Nobody decides to wreck a brand. There’s no meeting where the room votes to make the logo worse, no feedback round titled “let’s lose the plot.” What actually happens is quieter: four hundred small approvals, each one reasonable, each one judged against last week’s version instead of the thing everybody said yes to at the start. The work never breaks. It drifts.
I’ve watched this from every seat. Freelance, in-house, agency side, and now with a machine for a design partner, which turns out to be the seat where the whole mechanism finally becomes visible. But start with the oldest version, because I got to live inside one.
TMZ, 2010: matching the last thing
When I joined TMZ in 2010, the site moved at a pace most brands never experience. Stories broke overnight, pages went up in hours, promo art got built the same day it ran. And there was no brand system for the website. Not a thin one. None. So new work got made the only way work can be made at that speed: open the last thing that shipped, match it, go.
That sounds like consistency. It’s actually the disease. Every match introduces a little error (a color eyedropped off a compressed JPEG, a headline set a touch tighter, a button borrowed from a page that borrowed it from another page), and the next person matches THE ERROR. Nobody compares today’s page to the brand. They compare it to yesterday’s page, and yesterday’s page was already a few percent off. Line up work from January next to work from October and you’re looking at three different websites wearing the same logo… and no single day is to blame.
We got it back on track. Building the brand’s first in-house creative team was partly about talent, but mostly it was about building a shared answer to “what does TMZ look like,” writing it down, and putting it where the work happened. I’m proud of that. I’m also honest about what it was: a rescue. The system should have existed before the drift did.
Drift has exactly one cause
You’ve seen the client-side version of the same thing even if you’ve never touched a newsroom. Round six of feedback, and nobody in the thread is asking whether the work still serves the brief. They’re asking whether it addresses the comments from round five. The logo has grown four times, each time by an amount too small to fight about. The palette has quietly absorbed a stakeholder’s favorite blue. Every single step was defensible. The sum is something nobody chose.
So here’s the whole mechanism, stated plainly: work drifts when each version is judged against the previous version instead of the original intent. That’s it. Not weak designers, not bad taste, not “difficult clients.” A comparison pointed at the wrong target, compounding. If you only ever compare today to yesterday, you’ll approve your way, one reasonable inch at a time, to somewhere you’d never have driven on purpose.
The book that can’t push back
Ask a designer how to prevent this and you’ll hear the answer I gave for years: a style guide. And then, half the time, what gets made is a book. A gorgeous, bound, ninety-page book (or its modern stand-in, the ninety-page PDF) with the clearspace diagrams, the color values, and a moodboard spread the client will want to frame.
I like those books. There are real occasions for one: an acquisition, a licensing deal, the gift copy for the founder’s shelf. But be honest about how often the book gets opened on a working Tuesday. The banner ships at 6:40pm. Nobody is flipping to page 34 to check tracking values at 6:40pm. A bound guide is a photograph of intent, and drift isn’t afraid of photographs. The book sits outside the loop. Drift happens inside the loop.
What kills drift is a guide that lives where the work happens and gets consulted by default instead of by discipline. Tokens instead of swatches. Components instead of specimens. A source of truth the work is checked against while it’s being made, not a reference the work is supposed to loosely resemble. I’ve written about translating a brand into constraints a machine can’t casually wander away from (the brand kit was the easy part); the point here is narrower and older than AI. The format of your style guide is a drift decision.
The machines do it too
Now the part I find genuinely funny. Iterate on a design with AI and watch what happens across ten generations. The model sees the latest version and your latest note, and it optimizes against exactly that, the way a client call optimizes against the last deck. Generation ten looks nothing like generation one, and no single generation did it. The corners got rounder every round because rounder corners kept getting a nod. The palette wandered. Same disease, different patient. Researchers call the severe version model collapse: models trained on their own output slowly forget what the real thing looked like, one generation at a time.
AI didn’t invent drift. It runs the agency feedback loop at a speed where you can finally watch it happen, and that turns out to be clarifying in a way twenty years of postmortems never managed. The model has no ego, no politics, no committee, no nephew who “knows design”… and it still drifts. Which settles an old argument. It was never really about the client. The cause was structural the whole time: intent that lived in someone’s head instead of in the loop.
That’s also why the fix for the machine turned out to be the fix we already knew. Look at how anyone works seriously with these tools now: a written system the model has to ground every decision in, foundations locked in files, output checked against the brief instead of the last generation. It’s agency discipline, formalized, because the tool won’t pretend to remember what you meant.
Keeping the brief in the room
The practices, then. Notice that every one of them works on humans and machines for the same reason.
Lock the foundations early, and make them expensive to reopen. Type, color, grid, spacing. Feedback that happens inside a frame produces variation. Feedback that can touch the frame produces drift.
Write the rationale down at decision time. One sentence per decision: “the mark stays small because this brand is the whisper in a loud room.” Six rounds later, feedback has to argue with the reason instead of the pixels, and most drift can’t survive contact with a written reason.
Diff against round zero, not round five. The cheapest anti-drift habit I know. Before approving anything, put it next to the originally approved direction, not the last revision. Drift is invisible step to step and unmissable end to end; you just have to point the comparison at the right target (the eye is the spec is about training that catch).
Put the source of truth in the loop. Tokens in the build. A check at export. A model grounded in the system files. Whatever the form, intent has to be a participant in the work, not a book about the work.
At TMZ we earned our system the expensive way, after the drift, as a rescue. These days I build the guide before the first round of feedback, because I finally understand what it’s for. It was never decoration, and it was never really documentation either.
The brief leaves the room the moment it isn’t written down. Everything after that is the drift.