Contra did something quietly revealing. The platform "the network for creative intelligence," spun up a research arm and its first project is the Human Creativity Benchmark, a system meant to become the standard that every creative AI tool gets judged against. To build it they had to do something the AI labs mostly skip: define what creative quality actually is. And the most interesting line in the whole framework is the one where they admit a large part of it can't be scored objectively at all.

Their evaluators split every creative decision along one axis. At one end is convergence, the stuff professionals agree on, where there's a verifiable right answer. Did the output follow the prompt. Is the kerning broken. Are there artifacts. At the other end is divergence, where expert evaluators legitimately disagree, because the question has stopped being correctness and become taste. Does this feel right. Each output gets scored by three or more working creatives, and the benchmark exists precisely because that second end won't reduce to a number.

An AI is optimized to hand you the answer the most people would approve of. Taste is the willingness to hand them the answer most people won't, for a reason you can defend, and the better the model gets the more that refusal is the entire job.

What the model is actually optimized to do

Something most of the replace-the-designers discourse gets wrong in both directions. The lazy dunk is "it's just predicting the next token, so it can't really create" but that's a weak explanation and serious researchers have torn it apart, because explaining a model by the token it emits is like explaining a jet engine by pointing at the exhaust. These systems abstract, combine, and generate things that were not sitting in the training data verbatim. Novelty is not the thing they lack.

What they're optimized toward is the thing that matters. A base model learns to predict the most probable continuation of human text, which already pulls it toward the center of gravity of everything written before. Then we run reinforcement learning from human feedback on top, which tunes the model to produce the response an average human rater would rate highly. Researchers have a precise name for what that does: it's mode-seeking. It narrows the output distribution, lowers the entropy, and makes the model more deterministic with each alignment pass. There is a paper, not from a brand studio but from machine learning researchers, titled "Creativity Has Left the Chat," documenting how the very debiasing that makes these models pleasant and safe also flattens their creative range. The optimization target is, almost literally, the consensus-acceptable answer. The system is built to minimize surprise.

Now lay that over Contra's two ends. On the convergent end, where there is a right answer professionals agree on, a model optimized for the consensus answer is going to be devastating, and it should be. Layout conventions, accessible contrast, on-trend polish, the grammar of a competent interface, all of that is convergence and the machine will eat it but the divergent end is defined by the absence of a consensus to predict. It is the zone where the average is, by definition, the wrong answer. Aiming a consensus engine at it produces the most expected version of an unanswerable question, which is the one thing taste is supposed to never be.

Taste is not novelty, and that's the whole misunderstanding

People keep trying to settle this by asking whether AI can be "creative" and it's the wrong question. Creativity in the loose sense, surprising output, is cheap. Turn the temperature up and you get plenty of surprise. What you don't get is judgment, and taste is judgment, not novelty. Taste is a directed deviation from the expected, a specific choice that a meaningful number of people will dislike, made on purpose because it's right for this one company and this one customer. Randomness diverges in every direction at once. Taste diverges in exactly one, and can tell you why.

That "why" is the tell, and it's the thing the model has no access to. It has no stake in the outcome, no client who will be in the room when the call gets defended, no reputation riding on the unpopular choice turning out correct. It can generate the option. It cannot commit to it. When AI output does land as genuinely tasteful, look closely and there's almost always a human at the controls who prompted toward it, recognized it among a hundred bland siblings, and chose it. The model widened the option space. The person supplied the judgment about which option was right. Strip the person out and you don't get taste, you get the average wearing this season's texture.

Invention is a bet against consensus

The same logic settles the bigger claim, the one about whether these things will invent. Invention is not recombination, it's a bet against the consensus that turns out to be right. Every real one looked wrong to the average evaluator at the moment it was made, which is exactly why it was worth something. A system whose entire objective is to converge on what the average rater already approves has no mechanism for that bet and no reason to hold it when the room pushes back. It will give you the brilliant synthesis of everything that already exists, instantly, for free. It will not give you the thing the existing data argued against, because the thing the data argues against is precisely what it was trained to avoid.

So if you actually understand how the machine works, the conclusion isn't the comforting one where humans are mystically special. It's narrower and more useful. The machine collapses the cost of everything on the convergent end of creative work to roughly zero, and in doing so it relocates the entire value of a designer onto the divergent end, the defensible no-consensus call. That was always the scarce part. It just stopped being subsidized by all the convergent labor that used to come bundled with it.

The so-what for founders

Steal Contra's axis and run your own brand decisions through it. Sort every creative call into convergence or divergence. The convergent ones, where there's a known-good answer, automate hard and feel no guilt about it, because a human sweating those is a human wasted. The divergent ones, the positioning sentence, the single risky color, the joke in the headline that one specific buyer will love and another will hate, never hand those to a consensus engine, because on that axis the average is the failure mode, not the goal.

The mistake isn't using AI. It's not knowing which axis you're on when you use it. Outsource the convergent layer and you move faster. Outsource the divergent layer and you will ship the most agreeable version of your category, on time, under budget, indistinguishable from three competitors who made the same efficient choice. Inoffensive is a thing the machine does better than you ever could. It's also the one outcome a brand cannot survive.

If you can't tell which of your brand decisions are convergence and which are taste, drawing that line is most of what I'm actually hired to do. Reply with the call you're stuck on, or see the work at bybrightstudios.com.