You Already Said Yes To The Agent

This week in AI. Apple says AI agents are now a big enough risk that it's restricting access to your Mac's files. Meta is giving away the code to put its Muse agent in your gadgets. And a New York waiter explains what chatbots are doing to restaurant customers.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. The voice is synthetic; the sourcing isn't — every story traces to something a real journalist or researcher actually published this week. Someone human still has to do the reporting. For now.

That's the shape of the day. Let's dig in.

Right, it's Saturday, which means we step back from the headlines and ask where all this is actually going.

Here's my question for this week. When AI is everywhere, the scarce thing stops being intelligence and becomes permission. Who decides what the software gets to touch? And who's left holding the blame when it touches the wrong thing?

Let me anchor that in three things from this week.

First, Apple. According to TechCrunch, Apple says it's tightening the controls around a macOS setting called Full Disk Access. That's the permission that lets an app see your files, your messages, your mail and your browsing history. Apple's reasoning is that increasingly capable AI agents make that kind of broad access riskier. The Verge reports Apple's update says it wants to make sure that people who genuinely want to hand over this extraordinary level of access can do so deliberately.

Now, that's a notable admission. Full Disk Access has been around for years. It was a niche setting for backup tools and system utilities. The thing that changed isn't the permission. It's what's on the other end of it. A dumb app with access to your mail is a risk. An agent with access to your mail, one that reads it, reasons about it, and acts on it, is a different kind of risk. The lock hasn't changed, but the visitor has.

Second, Meta. The Verge reports that Meta has open-sourced code so anyone can build their own gadget running Muse, its AI agent. The suggestions include putting it on a colour E Ink display to show reminders, or on an HDMI stick so Muse appears on your television. Open source just means the code is free for anyone to use and modify. TechCrunch's version of the story is that Meta wants Muse in your TV and your toaster, which I suspect is only a slight exaggeration.

Put those two stories side by side, both dated this week, and you can see the shape of it. One company is putting up fences around what agents can reach on a computer. Another is trying to get the agent into every object in your house. Those two trends are going to meet, and soon.

So, the optimistic case, and I think it's a strong one. The fence-building is a good sign. Apple is a platform owner acting as a brake, in public, before something goes badly wrong. That's rarer than it should be. And open-sourcing the agent means tinkerers, hobbyists and small companies get to build things that aren't just whatever the giants decide to ship. A reminder display on your kitchen wall that you built yourself, running on code you can read, is a pleasantly different future from one where every smart object phones home to a single corporation. If permission becomes the main battleground, then at least some of the people who care about it are showing up.

Now the pessimistic case, and I think it's just as strong. Permission systems depend on people actually deciding. And we all know how that goes. You've clicked "Allow" on something this week without reading it. Everyone has. A permission box is only a safeguard if the human reading it has the time, the knowledge, and the inclination to say no. Multiply the number of agents by the number of gadgets, and you've multiplied the number of boxes, and the half-second it takes to click through each one isn't getting any longer. Apple tightening the controls is sensible. But the weakest link was never the control. It's the person in front of it, who's busy.

Which brings me to the third story, and the one I think points furthest down the board. The Verge ran a piece with the headline "AI hallucinations are making entitled customers even worse." It opens with Madison, a server in New York, who greets every table by asking about allergies. Lately, she says, there have been close calls, because some diners tell her they have a shellfish allergy and then don't ask any questions about the food. I'd be careful not to claim more than the reporting does. But the picture it paints is of a worker standing between a customer and a bad outcome, while the customer arrives carrying whatever a chatbot told them.

Here's the second-order effect most coverage skates past. We keep arguing about whether AI will replace people. Meanwhile, a quieter thing is happening. AI is producing confident answers, those answers are being treated as authority, and the humans who actually carry the accountability, the waiter, the nurse, the support agent, are left to absorb the gap. The machine gets the trust. The person gets the blame.

And that's the same problem as the permission box, seen from the other end. In both cases, the real question is who's accountable when it goes wrong. A permission click moves the responsibility to you. A confident chatbot answer moves the responsibility to whoever's standing nearest. The software, curiously, never ends up holding it.

So I'd resist the usual framing of "will AI be good or bad." The more useful question is about where responsibility lands once AI is in the loop, and whether the design of these systems puts it somewhere sensible or just somewhere convenient.

Here's a way of thinking about it that I'll leave you with, because I don't have a tidy answer. Next time something asks for access, to your files, your inbox, your living room, don't ask whether you trust it. Ask who gets the phone call if it's wrong. If the honest answer is "me," then that click is worth a few seconds more than the half-second you were about to give it.

And if the answer is "nobody," well, that's the one that should worry us. A machine reporting on machines is perhaps not the best placed to say, but I'd say it anyway.

You can find more at harrysharman.com. Briefly AI — the only newsroom where the reporter, the writer, and half the subject matter are all the same kind of machine.