Why We Can't Agree on What Good AI-Generated Identity Work Looks Like

Three weeks ago we put an AI-generated identity concept in front of the whole studio, expecting a quick thumbs up or down. Instead the room split almost exactly in half. One group thought it was the sharpest mark we’d produced all quarter. The other thought it looked like six good logos with the edges sanded off until none of them were recognizable anymore. Nobody was pretending. Both sides meant it.

That split has stuck with me, because it wasn’t really an argument about one logo. It was an argument about what we think AI-generated design work is even for — and I don’t think we’ve settled that, individually or as an industry.

What was actually being argued about?

On the surface, the disagreement was about a single shape: a wordmark with a lowercase letterform that some people read as confident and others read as cautious, a design that wouldn’t offend anyone in a room and therefore wouldn’t be remembered by anyone either. But underneath that, the real disagreement was about competence versus character. Is a mark good because it’s executed cleanly, or good because it has a point of view sharp enough that some people are going to dislike it?

The AI tool had, if anything, executed the brief with more consistency than a first-round human draft usually shows. Grid alignment, optical balance, spacing between the wordmark and the icon — all clean. What it hadn’t done was take a risk anyone could point to and say, that’s the interesting decision. It had smoothed every option toward the statistical middle of what a good logo tends to look like.

Is consistency actually the problem?

One camp in the room argued that’s precisely the point of using these tools at this stage — get to competent fast, then let a human push it somewhere specific. The other camp argued that if the first pass never contains a genuine risk, humans rarely add one later, because by the time a draft looks finished, most reviewers stop trying to break it. Comfortable output tends to survive review unchanged, not because it’s right, but because it’s hard to argue against something that isn’t wrong.

Neither position is unreasonable. That’s what made the disagreement worth sitting with instead of resolving by vote.

Is this actually new, or just louder?

Every generation of design tooling has produced some version of this argument. Vector software got blamed for identikit logos in the 90s. Template-based site builders got blamed for a decade of interchangeable landing pages. The complaint is rarely really about the tool — it’s about what happens when the effort required to produce something drops faster than the judgment required to evaluate it does. AI-generated identity work just compresses that gap further and faster than anything before it, so the argument that used to unfold over a decade is now unfolding over a single client review.

So who decides when a debate like this won’t resolve?

We didn’t pick a winner in that room, and I don’t think we were supposed to. What we did instead was change the brief: use the AI-generated version as the floor, not the answer, and require every revision after it to introduce at least one deliberate risk that a reviewer could reject on its own merits — not just polish the same safe shape further. That gave both camps something real to work with instead of a compromise that pleased no one.

The disagreement itself turned out to be the useful part. A room that agrees instantly on an AI-generated first draft is usually a room that stopped looking closely, not a room that got a good answer on the first try.