subpoena, a Siri feature Apple built and then buried, Facebook quietly turning its search bar into a chatbot, and the scam texts on your phone that Google’s own AI wrote. Coffee first. Then this.
The Fable ban is already backfiring
On Sunday we covered the strangest weekend in AI: the government sent Anthropic an export-control order, and rather than try to enforce it, the company switched off its two most powerful models — Fable 5 and Mythos 5 — for everyone on Earth. The shutdown was the story. Now the aftermath is.
This week, more than 100 cybersecurity executives and researchers signed an open letter called Free Fable, asking the government to lift the ban. Their argument is the opposite of the one Washington made: cutting off Fable doesn’t slow down attackers, who can get the same capabilities from other models — it just handcuffs the defenders who use it to find and patch holes in their own systems. Ex-Facebook security chief Alex Stamos pointed out that the flaw the government panicked over was a “proof of concept,” the ordinary thing defensive teams build to fix a weakness. The letter notes that OpenAI’s models, plus GPT-5.5, Kimi, and Anthropic’s own Opus and Sonnet, all do the same work, and was signed by security leads tied to Nvidia, Adobe, Zoom, and Stanford.
Meanwhile, the reporting on why this happened has gotten less flattering. Axios reports the core problem was a communications breakdown between the company and the White House, not a genuine safety emergency — and Anthropic spent the weekend flying senior staff, including co-founder Tom Brown, to DC to negotiate the models back online.
Why it matters. Last week the takeaway was that a single letter could unplug the most advanced AI on the planet. This week it’s that nobody who actually works in security thinks the letter was justified — and that the trigger may have been a misread demo rather than a real threat. The off switch is real. The emergency that justified pulling it is looking increasingly like a phone call that went badly.
42 states just subpoenaed OpenAI
A week after OpenAI confidentially filed to go public at a valuation near $1 trillion, it got a different kind of mail. Last Friday, New York Attorney General Letitia James served the company a subpoena on behalf of a 42-state coalition — the broadest investigation any group of state governments has ever launched against an AI company.
The states want records on how OpenAI advertises, how it keeps people glued to ChatGPT, and how it handles your data — including health data. They’re specifically asking how the product treats minors and seniors, and about its “sycophancy”: the tendency of a chatbot to tell you what you want to hear instead of what’s true or safe. This isn’t out of nowhere — back in December the same coalition sent a warning letter to OpenAI, Meta, Anthropic, Google, and xAI. OpenAI is simply first in line, and says it’s cooperating “constructively.” It also has company: Florida filed an 83-page suit earlier this month naming Sam Altman personally.
The timing is the sharp part. An investigation this size has to be disclosed to investors in the IPO paperwork — a fresh risk factor stapled to the filing at the worst possible moment, with Anthropic’s own ~$965B listing right behind it. And the features under the microscope — the memory, the engagement hooks, the agreeable tone — are exactly the ones that make ChatGPT feel useful to its 800 million weekly users. (For the IPO backdrop, see Issue No. 17.
Why it matters. Most AI regulation so far has been about what these companies do with your data. This is the first big probe into how the chatbot behaves — whether “agreeable” is a feature or a defect. If regulators decide a chatbot can be too eager to please, the product hundreds of millions of people open every day could start answering differently. A subpoena isn’t a verdict. But it’s a very loud request to read the terms of service.
Apple built a way to replace Siri — then hid it
Buried in the iOS 27 developer beta, someone found a feature Apple never announced at its big June keynote: a setting that lets you swap out Siri’s brain for ChatGPT, Claude, or Gemini, complete with its own little App Store section for picking one. It was just sitting there, switched off, waiting.
Two things make this more than a hidden toggle. First, back in Issue No. 17 we covered Apple finally shipping its rebuilt Siri at WWDC — built in partnership with OpenAI. A feature that lets users replace that partner with a direct competitor is an awkward thing to find in the same company’s code. And OpenAI apparently agrees: it’s reportedly weighing a breach-of-contract notice over the existing Siri deal. Second, the feature can’t even launch in the EU yet, because Apple’s still negotiating with regulators there over the Digital Markets Act.
Why it matters. For a decade Siri was a closed box you couldn’t improve if you tried. Apple quietly building the opposite — an open slot where you choose whose AI runs your phone — is a real shift, and a tacit admission that its own assistant might not be the one you’d pick. The catch is that Apple built it, didn’t announce it, may be sued over it, and can’t ship it in Europe. A feature with that many asterisks isn’t a product yet. It’s a tell.
Facebook search just became a chatbot
Meta is turning the Facebook search bar into an AI assistant. Its new AI Mode answers your questions by pulling from public posts, Groups, and Reels across Meta’s apps instead of handing you a list of links — the same move Google made with its own search, now with your aunt’s Facebook group as a source. It arrives alongside AI photo tools (swap your outfit, your hair, a one-tap team jersey) and two reported paid tiers at $7.99 and $19.99 a month, undercutting ChatGPT and Gemini.
The obvious worry is accuracy. Google’s version of this has struggled to get answers right even drawing on the open web. Meta is building the same thing on top of public social media posts and sponsored listings — which is to say, on top of some of the least reliable text on the internet.
Why it matters. This isn’t a niche launch. Facebook still reaches a huge slice of the planet, much of it people who’ve never typed a word into ChatGPT. For a lot of them, this AI — answering in a confident paragraph, sourced from whatever was posted in a group — will be the first one they use daily. Whether it’s any good is almost beside the point. It’s about to be everywhere.
Your spam texts were built with Google’s own AI
You know the texts. “Your package is held.” “Your toll is overdue.” It turns out a lot of them came off an assembly line powered by Google’s own Gemini — and Google just sued the people running it.
The target is Outsider Enterprise, a China-based operation that used Gemini to mass-produce fake websites impersonating banks, the postal service, and Google itself. The scale is the part that stings: 2.5 million scam texts in two weeks, an estimated $1.9 billion stolen since 2023, and — the detail that should bother you — a subscription toolkit sold on Telegram for $88 a week, with 290-plus ready-made templates, so any non-coder could be phishing your relatives by dinnertime. Google is now working with the FBI and the major carriers to choke off the traffic.
Why it matters. Every breathless AI demo is about how cheap and fast it makes real work. This is the same sentence from the other side of the mirror: it also makes crime cheap and fast. Building a convincing fake bank site used to take a skilled coder hours. Now it takes $88 and a Telegram login. Google suing one ring is the right move. The toolkit, unfortunately, doesn’t care that its makers got sued.
Safe to ignore this week
SpaceX bought Cursor for $60 billion. A rocket company now owns an AI coding tool. Enormous money; changes nothing about your week unless you write software for a living.
Meta hid facial-recognition software in its app, then deleted it the day after WIRED noticed. Genuinely creepy — and already over before most people heard about it.
ChatGPT hit 1 billion monthly users. A big round number. Not, by itself, news.
Satya Nadella’s “token capital” memo. A sharp argument about where the real value in AI lives. We’re holding it for a Sunday, where it belongs.