AI Design Tools Don't Generate Creative Work — They Generate Averaged Work

Every Midjourney portrait from 2023 looked like it came from the same photographer who had never met a real person. Pores rendered in obsessive detail. Light falling at the same three-quarter angle. Eyes with that particular hyperreal sheen — technically perfect, structurally identical, aesthetically hollow. You could tell the tool, not the intent.

That was the tell. And most people celebrated it.

The Approval Loop Is Not Taste

Here is what diffusion models actually do, stripped of the marketing language: they learn a probability distribution over visual space from training data. When you prompt one, it samples from that distribution, weighted toward regions of high density — areas where a lot of training images clustered together.

Training data is not a neutral cross-section of all visual work ever made. It is weighted heavily toward images that were shared, published, approved, clicked, and copied. Viral images. Stock photography that licensed well. Design work that won awards and got featured on Behance. Brand identities that got written up in Fast Company.

In other words: work that already won approval.

The model learns what approval looks like. Then it generates approval-shaped outputs. This is not a flaw in the system — it is the system working exactly as designed. It is also why every output trends toward the center.

The statistical term for what happens is regression to the mean. In visual terms, it means: when you use a diffusion model to design something, you are generating a weighted average of every design that was ever popular enough to make it into the training set. The output is not derivative in the way that a student copying a master is derivative. It is derivative in a deeper way — it has no master to copy from. It has only the aggregate.

The Midjourney Portrait Problem

In 2023 and into 2024, a specific aesthetic saturated social feeds: the hyperreal AI portrait. Skin with pore-level resolution. Bokeh backgrounds that no lens would actually produce at that focal length. Perfect bilateral symmetry in faces. A softness in the light that felt post-processed even when nothing had been post-processed.

Every version of this image looked like every other version. Users with wildly different intentions — fashion photographers experimenting, UX designers generating personas, brand teams building campaign imagery — all arrived at the same aesthetic because they were all prompting the same model. The model had learned that this kind of image was what got shared. So it generated it, again, regardless of what you asked for.

The interesting thing was not that the images looked similar. It was that the images looked similar across domains. The AI-generated startup founder looked like the AI-generated luxury model looked like the AI-generated medical professional — different prompts, different contexts, same underlying gravity pulling everything toward the center of what approval looked like in the training data.

Some photographers noticed before designers did. The hyperreal AI portrait was identifiable not because it looked fake but because it looked like a specific kind of real — a statistically optimized, uncanny real that no actual photographer would choose.

The aesthetic was not anyone's point of view. It was the absence of point of view.

The Canva-ification of Marketing Design

Canva did not start this problem, but it scaled it in a way worth naming precisely.

Template-based design tools select winners. A template gets featured when it performs well — when users who select it produce work that gets shared, downloaded, or duplicated by other users. The template ranking system is an approval loop: good performance → more visibility → more use → more data confirming good performance. Templates that encode the most broadly acceptable aesthetic rise to the top and stay there.

The result, visible across any industry's LinkedIn presence or email marketing archive from 2022 onward: a convergence toward the same set of fonts (Inter, Poppins, Playfair Display in combination), the same layouts (hero image left, text right, a single CTA button in a coral or sage accent), the same color palettes (the dusty muted tones that soft-launched from a certain set of wellness brands and then spread everywhere).

None of these choices are wrong in isolation. Inter is a good typeface. A left-aligned hero image works. Dusty sage is a coherent palette choice. The problem is that when every brand makes the same coherent choices through the same approval-optimized tools, the coherence stops meaning anything. Distinctiveness is a relative property. You cannot be distinctive by doing what the approval loop tells you to do.

This is the Canva-ification of marketing design: not that the work is bad, but that it is indistinguishable. A 2024 analysis by design research firm Lippincott found that brand recall dropped significantly for companies whose visual identities fell within what they called "category conventions" — the aesthetic norms that approval loops reinforce within any given industry. Brands that scored high on distinctiveness had recall rates roughly three times higher than brands scoring in the convention band. The work that wins the approval loop does not win the market.

What AI-Generated Brand Identity Actually Looks Like

Imagine a fintech startup — call them Klave — that launches in 2025 with a brand identity described as "AI-assisted." The logo is a geometric mark: two intersecting arcs forming a shape that suggests both a K and a sense of motion. The wordmark is set in a geometric sans with slightly rounded terminals. The primary palette is a deep navy with a gradient accent in violet-to-teal. The typography pairing is that geometric sans for UI copy and a humanist serif for marketing headlines.

Every element of this identity is a correct choice. Nothing is wrong with it. The geometric mark, the rounded terminals, the gradient accent in the blue-to-violet range, the humanist serif headline — these are all defensible decisions, each one backed by approval data showing that similar choices have tested well.

The problem is that this brand description matches, almost exactly, two dozen other fintech brands that launched in the same period. Different company names, same gradient range, same mark geometry, same type pairing. Same center of mass.

That is not what a brand identity is for. A brand identity is for creating a persistent, distinctive signal in a noisy market. When the tool that generates your identity is optimized for what has historically received approval in your category, it will generate the category average. The average is not a signal. It is the noise.

So Actually, the Tool Is Doing Exactly What You Asked

Here is the reframe that most enthusiasm about AI design tools avoids.

The problem is not that the tools are low quality. Many of them produce outputs of remarkable technical quality — more technically accomplished than most designers could produce in the same time. The problem is what "quality" means when the optimization target is approval.

If you use a diffusion model to generate imagery, you are asking: what image would have received the most approval from the population of people who shared, liked, and pinned images in this domain? That is the question it answers. If that is the question you want answered — if you want to understand the category center, identify conventions, or generate placeholder work — these tools are excellent.

If you want to ask a different question — what visual decision has not been made here, what aesthetic position is distinctive in this specific context, what would this brand look like if it had an actual point of view — the tool cannot answer that. Not because the technology is immature, but because the question is not about historical approval. It is about something that has not yet been approved.

Distinctiveness, by definition, does not exist in training data at high density. The work that eventually changes what is normal has to start somewhere outside the center. AI tools cannot find that place for you. They are optimized for the center. The farther from the center the better the work needs to be, the less useful the tool becomes.

Using the Tool Without Being Flattened by It

None of this means the tools are not useful. They are. Speed, ideation, iteration — the practical cases for AI-assisted design are real and the productivity arguments are not wrong.

The discipline is knowing what the tool optimizes for and designing your workflow around that constraint.

Use diffusion models for exploration of known aesthetic territory — generating references, surfacing category conventions, understanding what the center looks like so you can deliberately move away from it. Do not use them to answer the question "what should this look like?" That question requires a point of view the tool does not have.

Build your own reference systems that sit outside the approval loop. Physical archives, obscure digital collections, art historical references, industry-adjacent work that never made it into viral circulation. The training data is the internet's greatest hits. Your job as a designer is to know what exists outside the greatest-hits list.

Test for distinctiveness explicitly. Design critic and educator Steven Heller has noted that the question most brand identities never ask is "distinctive from what?" Before approving any visual direction, map it against the category. If it clusters with the category center, that is information, not a pass.

The tools will keep getting faster, cheaper, and more capable. The aesthetic gravity will not go away — it will get stronger as more output floods the training sets. The designers who matter will be the ones who understand what the gravity is pulling toward and choose, deliberately, to pull somewhere else.

The best design work was always the record of a specific set of decisions made by someone with a specific point of view. AI tools do not eliminate the need for that. They make it the only thing that counts.