Generations Are Cheap.
Taste Is Not.

I can generate a thousand drafts of this post in under a minute for less than a dollar. I don’t send it out that way though. The companies, the products, and the sense of good taste are all built by people, not inference. That’s why we’re collectively developing an “AI sense.” It’s triggered when the photo is slightly off, or the grammatical tells start to show up everywhere. It’s not just you.

Cheap Generations, Expensive Discernment

Generations are cheap today. Small models capable of real work go from 10 to 40 cents per million tokens. We’re being flooded with so many half-cooked outputs that we can actually taste what’s still raw. As Meltwater puts it: “Brands that have made craftsmanship and quality a key part of their identities have a huge advantage as public sentiment turns against low-quality genAI content. Emphasizing qualities like expert design, human curation, and interpersonal connection is helping such brands appeal to discerning, AI-slop-averse audiences.”

The landscape hasn’t changed since then; it’s only accelerating.

In a world where generations are cheap and answers are easier than questions, good taste is the value. Retaining your ability to say “no, this simply won’t fly,” despite your agent’s kind disposition, is now a survival instinct.

After the Camera

After the invention of photography, you would hire a skilled portrait artist for a full day of sitting, while the artist captured every detail. But once people had instant-capture in their hands, billions of pictures could flow out of a finger resting on a button. Portrait artists didn’t completely disappear, but the skill of photography deepened with respect to framing, timing, editing, sequencing, curation.

I’d imagine we’re at that stage currently, where the people working with AI tools today are either developing the perfect lighting techniques, or just taking a candid selfie.

What Models Can’t Do

Now that we’re understanding AI capabilities better, we need to take a look at the wide shot. Can matters of taste and judgment be the last bastion of human expertise?

“Tools that once assisted human analysis now take the lead while we sleep, sorting, summarizing, and sifting through what matters. These systems never tire, but they also don’t yet care in the way people do. Caring, for now, remains a human responsibility.” — Harvard Chan, March 2026

You can’t program a machine to care like a human does. A model optimized for ‘correct task completions, as efficiently as possible’ isn’t necessarily thinking about which metrics are most important for you or a client. Sticking another LLM on top just to make the decisions for you is actually just kicking the can down the road at double the price.

The science suggests human specialists are still king: when researchers built RubricBench specifically to test whether frontier models could write their own evaluation criteria, they found a “substantial capability gap between human-annotated and model-generated rubrics” (arXiv:2603.01562, 2026). Models can grade against a rubric; they cannot reliably author one. Expert judgment is the input that the whole stack depends on.

Operationalizing Taste

We’ve taken a similar approach in authoring our own rubrics. The methodology is simple: rather than ask “is this output good,” decompose what “good” actually means first, put it into a structured set of criteria an expert can defend, then evaluate against that. This is the operationalization of taste.

Some aspects of taste are generally shared, and that’s what generic models do well. But the only one who can deconstruct what taste is in your perspective, is yourself.

If you’re leadership, operating a business, or the head of brand development, this should matter to you. Are you letting it ride on AI self-improvement and accepting generic outputs, or are you authoring your own future?