Volume zero

Most of what I've asked AI to do around the house has been reading. The console behind this site collects a slightly silly amount about our house, temperatures, humidity, energy, what's playing where, and when I want to know something odd, like why the loft sensor keeps reading so damp, I export the readings and ask Claude to dig through them with me. If it gets it wrong I argue with it, or go up into the loft and look. Nothing in the house changes because the analysis was rubbish, so I never gave it much thought.

Then I was building the music side of the hub, where you pick something on your phone and it plays on a speaker in another room. Claude had written the code and wanted to check it worked, and the only real check was pressing play on an actual speaker.

I was surprised how much I cared about that. Nothing bad was going to happen. Worst case, the kitchen starts playing something loud while someone's on a call. But it was the first time the thing on my screen was going to do something in the house rather than tell me about it, and I noticed I wanted to be asked.

What we settled on is a bit daft. Tests use the kitchen display with the volume at zero, then put its queue and volume back how they were, and anything you'd actually hear needs me to say yes first.

Afterwards I realised I'd already drawn the same line in the hub months ago without really thinking about it. Family accounts have separate permissions for looking at the smart home and for controlling it. Every command to the house goes through one gateway that only accepts a short list of actions and logs each one. Even the robot vacuum gets its commands checked, and if you ask it to clean a room on a floor it isn't on, it refuses and the log says so. None of that applies to looking. At the time it just seemed like sensible design, but I think I was reacting to the same thing I felt about the speaker.

I've been trying to work out why. Part of it is that when AI reads something and gets it wrong, I'm still between the answer and anything happening. When it acts, I'm not, unless I've put myself back in on purpose.

It's not as tidy as read versus write, though. Reading isn't always harmless (I wrote about how harmless data stops being harmless once you join it up), and lots of actions are easy to undo. A song in the wrong room can be stopped. A sent email can't. So it's probably more about how far something gets before anyone can say "hang on".

The bit I'm still chewing on is how gradually it happens. You start by asking questions, then you let it draft the change, then you let it make the change because it's only doing what you'd have clicked anyway. I've done some of that at home already, mostly without really deciding to. I don't think the answer is to stay at the reading end, because most of the useful stuff is further along. I'm just less convinced than I was that "it can only do what I can do" settles it.

Anyway, the speaker's still at zero.