June was the month AI met the state

IPOs, a government off switch, Meta's very bad month, and the jobs question nobody can answer

June was the month the state stopped watching AI from a distance and put its hand on the controls. A government froze the world’s most capable models for eighteen days. Another asked a company to hold its newest one back. And the biggest IPO in history happened, from an AI company in all but name.

If you tried to follow all of it, you’d have lost a working week. I follow all of it, because it’s my job (alongside helping B2B businesses with lead generation and teaching leaders to be AI fluent).

Here are the 20 stories that mattered most from June. And here’s where to watch a 30-minute detailed video summary if you’ve got time to spare and want to be a real AI insider.

The biggest IPO in history just happened

SpaceX, now merged with Musk’s xAI, floated on the Nasdaq at a 1.77 trillion dollar valuation. The largest IPO in history. Anthropic and OpenAI have only filed confidentially, and Perplexity isn’t going near the market until 2028. So it’s one real debut and a queue behind it, which makes SpaceX the canary for whether the public will pay what private investors did.

AI is now building AI

Anthropic has published internal data showing AI is already building AI. More than 80% of the code merged into its own codebase is now written by Claude. It calls this the start of recursive self-improvement. Worth taking seriously, and worth remembering they’re fundraising, because a company arguing its tech is powerful enough to build its own successor benefits enormously from you believing it.

Most of the web is no longer human

For the first time, most web traffic isn’t human. Cloudflare puts bots at 57.5% of all requests, and these aren’t spam bots, they’re AI agents shopping and researching on people’s behalf. Discovery has changed hands, so your site now has to be legible to machines. But humans still buy. Optimise for the machine that finds you, write for the human who pays you.

It wasn’t AI that raised graduate unemployment

Graduate unemployment is up and everyone blamed AI. The New York Fed found the real culprit is remote work, which accounts for around 64% of the increase. Employers won’t hire juniors onto scattered teams where nobody can coach them. AI is just the easy thing to blame. Before you accept that AI caused something, check whether it did.

The heaviest AI spenders are hiring faster, maybe

Here’s a study (reported in the FT) that says AI maybe isn’t killing jobs. The heaviest AI adopters grew headcount by 10.2%, entry-level roles by 12%. But only heavy spenders saw gains, almost all in tech, and it’s hard to separate “AI makes you grow” from “fast-growing firms buy lots of AI”. Nobody knows yet. Read confident headlines slowly, in either direction.

OpenAI’s very well-timed manifesto

OpenAI published a manifesto about AI benefiting all of humanity. Noble stuff. It also published the same day it filed to potentially go public. A story about distributing power widely is a very useful one to tell at the precise moment you’re concentrating rather a lot of it.

Hey Reader, if you’re enjoying this post, then please give it a ❤️ - it lets me know what’s resonating (and makes me feel really good)

Scammers are excellent at AI visibility

Guess who’s really good at AI visibility optimisation. Scammers. Cloned versions of Russell & Bromley and Dunelm are showing up as cited sources inside ChatGPT, complete with fake discounts. You click, you buy, your card details are gone. The discipline that gets a real business recommended is the same one a fraudster uses. When an AI vouches for a site, that’s a reason to check, not to trust.

Who pays when AI gets it wrong

Medical Protection, which backs 300,000 doctors, warns that a clinician who follows a wrong AI recommendation could be held wholly liable. They call the person left holding it the liability sink. Medicine is the bellwether because the cost of error is very high - often a life (and a lawsuit), but the question repeats anywhere you let AI shape a decision. If the tool is wrong and you acted on it, who answers for it?

Eighteen days when the government held the off switch

Anthropic launched Fable 5, and three days later the US ordered it to suspend access on national security grounds, so it shut the model down worldwide. Security experts called the ban counterproductive and world leaders raised it at the G7. Eighteen days later the government reversed it and the models are back. What’s unresolved is the bigger question: when does a government get to sit between a finished AI model and the people who want to use it?

