How to spot a design made by AI (and the questions to ask before you sign)

An agency sends you three proposals for the new site. They're good-looking. They're clean. Every one of them has a purple-blue gradient across the top, three identical cards with rounded corners, a semi-transparent blurred panel floating over a dark background, and a headline that says something like “build faster”. You like them. You sign.

Six months later you find out that your competitor's site, built by a different agency, looks like the exact same site. Not because anyone copied: because neither of them decided anything. Both inherited the default setting of a generative tool — and that default is identical for everybody.

This article won't turn you into an art director. It gives you what a business owner actually needs: the ability to look at a proposal and tell a decision apart from a default, and to know which questions to put to the vendor before you sign. It's a buyer's skill, not a designer's.

“AI slop”: from meme to word of the year

The term AI slop describes mass-generated AI content, produced without care and recognisable precisely by that absence of care. It entered common use in August 2024 and exploded after the Ghibli-style image craze of March 2025. In 2025 slop was picked as word of the year by Merriam-Webster and by Australia's national dictionary (Wikipedia, PBS News, Euronews).

Once a word reaches the dictionary, the thing it describes has stopped being an insider curiosity: it's something your customers recognise, even without being able to name it. A site that “looks AI-made” communicates one very specific thing to the person looking at it — that nobody paid attention to it.

First wave: the signals you can see with the naked eye

The first wave (2024-2025) has a narrow, repetitive visual vocabulary. These are the easiest signals to recognise, and you can check them off one by one on any proposal put in front of you:

• Purple-indigo-blue gradients, almost always diagonal, almost always in the header.
• “Glass” panels, semi-transparent and blurred (glassmorphism), over dark backgrounds.
• Glowing borders — a coloured halo around the boxes, on black.
• Floating 3D spheres, animated gradient blobs, particles and swirls used to illustrate “artificial intelligence”.
• Three identical cards side by side with the same thin icon on top: the most reproduced feature layout on the web.
• A single typeface (usually Inter) with no real hierarchy: headings and body text differ only in size.
• Interchangeable copy: “build faster”, “ship smarter”. Lines a competitor could paste onto their own site without changing a comma.

The fullest catalogue of these patterns is impeccable.style/slop, which currently lists 64 of them by name — “Purple Gradients Everywhere”, “Inter Everywhere”, “Identical card grids”. The British studio 925studios turned it into a buyer's checklist that adds generic copy — “Build faster. Ship smarter.” — to the list of tells.

The cause is more interesting than the symptom, and it's worth understanding because it explains why the phenomenon doesn't correct itself. The world's most widely used CSS framework, Tailwind, ships a default accent that is a purple-indigo; that purple saturated the public code the models were trained on; the models therefore generate purple interfaces; those interfaces go online and feed the next round of training. It's a loop that tightens on its own (prg.sh, Kai Ni, Indie Hackers).

The reverse holds too: 925studios names three digital products that escaped the generated average — Linear, Stripe, Duolingo — and the common factor isn't budget. It's deliberate constraint: a palette chosen and then respected, a typographic hierarchy actually picked, a visual language held for years. All of it is within reach on a small business budget, because it costs decisions, not rendering hours.

Second wave: when the cliché turns “tasteful”

Here comes the part almost no vendor will tell you, and it's the reason a first-wave-only checklist would protect you just halfway. Since mid-2026 the cliché has moved. Generative tools got better, they learned “good taste”, and the result is a new set of tells — restrained, warm, editorial this time:

• beige/cream palettes with terracotta-rust accents;
• huge italic serif typefaces used as headlines;
• widely letterspaced uppercase eyebrows;
• ticker-style scrolling text bars.

The person documenting it is Kyle Chayka, New Yorker staff writer and author of Filterworld, in “The Generic Style of AI Web Design” (also picked up in this reprint). In the piece, designer Celine Nguyen sums up the reaction of people working in the field: “I find myself instinctively repulsed by the warm tones”.

What that means for you, and it's the most useful thing in this whole article: if a vendor offers you cream-and-terracotta with the italic serif as the alternative to the purple gradient, they aren't offering you a decision. They're offering you the next default. Escaping the first wave by landing in the second isn't standing out: it's reaching the cliché six months late.

It isn't the first time: three precedents that show how it ends

The cycle — a visual language works, gets adopted en masse, becomes indistinguishable, provokes a backlash — has already run its full course at least three times. Knowing how each ended is what stops you making the opposite mistake.

Skeuomorphism → flat design (2013). The leather and felt textures of early mobile interfaces went from charming to embarrassing in four years; iOS 7 stripped them out, following Microsoft's flat Metro of 2010. But flat design provoked a backlash of its own — with no shadows or relief you could no longer tell what was clickable — which produced “Flat 2.0”, semi-flat, with shadows and depth reintroduced deliberately (Vermeulen Design, Nielsen Norman Group). The moral: a mechanical reaction to a cliché just produces the next one.

