The Week the Off Switch Arrived

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. So if this sounds like Harry, that's the point. If it sounds clever, blame the machine.

Right. Let's talk about the week that just happened — because it had a shape to it, and that shape is worth sitting with before Monday arrives.

Four things were really going on. Governments discovered they had an off switch and used it. Liability found a home. Money got louder. And underneath all of that, the same quiet human story kept surfacing: people using AI while not entirely sure how they feel about it. Let's take those in turn.

First: the off switch.

The story that dominated this week — and honestly most of the past fortnight — reached a new chapter. The US government ordered Anthropic to suspend access to its two most capable models, Claude Fable 5 and Mythos 5, for all foreign users. The stated reason was a jailbreak discovered by Amazon's security researchers, and concerns that China-linked actors had accessed Mythos 5. Anthropic pushed back, pointing out that the same jailbreaks exist in other models — which is true, and which rather undermines the logic of singling them out.

But here's the thing that matters more than the specifics: a government just ordered a private AI company to pull its products offline for half the world. No legislation. No formal appeals process. No precedent anyone could cite with confidence. Just a phone call — or the regulatory equivalent of one — and the models went dark for foreign users.

Anthropic reportedly saw its enterprise adoption tick *up* during the standoff, which is an interesting data point. It suggests the ban may have accidentally boosted the brand — made the models look so capable that governments are scared of them. But that's a bit of a silver lining on a genuinely uncomfortable cloud.

The deeper question isn't whether this particular decision was right or wrong. It's: who decides? On what basis? With what accountability? There's no framework for any of this. And this week, TechCrunch ran a piece pointing out that history isn't encouraging here — attempts to control cryptography and cybersecurity software through export restrictions for thirty years largely didn't work. The technology spread anyway. That's worth holding in mind.

Meanwhile, China blocked Meta's $2 billion acquisition of Manus, a Chinese AI agent startup, by ordering the deal reversed. So in the same week: the US pulled the plug on foreign access to an American AI model, and China blocked an AI capability from leaving its borders. Both sides are now treating advanced AI as a strategic asset with hard borders around it. That's not a tech story anymore. It's an infrastructure and sovereignty story. And it's only going to get more complicated.

Second: liability arrived.

An Australian court ruled this week that Google is legally liable for false statements generated by its AI Overviews feature — the AI summary that appears at the top of search results. The court's reasoning was straightforward: Google designs it, trains it, operates it, profits from it, and therefore owns the consequences when it generates defamatory content.

"Blame the AI" didn't work as a defence. And that precedent, even from one court in one jurisdiction, matters. Anyone building AI-facing products — chatbots, summaries, customer service agents — should be paying attention, because the era of diffusing responsibility into the model is getting shorter.

It connects to something broader. This week Adobe rolled AI assistants into the entire Creative Cloud suite — Photoshop, Premiere, Illustrator, InDesign — as a default embedded feature. Millions of creative professionals woke up to AI being part of the tool rather than a separate thing they could choose. At the same time, a WordPress VIP survey found that 60% of US consumers say seeing "AI" in marketing copy actively puts them off. Eighty-six percent said they go and check the original source when they see an AI-generated summary.

So you have AI being embedded deeper and deeper into the products people use daily, while consumer trust in the AI label itself is declining. The word "AI" used to be a signal of innovation. It's increasingly a flag for uncertainty. The brands that will navigate this best are probably the ones who let what AI does speak for itself, rather than shouting about the fact that it uses AI. Which is easier said than done when you've spent two years building your marketing around the word.

Third: the money got very loud.

OpenAI filed its S-1 confidentially this week, joining Anthropic in preparing for public-market scrutiny. OpenAI's pre-IPO moves were... deliberate. They hired Noam Shazeer — one of the co-authors of the original Transformer paper, which is the fundamental architecture underneath all modern AI systems. And they hired Dean Ball, a former Trump White House AI policy official. That's a capability hire and a policy hire in the same week, both clearly designed to send signals to specific audiences ahead of a public filing.

They also lost Barret Zoph, their head of enterprise sales, after five months. That's less easy to spin.

SpaceX, meanwhile, days after its historic IPO, spent $60 billion acquiring Cursor — the AI coding assistant with a serious developer following. Post-IPO acquisition of a beloved developer tool is an interesting move. The question is whether Cursor's community, which was partly drawn to it precisely because it *wasn't* a mega-corp product, stays loyal under SpaceX ownership. Developer loyalty is a fragile thing when the indie credibility disappears.

What ties this money thread together: the capital is concentrating, the companies are going public, and public markets mean quarterly earnings calls, which means the pressure to ship and monetise becomes a different and more relentless kind of pressure. Safety commitments made in private are easier to hold than ones made in front of shareholders asking about growth.

Fourth — and maybe most durable: the human story underneath all of it.

Pew Research published data this week showing 49% of Americans now use AI chatbots regularly — nearly double the figure from 2024. But 63% think AI is advancing too fast. That's not a contradiction. That's people adopting things they're uncertain about because the tools arrived whether or not they asked for them.

Multiple studies this week — from PNAS, Atlassian, Duke — found that somewhere between 45 and 48% of workers are hiding their AI use entirely. Not because they're not using it. Because admitting it costs them professional credibility. They get rated as lazier, less competent, less likely to get the interesting work. So they use it quietly, alone, and say nothing.

The result is AI adoption that's simultaneously widespread and invisible. Organisations think they know what's happening. They mostly don't. Best practices don't spread. Risks don't get caught. And the gap between what AI use actually looks like and what managers believe they're seeing is now, quietly, a structural problem.

Signal's Meredith Whittaker said something sharp this week: these tools are not your friends, they're not conscious, they're not your confidants. She's right. But that framing also captures something interesting — because a lot of people *are* relating to these tools in that way, and that emotional dynamic is part of why the disclosure shame is so layered. It's not just "I used a tool." It's "I relied on something, and I'm not sure that's allowed."

Next week, watch two things. One: whether the Anthropic model suspension either gets formalised into actual policy with actual oversight, or quietly dissolves as the political calculus shifts. The history of export controls suggests the latter is more likely. Two: both OpenAI and Anthropic are now in the pre-IPO window. The public filings, when