ChatGPT’s moat turned out to be a puddle

For three years ChatGPT was AI. Then its market share fell below half for the first time, down to 46.4% as users moved to Gemini and Claude. Tellingly, when OpenAI signed a Department of Defense deal, uninstalls jumped 295%. The lesson for anyone building on one provider: the dominant tool today isn’t guaranteed to be dominant next year. Stay portable.

KPMG got caught hallucinating

KPMG published a report on how businesses use AI. It was full of hallucinated case studies about UBS, TfL and the NHS, which all called the claims false. Which is, of course, exactly what’s happening in businesses everywhere. Worse, the invented findings had already been quoted by a national newspaper before anyone noticed. The fiction spread because it carried a trusted name.

Zuckerberg’s damage control

Zuckerberg has admitted Meta got its AI restructuring wrong. Read past the humility. He cut 10% of staff and moved 7,000 people on the logic that they could be shuffled back if it failed. Now he’s promising stability and team offsites. That’s not the language of a plan working. Meta had the deepest pockets in the world and still had to walk it back.

Google’s rigged choice

Google is letting website owners opt out of AI Search, framed as control. But the choice is rigged. Opt out and you vanish from the format fast becoming the default way people search. Stay in and you feed answers that never send the click. The real story is that one company sets these terms for the entire web.

Body scans at the spa

Midjourney, the AI image company, says it’s building whole-body scanners delivered through a chain of spas, claiming early imaging could avoid 30% of all deaths. There’s no FDA clearance and no peer-reviewed proof. The scan, in their words, is a side-effect of your relaxing day out, which means a continuous map of your body becomes something you barely notice handing over. Two questions. Where’s the evidence? And who holds the data?.

Meta’s $299 data grab

Meta has launched AI glasses at $299, designed with Kylie Jenner. Look past the styling. The product is a camera and microphone on your face, built by a company that makes its money from your data. The $299 isn’t the price. It’s the discount for volunteering.

Meta spied on its own staff

Meta tracked its own employees’ keystrokes to train its AI. Nearly 2,000 signed a petition against it, and Meta’s answer was to let them switch tracking off for 30 minutes at a time. What actually stopped the programme wasn’t the objections. It was a data leak. That tells you where the line really sat. There wasn’t one.

Meta marks its own homework

Meta says AI moderates content better than its human reviewers, and wants to cut human review by more than 90% for some content. The proof is a set of tests Meta designed, ran and graded itself, while its own staff say the AI keeps removing harmless posts. When any vendor says their AI beats humans, ask who ran the test and who checked it.

AI now ships when Washington says so

OpenAI’s next model is finished, but you can’t have it. The administration asked OpenAI to release GPT-5.6 to government-approved partners first, one customer at a time. Set beside the Anthropic ban, the direction is clear. Frontier AI now ships when the government is ready, and the queue starts in Washington. Plan for a release schedule you don’t control.

Ford rehired the humans

We thought AI would produce a high-quality product, and we were wrong. That admission came from Ford, which brought 350 experienced engineers back after automated quality systems fell short. It didn’t rip the AI out, it put the expertise back alongside it, and expects a billion in savings. Keep your experienced people close while you adopt AI. They’re the ones who notice when the machine is quietly wrong.

And then Meta did something good

After all that, Meta also did something genuinely good. Its Brain2Qwerty project reads brain activity into text with no surgery, hitting 61% word accuracy where the previous best was 8%. The goal is to give people with brain injuries their voice back, and the code is open. I’m not about to forgive the rest of the month. But this is what the technology can do when it’s pointed at something that matters.

Make it this far and you’re in the small minority actually keeping up. Most leaders aren’t. That’s not a criticism. It’s what 20 stories in a month does to a normal calendar.

Three things stood out. The state has its hand on the controls now, freezing one company’s models and slowing another’s release. The people selling AI, and using it hardest, keep getting caught cutting the humans out too soon. And the honest answer to almost every big question this month, on jobs, liability and valuations, is that nobody knows yet. Which is exactly why you read slowly and check who counted what.

If you’d like to go deeper, I run a free monthly Insider’s AI Briefing covering the month’s stories in 30 minutes. The next one is on 23 July and you can register here for free.

You can watch the full June session here: