Scroll a feed of new brand work and something starts to nag. The gradients are the same gradients. The serif is doing the same thing it was doing on the last one. The photography has the same slightly weightless quality, lit from nowhere in particular. Everything is competent. Almost nothing is memorable.

This is not a conspiracy or a decline in taste. It is arithmetic.

Why the convergence happens

A generative model produces the most probable continuation of a prompt. That is not a criticism, it is the design. Trained on an enormous quantity of existing work, asked for a logo or a layout or a photograph, it returns something near the centre of everything it has seen.

The centre is a genuinely good place to be for most problems. It is why these tools are so useful. Competent, conventional, unsurprising output is exactly right for a slide deck, an internal doc, a first draft, a piece of scaffolding you will replace later.

A brand is the specific case where it is wrong. The entire function of an identity is to be distinguishable from the other things in its category. Asking a system optimised for the most likely answer to produce the least likely one is asking it to work against itself.

Push harder on the prompt and you get a more elaborate average. Not a different one.

Where it genuinely helps

We are not sceptics here. Two of the tools in this Lab were built conversationally, and they work, and they took an afternoon each rather than a fortnight. That is real.

The distinction that has held up for us is between generation and execution.

Generation is deciding what the thing should be. What it means, who it is for, why this direction and not the four others on the wall. That is a judgment call made with incomplete information, and it is the part clients are actually paying for.

Execution is everything downstream of that decision. Building the thing, testing it, adapting it across formats, catching the edge cases. Enormous amounts of that work are now faster, and pretending otherwise helps nobody.

The tools collapsed the cost of execution. They did not touch the cost of judgment.

What this changes

If competent execution is close to free, competent execution stops being a differentiator. That is uncomfortable for anyone whose value was in production speed, and it is straightforwardly good for anyone whose value was in deciding what to produce.

It also raises the floor on what is expected. A rough draft is no longer an acceptable deliverable when a rough draft takes ninety seconds. The bar moves to the thing that could not have been generated.

Which is, in practice, the thing that comes out of understanding a specific client, in a specific market, with a specific problem that does not resemble the average of anything.

The useful question

Not whether to use these tools. That question is settled and the answer was obvious.

The question is which parts of the work were ever supposed to be average, and which parts exist precisely because they are not.

Get that division right and the tools are extraordinary. Get it wrong and you ship something indistinguishable from everyone else who also got it wrong, faster than ever before.