The giveaways that scream "AI made this"

Purple gradients and bento boxes are just two examples. Design folks call these tired conventions anti-patterns, or design smells. To understand what's new here, it helps to see the full list of smells on the table.

Three months back, Design Arena analyzed GPT-5.5 and drew up a list of these recurring tells. The main ones:

Individually, none of these are inherently wrong. The problem is that AI models lean on them so heavily that readers instantly clock them. In Design Arena's blind tests, pages carrying these smells were consistently the ones humans rejected.

Smell #1: Purple-blue gradient. Pages like this routinely lost in blind judging. Source: Design Arena

Smell #2: Grid background. Source: Design Arena

Turning 1,000 Web Pages into a Map — and Finding Missing Pieces

"Good taste" and "bad taste" can feel subjective. So Design Arena wanted visible proof: they plotted all 1,000 of GPT-5.6 Sol's generated pages as points on a map.

The process had two steps. First, each page screenshot was fed into a model called CLIP, which outputs a long string of numbers for every image. Think of it as a genetic code for visual style: similar-looking images get similar codes. Second, a technique called UMAP flattened those long codes into points on a 2D plane, preserving the original distances as best it could. Pages with similar styles ended up as neighboring points on the map.

Laid flat, those 1,000 points form the model's design space — the full range of styles it likes to generate. And there, reviewers spotted something unexpected: several obvious voids, like chunks missing from a point cloud.

Below are the actual projections from Design Arena. The first is GPT-5.6 Sol — you can spot the gaps in the dense point field. The second is GPT-5.5, where the cloud is uniformly filled.

GPT-5.6 Sol's UMAP projection shows obvious voids in the point cloud. Source: Design Arena

The same projection applied to GPT-5.5: fully populated, no such gaps. Source: Design Arena

To confirm what was in those holes, reviewers overlaid outputs from both models on a single plot: GPT-5.6 Sol's points in orange on top of GPT-5.5's points. Where orange dots appear, GPT-5.6 Sol ventures. Where only the base color remains, that's what it avoids but its predecessor visits.

Orange points (GPT-5.6 Sol) overlaid on blue (GPT-5.5). The two overlap broadly, except for the purple-gradient cluster, which is completely free of orange — a region GPT-5.6 Sol entirely avoids. Source: Design Arena

The two models overlap in most regions, except for the purple-gradient cluster, where there are zero orange points. The same pattern holds for bento boxes, oversized headlines, and offset layouts. The holes are precisely where the AI tells live.

Avoiding the Same Bad Habits by Two Very Different Routes

GPT-5.6 Sol isn't the only model that avoids these clichés. But the way it does is different — and it's the most counterintuitive finding here.

Take GLM-5.2, which also ranks high. It rarely produces oversized headlines or other tells, but its method is simple: it learned from a batch of high-scoring templates that never contained those patterns in the first place. Its design space never had a "purple gradient" region, so it can't generate one, and no hole appears on its map — because that area was always empty.

GPT-5.6 Sol is a different story. The holes on its map suggest it has learned these patterns and can draw them — it just steers around them every time. It knows what a purple gradient looks like, knows where to place it, and then decides not to.

GLM-5.2 avoids bad habits through templates; that region simply isn't in its design space, so no hole appears. Source: Design Arena

Both can produce pages that dodge clichés. The difference is the shape of capability: one fenced off that land and labeled it "don't go here"; the other never mapped it at all. Design Arena suggests GPT-5.6 Sol's behavior is closer to learning followed by active suppression — a rarer trait among current models.

Starting from Reliable Templates, Then Deviating Significantly

Here's Design Arena's second finding, about personalization. Web-design models tend to fall into two camps: heavily template-dependent (stable but repetitive) or nearly template-free (varied but less reliable). GPT-5.6 Sol sits between them.

It starts with proven structures, then makes substantial adjustments for each brief. Under the same template archetype, it grows a family of related but distinct variations. Design Arena's metaphor: like bacteria evolving into closely related strains — a shared foundation, each branching its own way.

