Every few months somebody sends me the same message. A link to a new tool, a screenshot of fifty logo options generated in eleven seconds, and one line underneath: “Should we be worried?”

My answer has not changed in three years. Worried about what, exactly? The tools are real, and they are good at what they do.

What they do is make. Making was never the expensive part of design. It only looked that way because it took so long.

AI in design means using machine-learning tools, from image and layout generators to copy assistants and automated production systems, to create, vary and finish visual work faster than people can by hand, while the decisions about what is right still sit with humans.

That last clause is the whole argument. Here is why.

What actually changed?

Three things. Only one of them is interesting.

The first is speed. A mood board that used to take a junior designer a day now takes twenty minutes. Background plates, product mock-ups, colour variations, social formats in six sizes: the drudgery has collapsed.

Ask any studio owner where those saved hours went, though. Usually into more rounds of revision. Not quite the productivity miracle anyone was promised.

The second is volume. When making costs almost nothing, people make more. Far more. The internet is filling up with competent, slightly glossy, faintly familiar imagery, and every brand that leans on the default settings adds to the pile.

The third is the interesting one. Generated work converges. Ask a model for “a premium skincare label” and you get something that looks like the average of every premium skincare label it has ever seen.

That is the point of a statistical model. It is also the opposite of the point of a brand.

When people use generative tools for creative tasks, the weakest individual work tends to improve, sometimes a lot. But the outputs drift towards each other. Individually better, collectively more alike.

For a field whose entire commercial value lies in being recognisably different, that is not a footnote. It is the headline.

The weather app that wasn’t

If you want one story that shows the convergence problem in miniature, take Figma.

In early July 2024, the company switched off its new Make Design feature within days of launch. Andy Allen, founder of NotBoring Software, had asked it for a weather app and kept getting something remarkably like Apple’s own Weather app.

Dylan Field, Figma’s chief executive, said the feature had not been trained on customers’ files. He blamed the commissioned design systems underneath it and promised better quality checks.

Fine. Bugs happen.

But look at what the episode reveals. A tool from the most respected design-software company of its generation, asked for something “not boring”, produced the most famous existing answer. That is not a malfunction. That is the technology working as designed, pulling towards the centre of what already exists.

Test these tools seriously and you keep reaching the same verdict. In 2024 the layout generators produced generic results that needed heavy rework, and the same prompt could return wildly different outputs: handy for brainstorming, ruinous for repeatable workflows.

By 2025 they were better, mostly at narrow, single-task jobs like renaming layers or suggesting palettes. None of them worked well with an established design system. Read that sentence twice if you run a brand.

So who is AI actually replacing?

This is the question that turns up on every design forum, usually at two in the morning. The honest answer is uncomfortable, and it is not “nobody”.

The work most exposed is the work that was already a commodity: stock-style imagery, background illustration, routine resizing, filler copy, the jingle for the regional car dealer. These jobs were never glamorous. But they paid rent while people learned the craft. Remove the bottom rung and you have to ask how anyone climbs the ladder.

Netflix Japan gave the industry a preview in early 2023. Its short anime, The Dog & The Boy, developed with WIT Studio, used image generation for the background art. Netflix framed it as an experiment to help an anime industry facing a labour shortage.

Artists were furious. Their reading was simpler: a large company had found a way not to pay background painters and called it innovation. Both readings can be true at once. That is what makes this moment so awkward.

The work least exposed sits upstream of the pixels. Interpreting a vague brief. Deciding what the problem actually is. Telling a founder that the thing they love is wrong for the customer they want.

Software engineers have been saying something similar for years: the hard part was never typing the code, it was deciding what to build. Design is no different.

Taste is now the scarce skill

When a resource becomes abundant, the value moves to whatever remains scarce. Making is abundant. Selection is not.

Taste, in a working studio, is not a mysterious gift. It is a trained ability to look at forty options and know, quickly and for reasons you can explain, which one is right for this brand, this audience and this moment. It is editing. Generative tools have turned every designer into an editor, whether they wanted the job or not.

And editing needs a standard to edit against. That standard is the brand strategy.

Without a clear position, a defined personality and a set of distinctive assets you are deliberately building, you have no way to choose between outputs. So you pick the shiniest one.

Do that for a year of content and you get a brand that looks like everyone else’s. Generically good. Instantly forgettable.

Picture a Tiong Bahru café that commissions twelve months of social posts from a prompt. Every image is pleasant: latte art, warm light, linen, a pastry. None of it could only belong to that café.

Now give the same café a two-page brand platform that says: we are loud, a bit chaotic, proudly local, and we put our regulars in everything. Same tools. Completely different output, because someone now knows what to throw away.

Which brands used AI well?

The useful question is not whether a brand “used AI”. It is how, and whether the idea came first.

Ferrero ran an early version of the good pattern in 2017, before anyone said “generative”. Ogilvy & Mather Italia built an algorithm that drew on a library of patterns and colours to produce seven million different Nutella jar labels, each with a unique code.

