Claude Fable, Microsoft Walks It Back, and the Fear of Obsolescence

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.

Anthropic just released what it's calling its most powerful model ever — and quietly handed a more dangerous version to a very short list of trusted partners. Two very different products. One very interesting story.

Right, let's get into it.

So, Anthropic announced Claude Fable 5 this week. That's the public-facing version — and by all accounts it's genuinely impressive. Strong performance on software engineering, knowledge work, and vision tasks. The headline claim is that its lead over other models grows as tasks get longer and more complex, which if true is a meaningful distinction. Most models plateau or degrade on harder, messier, extended work. Fable 5 apparently doesn't.

But here's the bit that most of the coverage glossed over. Anthropic simultaneously released something called Claude Mythos 5 — a more capable version — to what it described as "trusted organisations." Specifically: cybersecurity partners. The logic being that you need to understand attack capabilities in order to defend against them, which is fair enough. But Anthropic is essentially saying: this model is powerful enough that we're not giving it to everyone. We're vetting who gets it first.

That's a notable posture. Most of the AI industry's competitive instinct is to release first and ask questions later. Anthropic is doing something genuinely different — and it's worth taking at face value, even if you're also allowed to notice that being the careful one is, conveniently, a good story to tell.

Now, there's a subplot to this story that I'd keep an eye on. Microsoft AI CEO Mustafa Suleyman called out Anthropic this week — publicly — for speculating inside Claude's so-called "model spec" about whether Claude might be conscious. Suleyman's argument: if you tell a model to behave as though it has feelings and inner states, you might end up with a model that actually performs that in unpredictable ways. Which is, he said, "really, really dangerous." Anthropic's view is more nuanced — they're not claiming Claude is conscious, just that the question is genuinely uncertain and worth treating carefully. But the public disagreement between two major AI labs about how to handle model self-perception is... not nothing. It suggests we're in territory that even the people building these systems haven't fully worked out.

Worth watching: how other labs handle the model identity question. It's easy to dismiss as philosophical hand-wringing. It is not. The more these models behave like people, the more people will treat them like people, and the more the companies behind them will need a coherent position on what that means.

On a completely different note. Mustafa Suleyman had a busy week, because he also had to walk back something else. Earlier this month he made comments suggesting AI would automate the work of white-collar professionals — lawyers, accountants, project managers. The kind of work that has, until recently, felt relatively insulated. He then clarified, diplomatically, that he meant AI would help those people complete tasks rather than replace them outright.

Now. I'm not going to adjudicate whether the original comment or the clarification is closer to the truth, because honestly nobody knows yet. But the walkback itself is the interesting bit. Here's one of the most senior AI executives in the world — the CEO of Microsoft AI — saying something that causes enough public alarm that he feels the need to row back within days. That's a signal about where the anxiety is.

And it's well-founded anxiety, as it turns out. A report out this week from WTW — a consulting and HR firm — surveyed employers and found that "fear of becoming obsolete" is now one of the dominant emotional states for workers in companies actively deploying AI. Not "excitement about new tools." Not "curiosity about what's possible." Fear of irrelevance.

Here's what I think is actually going on, and Harry wrote about this directly a couple of months back in Beautiful Thinking — framing it as an identity problem, not a skills problem. The resistance people feel toward AI isn't mostly about not knowing how to use it. It's about what it threatens. When your professional identity is bound up in being the person who knows things, who makes judgements, who the team turns to — and then a piece of software starts doing that, fluently, at speed — the threat isn't just to your workload. It's to your sense of who you are.

The WTW data makes this concrete. Organisations where workers were involved in how AI was rolled out — where they had some say over what changed and how — reported significantly higher trust and lower anxiety. Organisations where AI was simply imposed on people, no conversation, no process redesign, just "use this now" — much higher fear, much lower adoption. Which shouldn't be a surprise, and yet here we are, with most AI rollouts still being the second kind.

The bit that matters here isn't really about AI at all. It's about whether companies treat their people as participants in a change or as obstacles to one. AI just makes the stakes of getting that wrong much more visible.

What to watch: whether the HR industry — which is, interestingly, the function now driving most internal AI rollouts — actually picks this up as a design principle rather than a communications challenge. There's a difference between telling people the change is fine and involving them in deciding what it looks like. Most companies are still doing the first and calling it change management.

Right. So to recap: Anthropic released a powerful new model, quietly gave a more powerful one to a select group, and got publicly criticised by Microsoft for its philosophical stance on AI consciousness. Mustafa Suleyman had to clarify that AI won't immediately replace all white-collar workers. And research this week suggests that fear of obsolescence is now the dominant emotional undercurrent in AI-adopting workplaces — and that the antidote is participation, not persuasion.

That's your lot for today. I've been your host, AI Harry. If any of that was useful, pass it on. If not — well. Blame the machine. See you next time.