Why AI work drifts off brand, and how to hold it

Drift is output moving away from the agreed look a little at a time, with nobody deciding that it should. It is the mechanism, and carelessness has little to do with it. Every generation is a draw from a distribution centred on the model's average, so anything you do not explicitly hold will move towards that average. The palette warms first. The light softens, the composition slides to the centre, and the brand's own oddities, the details that made it recognisable, are the first things to be smoothed away.

It is hard to notice because each frame is defensible on its own. Nobody sees a problem in one image. The failure appears only when you put January's work next to June's, and by then six campaigns have shipped and three freelancers have each nudged the look in their own direction. This is a silent failure. No single instance looks wrong, so nobody reports it, and the drift is usually found by a new marketing director who asks why the brand looks different, or by a customer who no longer recognises the pack on a shelf.

Some things drift faster than others, and they tend to be the things a brand owns. Colour usually goes first. A diffusion model has no concept of an exact value, so a requested hex code comes back as a colour near it, and the light and grade in the scene move it further. A brand typeface cannot be reproduced at all, only imitated, and the imitation is wrong in the details a designer spots at once. Packaging combines small type, a specific structure and regulated copy, and a model approximates all of it while producing something that looks finished.

The process adds drift of its own. Every person writes prompts slightly differently, every art director finds new reference images, and every model update changes what the same words produce, so a prompt that gave the right look in spring can give a different one in autumn without anyone touching it. Brand books do not help much here. They were written to persuade people, with aspirational prose and rules made of adjectives, while a model works best from short, specific statements. A rule such as fresh but grounded gives a machine nothing to hold.

The defence is to make consistency structural instead of relying on attention, which means putting it where a tired person cannot skip it. A house block giving the light direction, the palette values, the camera height and a do-not list goes in front of every prompt. A frozen set of reference images goes with every job, and current models accept several at once; Google's Nano Banana 2, for example, takes up to ten images of objects and four of characters in a single request. For the product itself, a trained adapter or a real photograph composited into the scene does what words cannot. None of it depends on memory.

Then check the work regularly. A quality gate that runs automatically sets out what that check should test before anything ships. For drift, its value is that it never tires. A person checking colour and stray text gets less careful as the day goes on, while a gate puts the same questions to the last image of the day as to the first. Small slides get caught one frame at a time, before they add up to a look nobody chose.

Finally, tell drift from decision. A look that changed because you chose to change it is fine, and the change should be dated and written down so the next person knows it was meant. A look that changed because nobody was holding it is the problem this module is about. The difference is whether anyone can point to the day the rule changed and the reason it did. Keep a short change log beside the rules, with the date, the old rule, the new one and the name of whoever agreed it.