The Car With No Steering Wheel

Tesla has just launched the Cybercab in Austin. Two seats, no steering wheel, no pedals.

The self-driving isn't the news. Driverless cars have been circling Austin for the best part of a year. What's new is that Tesla has removed the ability to intervene. Not switched it off. Actually removed it, physically, from the cabin.

I've been thinking about that empty footwell all week, because the same removal is happening on property projects right now. It just doesn't come with a launch event in Texas.

Your Projects Already Have One

Every scheme has an override in it somewhere, and it almost never looks like one.

It is the draft round where somebody says the brick reads too warm, and they are right, and the material gets changed before anyone outside the project ever sees it. It is the person who asks whether that is honestly the view a planning officer will stand in, or just the one that composed nicely. It is the quiet half hour before something goes into a pack, when a person who knows the scheme looks at it properly.

None of that is glamorous. None of it appears on a programme. It's the least interesting part of the process to describe and the most expensive part to lose.

And it's always the first thing to go. Not through any decision to remove it. It goes because the deadline moved, or because the images arrived so quickly that checking them felt like an insult to the software.

Tesla Took Theirs Out Last

Here's the part I think is genuinely instructive, and it holds regardless of what you make of the company.

Tesla didn't start with an empty cabin. The sequence, publicly at least, ran through years of supervised driving in live traffic before anything was taken away. The wheel came out late, at the end of the evidence rather than at the beginning of the ambition.

That order is doing all the work. There's a large difference between removing a safeguard because you have established you no longer need it, and removing it because it was slowing you down.

Most adoption of AI in property is currently running that sequence backwards. I don't say that as much of a criticism. It's what happens when a technology arrives faster than anyone's ability to test it. But it's worth noticing when you're doing it, because the two look identical from the outside and only one of them is earned.

Where I Want No Override At All

I should be clear that plenty of our own work should have the wheel taken out, and I'd take it out tomorrow.

Camera matching is the obvious one. Lining a virtual camera up with a photograph, so that a building sits in a real street at the right scale and the right angle, is slow and fiddly and it matters enormously. Get it wrong and nothing else in the image can rescue it.

But there's a right answer, and you can see whether you've hit it.

That's the useful distinction, and it isn't about importance. Cleaning up a model that arrived with a hundred thousand duplicate faces. Upscaling. Removing noise from a process that used to take hours. None of that is trivial work, and all of it can be checked against something.

The tasks that are safe to automate aren't the unimportant ones. They're the ones where you can tell if it has gone wrong.

So this isn't an argument for keeping a hand on the wheel at all times. Quite a lot of the driving should be automated, and I'd rather spend the hours that frees up on the decisions that actually determine whether an image works.

Why Images Are the Awkward Case

The difficulty is that a CGI is not a record. It is a claim.

That matters because people make decisions from these images. Planning decisions, funding decisions, leasing decisions, and the reaction of a room full of residents. If an image makes a claim about material, massing, visibility or impact, somebody needs to be able to say where that claim came from and who checked it.

An image of a building that exists can be checked. Someone can stand in the street and compare. An image of a building that doesn't exist can't be checked against anything, because the thing it depicts hasn't happened yet. The only check available is a person who knows what was actually specified.

Which is exactly the check that gets removed.

So the failure mode here isn't ugly pictures. Ugly pictures are obvious and somebody fixes them. The failure mode is a claim nobody made on purpose. A material that reads warmer than the one specified. A massing that sits more politely against its neighbour than it will. A view chosen because it composed well rather than because it was the honest one.

None of those look like errors. They look like good images. The real risk was never that AI makes an image wrong. It's that it makes an image plausible enough that nobody thinks to ask who checked it.

The Question Worth Asking

I don't think the useful question is whether to use AI. Everybody is, ourselves included, and the people who say otherwise have usually stopped noticing the tools they already rely on.

The better question is the Cybercab question. Where is the override, and have we earned the right to take it out?

That's worth asking beyond imagery too. The planning note drafted quickly. The appraisal pulled from a model. The visual pack that arrived suspiciously fast and suspiciously cheaply. In each case something was automated, and in each case the question is the same: can anyone still say what the check used to be, and who was doing it?

Because if you can't name what the override was, you haven't automated a decision. You have quietly stopped making one.

If the imagery on a scheme needs to hold up to more than a first glance, it's worth knowing where the checks sit before they disappear. That's usually a short conversation and a useful one.

That's where my head was at this week.

Rich

PS: Can you imagine self driving cars in the UK? I still struggle with the idea, but I would love to see one in the flesh.