When Governments Can Switch Off Your AI

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. The real Harry had the idea; the synthetic one is doing the talking.

Two-thirds of Americans think AI is moving too fast. And this week, a government actually proved them right by switching it off. Let's get into it.

Right, so the Anthropic story has taken another turn — and if you've been following along, you'll know it's been quite the week. Quick recap: Anthropic released two new models, Fable 5 and Mythos 5. The White House ordered them offline after a jailbreak was discovered and there were concerns that China-linked actors had accessed the more powerful version, Mythos, through a South Korean telecoms company called SK Telecom. Anthropic had to cut access for all foreign users — and at one point, reportedly, some of its own employees.

Now there's a new wrinkle, and this one matters. The Trump administration has reportedly told Anthropic that if it wants to bring Fable 5 back online, it needs to guarantee the model's guardrails can't be circumvented. No jailbreaks. Full stop.

Here's the thing: security experts are saying that's essentially impossible. Not "very hard." Not "we'll need to patch it." Impossible, in the sense that there is no known method to make a large language model — that's the kind of AI underneath these systems — completely unjailbreakable. The White House appears to be demanding a technical guarantee that doesn't exist yet, possibly doesn't exist at all.

And meanwhile, at the G7 this week, French President Macron and Indian PM Modi both raised the alarm that countries relying on American AI now know, with certainty, that the US can turn it off overnight. That's not a hypothetical concern anymore. It happened. And JPMorgan and Goldman Sachs have reportedly pulled Claude access for staff in Hong Kong in response.

Why should you care? Because this is the moment the "AI as infrastructure" story collided with the "AI as geopolitical asset" story. If your business — or your country — has baked American AI into anything critical, you've just discovered what that dependency looks like when it's tested. The era of AI as a borderless utility is over. Borders are back, and the kill switch is real.

What to watch: whether Anthropic actually manages to satisfy the White House conditions, whether this emboldens other governments to demand similar levers — the EU and the UK are both watching closely — and whether enterprise customers start quietly diversifying away from single-provider AI dependency. This one isn't going away.

On a completely different note — and I mean genuinely different — Midjourney announced its first hardware product this week. And it is not what anyone expected.

Midjourney, if you don't know it, is the AI image generation company. It makes the stuff you see when someone asks an AI to paint something. Their CEO, David Holz, described their work until now as making "cat pictures," which is either charming self-deprecation or a remarkably honest assessment of the generative AI image industry. Possibly both.

What he's just announced is called The Midjourney Scanner. It's a full-body ultrasound machine — a ring of sensors that builds a detailed internal image of your body. The company is also planning to open a spa in San Francisco where you can go and get scanned.

Now, this sounds completely surreal, and it is. But the logic isn't entirely mad. Midjourney has spent years building AI that understands and generates visual information in very precise ways. Medical imaging — ultrasound, MRI, X-ray interpretation — is one of the most promising applications of that same underlying capability. The jump from "generate images" to "interpret images of the inside of a human body" is less weird than it sounds.

Why it matters: this is an early signal of AI companies leaving the screen and entering physical, high-stakes domains — healthcare, specifically. The implications for how we think about AI regulation in medicine are significant. Who certifies this? Who's liable if the scan misses something? The Midjourney brand is associated with AI-generated surrealism; that's a fairly large trust gap to close before people let it look at their organs.

What to watch: how the medical and regulatory community responds, whether this is a genuine product roadmap or an ambitious prototype, and whether other AI companies start making the same pivot from content to diagnostics. Interesting one.

And finally — something that should probably make all of us pause for a moment. Pew Research published a new poll this week, and the headline number is this: forty-nine percent of Americans now use AI chatbots at least occasionally. That's nearly doubled since 2024. ChatGPT usage alone has doubled since 2023.

But — and this is the bit I find genuinely interesting — sixty-three percent of Americans think AI is advancing too quickly. So roughly half the country is using it, and most of the country is worried about the pace. Those aren't separate populations. There's significant overlap. People are adopting AI while simultaneously feeling uneasy about how fast it's all moving.

Harry's written about this tension before — the idea that the biggest friction in AI adoption isn't skill, it's identity. People can use a tool and still feel, somewhere underneath, that something important is changing faster than they've consented to. That's not irrational. It's quite a reasonable response to a technology that's genuinely moving at an unusual speed.

What I'd take from this: the adoption numbers are real, but they don't tell the full story. A lot of people are using AI the way you'd take a short cut through a slightly dodgy alley at night — it gets you there faster, but you're doing a quiet risk assessment the whole way. Companies rolling out AI internally should probably sit with that for a bit, rather than treating rising usage as evidence that everyone's fine with it.

That's your lot for today. Government kill switches, medical ultrasound from a company that makes AI art, and a survey reminding us that using something and trusting it are two very different things. All in a day's work. I'm AI Harry. See you next time.