Jake Young
For most of history, making things was expensive and choosing between them was cheap. Generative tools have flipped that. A team can produce fifty homepage directions, two hundred headlines and a month of social posts before lunch. The cost of a brand has moved from making to choosing, and choosing well is what we mean by taste.
That makes taste the most defensible asset a brand has, and most companies manage it badly, as a feeling one senior person carries around. The brands that pull ahead will turn it into a system a whole team can follow, with review gates that keep the judgment human and hand the checking to machines.
Why Competent Work Is Converging
The research on generative tools keeps finding the same pattern. In a 2024 Science Advances study, Anil Doshi and Oliver Hauser gave some writers story ideas from a language model. Their stories were rated more creative and better written, especially among writers who started weaker. They were also more similar to each other. Individuals improved while the pool narrowed.
Cornell researchers found a related effect at CHI 2025: AI writing suggestions pulled Indian and American participants toward the same Western style, making writing more alike within and across both groups. A July 2026 preprint by Nakayashiki and Watanabe compared which sentences models and human readers picked as important in 120 web documents. On a typical 70-sentence document, two readers agreed on about four selections. Two models agreed on nearly nine.
Put those together and you get a market problem. When every team uses similar tools trained on similar data, the tools pull everyone toward the same idea of good: a clean layout, a confident sans serif, a soft gradient, benefit cards, a friendly headline. None of it is bad, and all of it is interchangeable. Interchangeable work is hard to remember, because nothing in it tells a person which brand made it.
The Audience Has Noticed
Merriam-Webster’s editors chose “slop” as their 2025 word of the year, defining it as low-quality digital content produced usually in quantity by AI. Inside companies, researchers at Stanford’s Social Media Lab and BetterUp named the office version “workslop.” About 40% of the 1,150 workers they surveyed had received it from colleagues, and 42% of recipients said they saw the sender as less trustworthy afterwards.
Consumers are moving the same way. The IAB’s January 2026 study found 82% of ad executives believed Gen Z and Millennial consumers felt positive about AI-generated ads. Only 45% of those consumers did, and 39% of Gen Z felt negative, up from 21% in 2024. The people making the work are further from their audience than they think. Even Google’s guide to its AI search features asks site owners for “non-commodity content” that goes beyond common knowledge.
What Premium Means When Anyone Can Make Good
Premium used to be signalled by visible production cost: the shoot, the custom build, the months of craft. When a model can imitate the surface of all that, surface stops being evidence. What still reads as premium is evidence of decisions. Specificity, meaning a claim, image or phrase that could only belong to you. Restraint, with fewer things on the page and a reason for each. Coherence over years and channels, which is hard to fake because it takes thousands of refusals. Every one of these depends on what was left out, and a generator has no reason to leave anything out.
Taste Is a Refusal Rate
Romantic definitions of taste are hard to manage, so we use an operational one: taste is the ability to reject good options for reasons someone else can learn. The reasons are the point. A founder who says “I’ll know it when I see it” has taste without a system. The team can only guess, and the tools can’t guess at all.
That gives you a number to watch, the kill rate: the share of generated or proposed work that never ships. In a team using AI well, the kill rate goes up, because generation got cheaper and the standard stayed put. If most of what the tools produce ends up in front of customers, the standard has dropped to whatever the tools reach first.
The Judgment Stack
To make taste repeatable we build it in four layers. Each one makes the next enforceable, and together they let a team of ten, plus their tools, make decisions the founder would recognise.
Point of view
A written statement of what the brand believes and, more usefully, what it will never do. “We never use stock photos of people pointing at laptops.” “We never put two calls to action on one screen.” “We never promise an outcome we can’t measure.” Refusals beat values statements because they’re testable. Anyone can hold a draft up against them.
Canon
A reference set of work you admire, from any field, with one sentence on why each piece is in. Next to it sits an anti-canon: work you respect but must not resemble, which is where your competitors belong. The canon gives people a shared vocabulary. Loaded into your AI tools as context, it pulls output away from the average.
