The Last Nine Per Cent

This week I made an artificial version of myself.

I used a tool called HeyGen, fed it a script I'd written about how we use AI, and let it produce a video of me delivering it. The script was right. I'd read it back in my own head and it sounded exactly like something I'd say.

The result sort of looks like me, and it sort of sounds like me. It is also completely wrong.

The mannerisms aren't mine. The delivery lands in the wrong places. It pauses where I wouldn't and pushes on where I'd stop. Nobody who has met me would be fooled for a second.

The jokes suffer worst. There are one or two in the script, and in my own voice they land somewhere between mild amusement and polite silence, which I have long since made peace with. Delivered by the artificial version of me they don't land at all. There's nothing to react to. A joke is mostly intonation and a pause in the right place, and those turn out to be the first things to go.

Almost Right

Here's the thing, though. That video isn't a failure.

It took minutes. A decade ago it would have needed a crew, a day and a budget. And for plenty of jobs it would be perfectly adequate: an internal briefing, a quick piece of content where speed matters more than personality, anything where nobody needs it to actually be me.

The trouble only starts when I ask it to do something harder.

Architectural imagery works the same way. There is a version of almost right that is genuinely fine, and I'll say so when I think that's the case. An early concept testing whether a scheme should feel warmer. A sketch to start a conversation. A quick look at a massing option before anybody commits to anything.

The question was never whether AI is involved. It's what the image has to survive.

Why I Spot It and You Might Not

Somebody who has never met me could watch that video and think nothing of it. I knew it was wrong within about four seconds, because I know what I'm supposed to look like.

Buildings are the same. An AI-generated image can be entirely believable to somebody who has never seen the drawings. The architect who designed it will spot a dozen things that don't belong, usually before they've finished their coffee.

The awkward part is the gap between those two reactions. A plausible image can be in circulation for weeks before anybody who knows the scheme properly looks at it.

What Two Years Has Taught Us

We committed to this properly around two years ago, when a good part of the industry was still deciding whether it counted as cheating.

Every week the whole studio sits down for a meeting we call WTF, which increasingly lives up to its name. One person here has a full-time job working out what these tools can and can't do, and most of what she tests never reaches the workflow. That's the job working correctly rather than failing.

What I'd take from two years of it is this. The tools have changed how the work gets done. They haven't changed what the work is.

Measured Shortcuts

Every image involves shortcuts. That isn't a confession, it's how anything gets delivered to a budget and a date. A tree on a distant ridge doesn't need individual leaves.

The difference is that an experienced artist chooses them, knowing which will survive scrutiny and which will be the first thing a planning officer puts a finger on.

AI takes shortcuts too. What it can't do is tell the difference between a detail that can safely be approximated and one that has to match the approved design exactly. The trade calls these hallucinations, which makes them sound rather exotic, when mostly they're a gap filled with something plausible instead of something true.

A shortcut you chose, you know where you left it. One you didn't has to be found, by a person, looking closely at an image that seemed perfectly fine at a glance.

Try Changing One Thing

This is the part that catches people out, and it has nothing to do with the first image.

Make the cladding darker on the left return only. Move a tree three metres. Change the signage without disturbing the elevation behind it.

The editing tools are improving at this quickly, and some of them make remarkably convincing local changes. But convincing isn't the same as controlled. The test is whether the change is accurate, repeatable, and still consistent with everything the project has already agreed.

I'd put revision at around half of everything we do. Not corrections, mostly, but absorbing comments from the design team, the client, the agent and anyone else with a legitimate view, and making one image satisfy all of them without breaking the architecture.

Which is why the cost of getting an image and the cost of getting it approved are two different numbers. I've seen plenty produced quickly and cheaply that then had to be rebuilt once somebody held them up against the drawings.

How Much Control Do You Need?

All of this is moving quickly, and I wouldn't bet against a good deal of what's difficult today being routine next year. Which is why I don't spend much time wondering what AI will manage eventually.

It's this. How much control will you need after the first version arrives?

For an early concept, perhaps very little. Take the quick route and spend the money somewhere it matters more. For something that has to match approved information, survive four rounds of comments and hold up in front of a committee, rather more than that, and it's better to know which one you're commissioning before it gets made rather than afterwards.

Both are entirely legitimate briefs. They just aren't the same job.

Until next week.

Rich

PS: My AI self is available for meetings I'd rather not attend. He nods a lot and never asks about budget.