Every time a tool democratises visual production, we go through a phase of collective ugliness, backlash, and stabilisation. It happened with WordArt, with PowerPoint and Canva templates, and now with generative AI. Only this time it’s not just a visual problem — it’s a data bias problem too.

On the surface, the AI aesthetic crisis looks a lot like when Word introduced pre-laid-out templates for all kinds of stationery — the famous WordArt, which sparked the Desktop Publishing Revolution. I still remember the moment I discovered the magic of curved titles for my school project covers and thought I was killing it — until my teacher docked my grade because she found it ugly and preferred my handmade posters. That may have been my first design lesson.

wordart is my passion

This tension between democratisation and quality is nothing new in design. In 1993, Steven Heller published “Cult of the Ugly” in Eye Magazine, criticising the “ugly” graphic design produced by Desktop Publishing. Ugly, yes — but adopted en masse because of the popularisation of tools like Word. Ten years later, Virginia Postrel observed in The Substance of Style the other outcome of that democratisation: that the spread of computer-aided design, despite its ugly early phase, ended up making more people sensitive to visual quality. The pattern seems to repeat itself: a new tool democratises production, goes through a phase of collective ugliness, and eventually matures — not because the tool improves, but because people develop taste.

Cut to 2026, and AI is playing much the same role. What used to be a PowerPoint collage with a landscape photo and Comic Sans text is now a Gemini-generated image — the tool changed, but the impulse and the result are the same. At the corporate level, I’ve been watching developers, PMs and data analysts explore new visual horizons with generative AI. In one sense, I find this interesting: it brings the two worlds — visual and written — closer together, giving people who primarily work with text and code a way to communicate visually. Part of my work as a design lead has involved asking the team to include brand guidelines in their AI prompts, because otherwise everything ends up looking like it belongs to whichever company owns the models.

But the tension today is more complex than just an ugly aesthetic — it’s also a cultural issue deeply embedded in the data economy.

LAION-5B, one of the largest datasets used to train image models, contains 89% Western cultural artefacts versus 12.7% non-Western, with 78% of English-language sources from North America and Europe and only 6% from Africa or Asia (Said et al., 2025). OpenAI itself acknowledged in internal documents that DALL-E 3 has a “tendency toward a Western viewpoint” (Washington Post, 2023). And a Carnegie Mellon study (Taylor et al., 2026) showed that the aesthetic filter used to curate these datasets — the LAION-Aesthetics Predictor — was trained on the tastes of Anglophone photographers and Western AI enthusiasts. This filter rates landscapes and realistic portraits highly while scoring Picasso, Warhol and Dalí low.

In other words, the system that decides what is “beautiful” for training generative AI would fail most of twentieth-century art history.

Gemini likely operates with the same logic, although Google’s datasets are proprietary and less transparent. When your teams create visuals in Gemini and everything looks like it came from Google, that’s not a coincidence — the model was trained on an aesthetic that reflects who built it.

Along the same lines, a Cornell study (Agarwal et al., 2025) showed that the homogenisation mechanism works in text too: Indian participants who used AI writing suggestions ended up writing more like American English — but the reverse did not happen. AI pushed Indian writing toward Western patterns, changing not just what was written but how. Most worryingly: the more users accept these suggestions, the more online content conforms to Western norms, feeding the training data for future models. The bias feeds itself.

The homogenisation is not just aesthetic, it’s cultural. And it’s not accidental — it’s structural.

In my day-to-day work as a design lead, I’ve been trying to use this context to build AI-agnostic design processes. I created AI usage guidelines for the team and established a Human-in-the-loop system at critical points: key artefacts need human approval before moving forward. What counts as a key artefact depends on the context. I almost always define these through a series of participatory workshops with the team, drawing on a historical analysis of what the company has delivered, what was contracted, and what we want to deliver in the future — in terms of both code and design. I also introduced a gatekeeping step in the development process: no code gets opened without first properly exploring the problem.

In practice, the effort is singular: getting people not to swap their critical judgment for AI convenience, but to use the tool for what it does well. Because if the historical pattern teaches us anything — from WordArt to Canva — it’s that maturation never came from the tool. It came from people who learned to use it with taste. This time, taste includes knowing where the data comes from and making a conscious effort to mitigate the negative effects of that bias.

References

Heller, S. (1993). “Cult of the Ugly.” Eye Magazine, Vol. 3, No. 9.

Postrel, V. (2003). The Substance of Style: How the Rise of Aesthetic Value Is Remaking Commerce, Culture, and Consciousness. HarperCollins.

Said, M.N. et al. (2025). “Deconstructing Bias: A Multifaceted Framework for Diagnosing Cultural and Compositional Inequities in Text-to-Image Generative Models.” ICLR 2025 Workshop on Synthetic Data. arxiv.org/abs/2505.01430

Taylor, J. et al. (2026). “The Algorithmic Gaze of Image Quality Assessment: An Audit and Trace Ethnography of the LAION-Aesthetics Predictor.” Carnegie Mellon University. arxiv.org/abs/2601.09896

Agarwal, D., Naaman, M. & Vashistha, A. (2025). “AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances.” CHI 2025. Cornell University. arxiv.org/abs/2409.11360

Tiku, N. (2023). “AI generated images are biased, showing the world through stereotypes.” The Washington Post. washingtonpost.com