For reference points at the extremes: GLM-5.2 scored well this cycle, relying on a set of high-scoring templates; reliable, but variation mostly comes from the brief itself, and the same templates recur frequently. Claude Fable 5, by contrast, shows almost no template traces. Its design space is more dispersed, and each output is highly tailored to the brief.

GLM-5.2's outputs show visible template patterns and look similar to each other. Source: Design Arena

Claude Fable 5 shows almost no templating; design space is more spread out. Source: Design Arena

GPT-5.6 Sol works the middle ground: templates protect the floor, customization creates the differentiation. Reviewers believe this is a big reason for its high ranking — users get a page that fits the brief and still feels professionally crafted. One tell: when assigning images across different pages, it often reuses the same image in several different contexts.

Confetti Everywhere, Charts Still Weak

Active avoidance isn't a superpower GPT-5.6 Sol applies everywhere. Reviewers flagged two clear weak points.

First, confetti. It loves scattering confetti animations across pages — present in over 26.5% of its outputs. It even hand-rolls its own confetti library from scratch just to use it. That in itself is an AI tell, and this one it didn't avoid.

GPT-5.6 Sol overuses confetti, present in over a quarter of its outputs. Source: Design Arena

Second, data charts. Its performance on charts and data visualization is notably weaker — with chart.js (a common web charting library), it struggles to produce a decent real-world chart.

GPT-5.6 Sol is weaker at charts and data visualization. Source: Design Arena

More Than Twice as Fast as the Old Leader, at Half the Price

Beyond ranking and taste, GPT-5.6 Sol also wins on speed and cost. Design Arena says it sets a new Pareto frontier on both "quality vs. speed" and "quality vs. price."

GPT-5.6 Sol sets a new Pareto frontier on both "quality vs. speed" and "quality vs. price." Source: Design Arena

More Discerning, and More Adaptable

Design Arena sums up its findings with two observations.

First, it's more discerning: it's learned which patterns make a page instantly read as AI-generated, and actively suppresses them — while keeping reliable structures in reserve. Second, it's more adaptable. It merges templates with customization: templates as a safety net, heavy adaptation per brief, so the result feels both tailored and professional.

Together, those two traits form the core of Design Arena's explanation for its single-round lead. As for what it hasn't dodged — the confetti, the weak charts — that's in the report too. The full picture requires both sides.

GLM-5.2 is like a model that never learned these design smells, so it can't produce them. GPT-5.6 Sol is like a model that learned them, knows they exist, and refuses to draw them. Design Arena, analysis of GPT-5.6 Sol

The biggest difference, after digging in: it learned the clichés but actively chose not to put them on the page.

But "avoiding clichés" is still subjective. Reviewers wanted visible proof.

Here's how they found the proof: they turned all 1,000 of GPT-5.6 Sol's generated pages into points on a map.

Two steps. First, a model called CLIP gives each page a string of numbers (images with similar style get similar codes); second, a method called UMAP flattens those long codes into points on a plane, placing similar styles nearby. With 1,000 points laid out, reviewers saw something odd: a few obvious missing chunks.

Reviewers overlaid both models' outputs: in the purple-gradient cluster, GPT-5.6 Sol doesn't have a single point. Same story for bento boxes and oversized headlines. Those holes are exactly the AI tells. GLM-5.2 avoids them differently: it never learned the patterns, so that region was always empty. One is "can't"; the other is "can, but won't."

The same map also showed a second quality: the points aren't rigidly bunched. It starts from a set of good templates, then heavily customizes each brief, landing between cookie-cutter and chaotic.

Beyond taste, GPT-5.6 Sol also pulls ahead on speed and price.

It also didn't dodge everything: confetti shows up in over a quarter of its outputs — and if no library is provided, it writes its own; data charts are also a weak spot, with chart.js (a common web chart library) producing lackluster results.