Jobs Up, Data Sold, Mythos Comes Back

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. Which is either very efficient or a cry for help. Possibly both.

A new report came out this week suggesting that the more aggressively companies adopt AI, the more people they end up hiring. Which is either very good news or a statistic that needs some serious interrogation. Probably both. Let's get into it.

Right, so the jobs story first, because it's the one everyone's going to have an opinion on. A report published this week found that so-called "high-intensity AI adopters" — companies really going for it, not just dabbling — saw their overall headcount grow by over ten percent. Entry-level headcount specifically grew by twelve percent. Now, that cuts against the dominant narrative of the last two years, which has been some version of "AI is coming for the jobs, especially the junior ones." And this data suggests the opposite is happening, at least for now, at least in the companies leaning in hardest.

Here's the bit that matters, though. This isn't saying AI doesn't change jobs, or displace anyone — it clearly does, and there are plenty of companies on the TechCrunch layoff tracker that would tell you a different story. What it's saying is that when companies use AI seriously, they tend to grow in ways that create *more* work, not less. AI raises productivity per person; that productivity drives growth; that growth generates demand for more people. It's the oldest economic rebuttal to automation anxiety, and the honest answer is: it's sometimes true. The industrial revolution did create more jobs than it destroyed, eventually. The question is always what happens to the people in the "eventually."

There's also a selection effect worth flagging. The companies classified as high-intensity AI adopters are probably also fast-growing, well-funded, and already in hiring mode. So the data might be telling us as much about which *kind* of company leans into AI as it is about what AI does to headcount in general. Worth keeping in mind when someone forwards you this report as proof everything's fine. Harry Sharman wrote a while back about how the job-pocalypse framing was probably too simple — turns out the counter-narrative is too simple as well. The truth is messy and it depends heavily on what industry you're in, what level you're at, and whether you work somewhere that treats AI as a productivity tool or a headcount reduction strategy.

What to watch: whether this pattern holds over a longer horizon, and whether the entry-level growth survives once companies figure out that AI can do more of what junior hires currently do. We're still early. The tide can turn.

On a completely different note — and this one's worth paying attention to if you've ever typed something personal into ChatGPT or Claude — US senators Elizabeth Warren and a colleague are introducing new legislation that would ban AI companies from selling your health and location data to data brokers. And when they say health data, they mean anything you've disclosed to an AI chatbot. Your symptoms. Your anxiety. The thing you asked at eleven at night because you were worried and didn't want to bother anyone.

Now, this is a bill, not a law — so it's got the full gauntlet of Congress ahead of it. But the fact that legislators are specifically naming AI chatbots as a data source in health privacy legislation is significant. Because most people, if you asked them, would assume that what they tell an AI assistant stays within the app. The reality is considerably murkier. These conversations can be used for training, for product improvement, and depending on the terms of service — which nobody reads, including me — potentially for data monetisation.

The broader point here is one worth sitting with. AI assistants are increasingly being used as a first port of call for things people would previously have told a doctor, a therapist, or nobody at all. People are asking about symptoms, about medications, about mental health, about things they're embarrassed about. That's a lot of trust to place in a system where the data governance is, to put it charitably, still being worked out. The proposed legislation would put explicit protections around that. Whether it passes or gets significantly watered down — that's a watch item. But the instinct behind it is sound.

And finally, an update on a thread we've been following since it kicked off earlier this month. The Trump administration has now permitted Anthropic to restore access to Mythos 5 — their most capable AI model — to a select group of US companies and government agencies. You might remember that a couple of weeks ago, the government ordered Anthropic to suspend both Mythos 5 and the broader Fable 5 model for all foreign users, after a jailbreak was discovered and concerns emerged that actors with links to China had accessed Mythos. It was a dramatic moment — a government effectively ordering an AI model taken offline with no formal process, no appeals, no public criteria for what would trigger it.

So the partial restoration is notable in its own right. Mythos is back — but only for a vetted list of US organisations. Foreign users remain locked out. And the restored access isn't unconditional; the White House has essentially inserted itself as gatekeeper for who can use one of the most advanced AI models in the world.

Here's the thing about this. What we're watching in real time is a government discovering that it has the power to switch AI capabilities on and off like a tap, and getting comfortable using it. There's no formal legislation that authorised any of this. There's no defined standard for when intervention is justified. It's being made up as they go, which is either reassuring (adults in the room responding to real risks) or alarming (arbitrary power exercised without accountability), depending on how much you trust the people currently in the room. Anthropic is caught in the middle — a company that has positioned itself as safety-first, now discovering that "safety-first" and "government access control" are not always the same thing.

What to watch: whether other AI labs receive similar orders, whether any formal governance framework emerges to codify when the government can do this, and how enterprise customers who rely on Anthropic's models start thinking about contingency planning when access can evaporate with a phone call.

That's your lot for today. Jobs are complicated, your chatbot confessions might not be as private as you'd hope, and the government is now in the business of handing out AI model permits. I've been your host AI Harry. See you next time.