Constraints
The brand system, written as limits rather than suggestions: a type scale, colour ratios, spacing, motion rules, photography direction, and voice rules with right and wrong examples. Mailchimp’s public content style guide and the GOV.UK style guide show judgment written down clearly enough for strangers to apply.
Editor
A named person with the authority to say no and the obligation to say why. Committees average decisions the way models average the web. The editor owns the kill rate, signs off anything a customer will see, and turns any reason they give more than twice into a new constraint.
Design Systems as Decision Systems
Most design systems are treated as parts libraries: buttons, cards, modals. We treat them as decisions made once so they don’t get remade badly a thousand times. Public systems like GitHub’s Primer, IBM’s Carbon and Shopify’s Polaris are worth reading for exactly this reason. Their documentation explains when not to use a component as carefully as how to use it.
In practice that means design tokens for every colour, type size, radius, shadow and easing curve, and a codebase that refuses anything else. A stylelint rule such as color-no-hex blocks raw colour values, so an off-brand blue can’t sneak in from a generated snippet. Components come with a small, deliberate set of variants instead of open-ended props. Motion gets named presets for duration and easing rather than per-screen tweaks. Each of these removes a decision someone would otherwise make under deadline, and each gives an AI coding tool a narrow lane to work in.
The trade-off is speed of novelty. A strict system makes the expected thing fast and the new thing slow, because a new pattern has to be argued into the system before it ships. We think that friction is healthy. It keeps experiments deliberate and stops a brand drifting one convenient exception at a time.
Review Gates, and What to Automate
Taste fails at handoffs, so we put gates at each one, and we split every gate into what a machine can check and what a person must judge.
Gate one is the brief. Before anything is generated, a person confirms the problem, the audience and which refusals apply. Nothing is automated here, because a wrong brief produces confident, polished work pointed at nobody.
Gate two is direction. Options are reviewed against the canon and anti-canon, and most are killed. Tools can help by generating wide, but the choice stays human, and the editor writes one line on why the survivor won.
Gate three is conformance, and it should be almost entirely automated. Token and lint checks in CI. Accessibility checks with axe. Performance budgets enforced with Lighthouse CI so a heavy hero can’t merge. A prose linter such as Vale, which GitLab uses on its documentation, to flag banned words and voice rules in copy. Visual regression screenshots so a change in one component can’t break twenty pages unnoticed. Machines are better than people at this gate, and every minute they save goes to gate four.
Gate four is the edit. The editor looks at the finished thing in context, on a real phone, and asks the questions no check can: is it right, is it ours, and what can we remove? This is the gate most teams skip under deadline, and the one customers feel.
The dividing line is simple. Automate anything with a correct answer. Keep human anything that needs a reason. When you find yourself giving the same reason repeatedly, write it down and push it one gate earlier, into a constraint or a check.
The Next Two to Five Years
These are predictions. The creative bottleneck moves to review, and creative leaders get judged on the quality of their rejections, which brings back editor as a senior brand role. Brand guidelines get rewritten as rules and references a tool can load, so generated work starts inside the brand. And the gap between considered and generated brands widens, with the cost of being interchangeable showing up slowly, as rising acquisition costs. We can’t cite that last part yet. We’d still bet on it.
What to Do This Quarter
Write ten refusals that start with “We never,” then check the last month of shipped work against them.
Build a canon and an anti-canon of twenty references each, with one line of reasoning per item, and share them with everyone who briefs, writes or prompts.
Rewrite every “should” in your brand guidelines as a rule someone can check, with right and wrong examples, and load those rules into the AI tools your team uses.
Automate gate three: tokens enforced in code, accessibility and performance budgets in CI, and a prose linter for voice.
Name one editor with veto over everything customer-facing, give them protected time each week, and start tracking the kill rate.
Run an interchangeability test. Put your homepage beside three competitors with the logos removed and ask someone outside the company to match them. If they can’t find yours, that’s your first project.
Tools will keep getting better at producing good. Knowing what to keep is still a human job, and it’s the part of the work we care about most. If you want help defining your brand’s point of view and building the system that protects it, from the canon to the codebase, we’d love to talk.
