Type a prompt, get a clean interface. Reasonable spacing, sensible hierarchy, a colour palette that will not embarrass anyone. Five years ago that output would have cost you a fortnight of a mid-level designer's time. Today it is effectively free, and that is precisely the problem. When everyone can generate a competent screen, competence stops being a differentiator and becomes the price of entry.
Here is the mechanic that most product leaders are underweighting. Generative models are trained to predict the most probable next thing. Their entire architecture optimises toward the common case, which means they produce the median by design. Ask for a dashboard and you get the statistical average of every dashboard the model has seen. That is genuinely useful for speed, and genuinely dangerous for strategy, because your competitors are typing similar prompts into similar tools and converging on the same answer.
The average is now free, and everyone is spending it
As competent interfaces flood the market, sameness becomes the default state. This is counterintuitive. You would think cheaper design means more variety, but the opposite is happening: the cost of the average has collapsed, so more products ship toward it. Differentiation gets scarcer, not easier. The interfaces that stand out are increasingly the ones a model would never have suggested, because a model has no reason to suggest them.
AI can render any design you can describe. It cannot tell you which design is worth describing.
That gap is where taste lives. And taste is not a mood or an aesthetic preference. It is a stack of contextual decisions, each of them defensible only if you understand something the model does not have access to:
- Who this specific user is, and what they are actually trying to do at the moment they open the product.
- What the business needs to be true, including the thing you are quietly optimising for that no competitor can see.
- What to leave out, which is almost always harder and more valuable than what to add.
- Which convention to break on purpose, and which to keep so the deliberate break actually registers.
None of those are pattern-matching problems. They are reasoning problems grounded in context the training data never contained, which is exactly why the model defaults to the average instead.
Where to spend the judgment you have left
The practical move is not to resist AI or to hand it everything. It is to let AI compress the execution, the drafting, the wiring, the endless competent variations, and to concentrate your scarce senior judgment on the handful of decisions that create defensibility. Most teams invert this. They spend human hours on production and let a tool make the strategic calls by omission.
We run this deliberately: small senior teams making the contextual decisions, AI-accelerated delivery handling the fidelity, prototype-first so the judgment gets tested against real users before anyone commits to a build. The point is not to design faster. Everyone can now design faster. The point is to design the thing the average would never reach.