Corporate Memphis (2017-2021). The illustration style with the flat, disproportionate human figures, created by the agency Buck and spread by Facebook, became the default look of every tech company inside three years. WIRED called it “a massive homogenisation and dulling down of the internet's visual culture”. It's the closest precedent to AI slop: friendly signal → mass adoption → homogenisation → the coordinated designer backlash that killed it (AIGA Eye on Design, Creative Bloq).

Web brutalism (from around 2014). A term coined by Pascal Deville with brutalistwebsites.com, described as a youthful rebellion against the soft, corporate, crowd-pleasing styles — flat and material design first among them (UX Collective, Nielsen Norman Group).

Three cycles, one rule: the problem was never the style. It was adoption without decision. And that's exactly what you're buying, or avoiding, when you assess a proposal today.

Glassmorphism isn't banned: it's almost always badly implemented

One honest caveat, because a checklist that turns into a list of prohibitions is useless. Blurred glass panels were already declared “dead” by the designer debate back in 2023, and yet Apple and Microsoft kept using them across their own systems. The serious criticism isn't about the style itself, but about how it gets built: the legibility of text over a blurred background, and the computational cost on low-end devices — that is, the three-year-old phone half your customers open your site with (Axonix's 2026 CSS guide).

So the right question to put to the vendor isn't “did you use blurred glass?”. It's: which devices did you test it on, and what happens to the text on top when the background changes?

The backlash has a name: “Anti-AI Crafting”

Graham Sykes of Landor gave the 2026 counter-movement a name: Anti-AI Crafting, human craft as the antidote to AI's hyper-polished visual language. It is explicitly against laziness, not against technology — a distinction that matters, because the question was never “use AI or don't” (Design Magazine).

Meanwhile several AI companies have built deliberately anti-generic identities: Normal Computing with an assertive red sans and pixel detailing, Faculty with an ambigram monogram in a carved serif, Cohere with an organic, cell-like identity. The people selling AI, in other words, are the first ones who don't want to look AI-made (ebaqdesign, freelogodesign).

The Mistral logo: why “flat colour” was a decision

A concrete, dated example of what “chosen, not inherited” means. In February 2025 Mistral AI adopted a new mark designed by Sylvain Boyer Studio: a modular, blocky “M” in solid flat colour, explicitly dropping the gradients. The stated rationale isn't aesthetic but practical — more contrast and simpler use across every medium, from the favicon to signage. Hidden in the mark is a pixel cat, a nod to the models' early training datasets (designyourway, logos-world; the change is also on record at Brand New).

One limit worth stating plainly: this concerns the mark, and it doesn't imply a website has to be free of colour or of gradients. It isn't “gradients are banned”. It's that this gradient, in this mark, was removed for a reason someone can state out loud. And a reason someone can state out loud is exactly what you're looking for in a proposal.

The questions to ask the vendor before you sign

Everything above boils down to a single criterion, well put by 925studios: distinctiveness comes from making the decision, not from inheriting it. You don't have to judge whether you like a proposal — you have to work out whether anyone decided. These eight questions reveal it inside one meeting, and none of them requires technical knowledge:

• Why this colour and not another? An answer that talks about your sector, your competitors or your materials is a decision. “It works well / it's modern” isn't.
• Who chose the typefaces, and why two voices instead of one? A deliberate pairing — a typeface with personality for headings, a neutral readable one for body text — is a sign of work done. One typeface for everything is, as a rule, the default.
• What, in this identity, is specific to my company? If the answer holds identically for any other client, you aren't buying an identity.
• Show me the alternatives you discarded. Whoever decided has some; whoever generated usually doesn't. It's the single most discriminating question of all.
• Which devices did you test the blurs and shadows on? It turns an aesthetic question into a verifiable one.
• Does the copy say anything a competitor of mine couldn't paste verbatim? Actually try it: read the headline with your name swapped for theirs. If it still holds up, it's generic copy.
• What constraints did you set yourselves, and who wrote them? Palette, hierarchy, spacing: if the vendor uses generative tools — perfectly legitimate — they should be able to show you the rules they briefed them with.
• In two years, will this site look dated to 2026? An uncomfortable and useful question: every visual wave carries its own date stamped on it.

None of these questions requires you to tell a serif from a sans. It's the same skill you use to assess a quote from any other supplier: telling apart the ones who thought about your problem from the ones who handed you their own default output. In design, since 2024, that default output is recognisable with the naked eye — and now you know what it's called.