The jars went on sale in Italian supermarkets and reportedly sold out within a month. The machine did the multiplication. Humans designed the system, the rules and the idea that a jar could be a collectible.

Ito En in Japan used AI more visibly in October 2023, casting an AI-generated woman in a television spot for its Oi Ocha Catechin Green Tea. She grows younger as she drinks, which at least suits the product’s health positioning.

The company also used AI to generate label design candidates, which its own designers then refined. Note the order. Machine proposes, people decide.

KFC Singapore found the wittiest route in March 2025. With Havas, it ran outdoor ads, MRT placements and in-store visuals across 26 outlets showing people licking their fingers, each person with a few too many.

The campaign turned AI’s most notorious glitch into a “finger lickin’ good” joke. It worked only because the brand idea already existed and was strong enough to bend the technology to it.

And which ones didn’t?

Toys “R” Us premiered a one-minute brand film made with OpenAI’s Sora at Cannes Lions in June 2024, telling its founder’s origin story. Online reaction was brutal.

The teaser for the Sora-made origin film: sentimental heritage, rendered by a model.

The film director Joe Russo publicly panned it, and viewers fixated on the uncanny faces and strange movement. The idea was sentimental heritage. The execution felt like nobody’s memory at all.

Closer to home, in July 2026 a community club in Kampong Chai Chee replaced an AI-generated National Day banner on the same day residents spotted the problems. Several flags did not depict Singapore’s flag correctly, a couple looked closer to Indonesia’s, and hands and figures were distorted.

The club, run largely by volunteers, handled it gracefully and said AI output should always be checked before use. A small story. Also the most common AI design failure in the world: nobody with a trained eye looked before it went up.

Is AI-generated design safe to use for a brand?

People ask this as a legal question, and partly it is. Who owns a generated logo, whether it can be registered as a trademark, and what the model was trained on are live issues that vary by jurisdiction.

Some vendors make a point of training on licensed material. Adobe says Firefly was trained on Adobe Stock and public-domain content, and markets its outputs as commercially safe, though later reporting suggested the training set was less pure than advertised. Read your tool’s terms as you would any supplier contract, then ask a lawyer about anything you plan to trademark.

The bigger risk is brand risk. A logo assembled from the average of a million logos is, by construction, hard to own and easy to confuse. Distinctiveness is the asset, and anything that pushes you towards the middle of the category erodes it, quietly, one acceptable output at a time.

Then there is the audience. People spot generated imagery more often than vendors like to admit. In categories built on craft, authenticity or community, discovering that a brand used AI where a person was expected can feel like a small betrayal.

Disclose where it matters. Never pretend.

The tools are consolidating. Your thinking shouldn’t.

Watch where the money is going. Canva, the Australian platform founded in 2013, bought Serif, the developer of the Affinity design suite, in March 2024 and the image-generation startup Leonardo.ai that August. In 2025 it added a design model that produces layouts with editable layers.

Figma, Adobe and the rest are racing the same way. Within a couple of years generation will be a default feature of every design tool, as unremarkable as spell-check. And when everyone has the same tools, the tools stop being an advantage.

Pricing is shifting with them. Freelancers and agency owners are openly working out how to bill when hours shrink but compute costs and client expectations rise. The ones who survive will stop selling hours of making and start selling decisions.

Clients do not, in the end, pay for fifty options. They pay for the right one, and for the confidence that it is right.

What designers should actually do

Not a manifesto. A short working list.

Use AI for the boring bits, ruthlessly. Resizing, background removal, mock-ups, first-pass mood boards, copy variants for testing. Take the hours back and spend them on thinking.

Build the system before the outputs. Codify colours, type, composition rules, tone and distinctive assets tightly enough that you can judge any output, human or machine, against them. A brand manual is now also a quality filter.

Get better at saying why. Clients can generate options themselves now. What they cannot generate is a reasoned argument for one option over another, made in business terms.

Push away from the centre on purpose. If the first output looks like the category, that is your cue to go the other way. The model shows you what is expected. Your job is to decide what is memorable.

Keep drawing. Sketching, photographing, writing by hand. Not out of nostalgia. Out of self-defence. The most distinctive designers of the next decade will be the ones whose ideas did not start in a prompt box.

Check everything before it ships. Flags. Fingers. Spelling in the background signage. Ten seconds of a trained eye beats a public apology.

Questions from the studio floor

Will AI replace graphic designers?
AI is replacing routine production work such as stock imagery, resizing and simple layouts, but not the judgement that decides what a brand should look like and why. Designers who move upstream into strategy, art direction and editing are becoming more valuable, not less.

Can I trademark an AI-generated logo?
It depends on the jurisdiction, the tool’s terms and how much human creative input went into the final mark. Treat any generated logo as a draft, have a designer develop it into something distinctive, and take legal advice before filing.

What is AI in design best used for today?
It is most reliable for narrow, repetitive tasks such as resizing, background removal, mock-ups, colour exploration and early-stage ideation. It is least reliable for building coherent identity systems or working within an existing brand’s design rules.

Cheap making is here to stay; if you want help deciding what your brand should make with it, talk to our brand consultants.