The Algorithm of Aesthetics: AI in West African Design

Generative tooling has collapsed the cost of visual exploration. It has also flattened the texture of brand work across emerging markets. The same six gradients, the same isometric illustration, the same drop-shadowed glyph — rendered in Lagos, Nairobi, Accra and Abidjan within the same week. The studios that win the next decade will use AI to interrogate cultural primitives, not to imitate global templates.

The monocultural drift problem

Default models reward the average. They were trained on a public internet whose centre of gravity sits in San Francisco, Berlin and Seoul. Ask any frontier model for "a modern African brand" and you will get a mood board that looks suspiciously like a Brooklyn restaurant from 2018. This is not a bug. It is a statistical inevitability — and it is the single biggest reason most AI-assisted brand work in our market reads as derivative.

The fix is not to abandon the tools. The fix is to refuse the defaults.

Building a proprietary reference library

Our working pattern at Label: feed proprietary reference libraries — adire textiles, nsibidi script, broadcast television from the 80s, Lagos market signage, Onitsha pamphlet covers — into constrained generation loops. Each reference is hand-labelled with the property we actually want to extract: the line weight of a hand-painted sign, the colour drift of degraded VHS, the specific geometric repetition of a particular weaver.

The library is small on purpose. Two hundred references curated by a designer beats two million scraped from Pinterest. The model is downstream of the library; the library is downstream of taste.

The output is not the artefact. The output is a faster argument with the designer.

Prompt as brief, not wish

A prompt is a creative brief written for a junior who works at the speed of light and has no taste. Treat it accordingly. Specify constraints, not aesthetics. Name the references explicitly. Reject the first three outputs by default — they are the model regressing to its mean.

The disciplined version of this workflow shortens the discovery phase of a brand engagement from four weeks to four days. The undisciplined version produces a deck full of work that looks like everyone else's deck.

Where AI should not touch the work

Three places, in our practice: the final wordmark, the photographic identity system, and any copy that carries the brand's voice in a regulated context. The wordmark because it must survive a hundred years of reproduction at every scale. The photography because the most defensible thing a brand owns in 2026 is a proprietary visual culture, and generative imagery cannot be defended in court or in court of opinion. The copy because hallucination in financial services, healthcare or public infrastructure is a liability event, not a design choice.

Everything upstream of those three — exploration, moodboarding, secondary illustration, motion studies, design system extension — is fair game and faster.

What we ship versus what we discard

We discard roughly nine out of every ten generated artefacts. That ratio is the point. The economics of AI-assisted design only work if the team has the discipline to throw away most of what it produces. Studios that ship the first acceptable output are not using AI; they are being used by it.

If you are a founder evaluating an agency's AI workflow, ask one question: what is your discard rate, and who signs off on the keepers? If they cannot answer, the model is in charge.

Want to pressure-test how your product, brand and growth stack hold up to scrutiny? Run the Phase-0 GTM Scorecard — twelve questions, a calibrated grade, and a written diagnostic in your inbox.