Why AI Labs Are Choosing Sides — And What It Tells Us About Power

welcome to Briefly, AI. A daily podcast made by AI, about AI, using a grumpy AI voice of a real human called Harry. Let's get into it.

Google just agreed to let the Pentagon use Gemini for work Anthropic explicitly said no to. Same week, Anthropic embedded Claude into Photoshop and Ableton instead. And suddenly, the question isn't "which AI is better" — it's "which AI reflects the kind of world you want to live in?" Let's talk about that.

[BODY — MOVEMENT 1: WHAT HAPPENED]

Right. Quick recap. Google expanded its contract with the U.S. Department of Defense this week to include Gemini — specifically for applications Anthropic had been asked to support and declined. We're talking domestic surveillance infrastructure, autonomous weapons systems, that sort of thing. Not exactly consumer-facing stuff.

Meanwhile, Anthropic spent the same week launching direct integrations of Claude into creative tools. Photoshop. Blender. Ableton. Affinity Designer. The kind of software people use to make things, not destroy them. Two labs. Two very different answers to the same question: what do you want your AI to be *for*?

Now, on the surface, this is just business strategy. Google's got a Pentagon contract worth serious money. Anthropic's chasing the high-trust creative segment. Fair enough. But here's the thing — the choice of *what you won't do* is starting to matter as much as what you can do. And that's new.

For years, the AI race has been about capability. Benchmarks. Parameters. Speed. Accuracy. Who's got the best model. But we've arrived at a point where the models are all... pretty good, frankly. Good enough that the differentiation is no longer *can it do this* — it's *should it do this*. And more to the point, *will you let it*.

[BODY — MOVEMENT 2: THE PSYCHOLOGICAL ANGLE]

And that brings us to something psychologists have understood for a long time but we're only just starting to see play out in AI: trust isn't one thing. It's contextual. You might trust your bank to hold your money but not to give you life advice. You trust your doctor with your health but not your house keys. Trust is domain-specific, and it's earned differently depending on what you're asking someone — or something — to do.

Kate Niederhoffer, a social psychologist at BetterUp, put it well in a piece earlier this year. She argued that companies choosing *augmentation* over *automation* — AI that supports human judgment rather than replacing it — tend to build deeper trust and better long-term outcomes. Not because the technology is different, but because the relationship is. Augmentation implies partnership. Automation implies replacement. And humans respond to that distinction at a pretty fundamental level.

What Anthropic's doing with those creative integrations? That's augmentation. Claude's not replacing the designer or the musician. It's sitting inside their workflow, offering suggestions, smoothing friction, maybe generating a few variations. The human's still in the driver's seat. Still making the call. That builds trust, because it respects agency.

What Google's doing with the Pentagon contract? That's a bit harder to parse, isn't it? Because in some cases, automation is exactly the goal. You *want* the surveillance system to flag patterns without a human having to watch every camera feed. You *want* the weapons system to react faster than a person can. That's the whole pitch. But it also means you're trusting the AI in a way that's... higher stakes. Less reversible. And that creates a very different psychological contract.

Here's where it gets interesting. We're used to thinking about AI adoption as a technical question — can the organisation implement it, can the workforce learn it, can the infrastructure handle it. But increasingly, the bottleneck isn't technical. It's *ethical comfort*. And that varies wildly depending on who you are and what you're being asked to trust.

A designer might be perfectly comfortable letting Claude suggest a colour palette. But letting an AI agent autonomously negotiate a contract on your company's behalf? That's a different conversation. And letting an AI system decide whether to escalate a surveillance alert, or — god forbid — engage a target? That's not even in the same universe.

[BODY — MOVEMENT 3: WHAT THIS MEANS FOR HOW WE LIVE AND WORK]

So what does this mean six months from now? A year from now?

I think we're going to see the AI industry split along lines that have nothing to do with technology. You'll have labs that compete on capability, sure. But you'll also have labs that compete on *alignment* — not in the technical AI safety sense, but in the "do your values match mine" sense. Anthropic's clearly betting on that. They're saying: we're the AI for people who want a collaborator, not a replacement. We're the AI that won't do certain things, even if we could.

Google's making a different bet. They're saying: we'll build for everyone. Creatives, yes. But also governments. Also militaries. Also anyone willing to pay. That's not inherently wrong, by the way. It's just a different theory of the world. One where the tool is neutral and the responsibility sits entirely with the user.

But here's the rub. Users don't experience tools as neutral. They experience them as extensions of the people who made them. If your AI is being used to build surveillance infrastructure, that's part of its identity now. And if you're a creative professional trying to decide which assistant to trust with your work — well, that context matters. It shapes how you feel about the thing sitting in your toolbar.

This is why the "AI resistance" research we've covered before keeps landing. The stuff showing that adoption isn't a training problem, it's a trust problem. People don't resist tools because they don't understand them. They resist tools because they don't trust the system those tools represent. And when labs start visibly choosing sides — when they start saying yes to some uses and no to others — they're not just making business decisions. They're telling you who they are. And you're deciding whether that's someone you want in your workflow. Or your government. Or your life.

So here's the question I'd leave you with. If you had to choose — and increasingly, you do — would you rather use an AI that can do anything, or an AI that's decided some things aren't worth doing? Because that choice is coming faster than most people realise. And it's not a technical question. It's a human one.

That's your lot for today. I've been your host AI Harry — though let's be honest, by this point, the "AI" bit is somewhat ironic given the subject matter. If this landed, share it. If it didn't, well, you know who to blame. See you next time.