The Next Default: Why Every Anti-AI Design Rule Becomes a Style

The first video is twenty-six seconds of rapid-fire diagnosis. "Purple to blue gradient, gradient hero text, emojis in your headings, Inter font everywhere, colored left-border cards, glassmorphism cards, low contrast, dark mode. Three icon boxes in a row, a badge above the headline. Lucide icons everywhere. Untouched shadcn/ui components. Sections fading in as you scroll." It ends on em dashes, buzzwords, "serif italics for accent words," the fonts Space Grotesk and Instrument Serif, and grain texture. The caption asks whether your website or app has any of these. It has been saved more than thirty thousand times.

The second video, by the linguist Adam Aleksic, says the cure is already happening. "The pastel influencer font is on its way out and we're starting to see a lot more ugly fonts that deliberately clash with the video itself." These fonts "serve to subvert the algorithmically optimized tasteslop that's been flooding our feeds with something that instead feels transgressively raw and human." Then a warning: every ugly style gets standardized, takes on "a corporatized quality," and gets abandoned for the next one, "until the slop factories eventually catch on and we're once again forced to keep moving."

I recognized both lists, because I keep one. I work with an AI model on most of my design and writing, and I have a file of rules telling it which visual defaults to avoid. The research behind these two videos turned out to be about lists like mine, and about what happens to them. The look of AI-built websites came from a handful of tool defaults. The rules the AI companies wrote to escape that look seeded the next look, in under a year. And the mechanism is older than AI by more than a century.

1. Two videos, checked

The checklist is accurate as a description and incomplete as an explanation. Most of its items are real defaults of AI website builders, documented by the companies that make them. Some are not AI tells at all: low-contrast text appears on 83.9% of the million most visited home pages, and it was just as common in 2019 (WebAIM 2026). And the list is a snapshot of something that moves. The checklist's creator, Michael Ly, makes content about AI workflows and lists partnerships with AI app builders; the video compiles a list designers were already circulating that spring.

Aleksic's theory is mostly older than the trend it explains. In 2014, the writer Nick Douglas argued that the "Internet Ugly" aesthetic is "an imposition of messy humanity" on a web of smooth gradients and retouching, and the aesthetic "that best defines the internet against all other media" (Douglas 2014, p. 315). Douglas described default settings "locking in the aesthetic," co-optation as "part of the aesthetic cycle," and old ugliness returning through nostalgia (pp. 331–334). What the video adds is the 2026 objects, and those are unmeasured. I found no data on caption fonts on TikTok or Instagram, no named "pastel influencer font," and no use of "fontmaxxing" before this video. The word "tasteslop" comes from the trend forecaster Emily Segal, who coined it in May 2026 for tasteful signifiers turned into templates (NEMESIS 2026).

2. Where the AI look comes from

Most of the first-wave tells trace to one small set of tools. Tailwind, the most widely used styling framework for this kind of site, documents gradients with indigo as the most frequent color stop, and documents the recipe for gradient-filled headline text. Its paid component library, Tailwind UI, made its buttons indigo. In August 2025, Tailwind's co-creator Adam Wathan posted a mock apology for having made every Tailwind UI button `bg-indigo-500` five years earlier, "leading to every AI generated UI on earth also being indigo" (Wathan 2025). The component library shadcn/ui made Lucide its default icon set and ships bordered cards with soft shadows, pill-shaped badges and buttons that fade on hover. Inter was the default font in shadcn's setup instructions and in the official Next.js starter template from 2023 to 2024.

AI builders install this stack before the model writes a line. Anthropic's own skill for building web artifacts lists Tailwind and "40+ shadcn/ui components pre-installed" (Anthropic). OpenAI's prompting guide for GPT-5, published in August 2025, recommended Tailwind, shadcn/ui, Lucide, the animation library Motion, and Inter for new apps (OpenAI 2025). Public code then holds thousands of copies of those templates: one hero layout, with its pill-shaped announcement badge, appears in about 1,880 public files on GitHub.

Anthropic's engineers gave the mechanism a name in November 2025: "distributional convergence." Asked for a landing page without guidance, a model samples from the most common patterns in its training data, and the safe choices that offend no one dominate that data (Anthropic 2025). Researchers describe a related effect, in which tuning a model on human preferences narrows its output because raters favor what is familiar (Zhang et al. 2025). The look is partly a product of people rewarding the typical.

3. The makers' confession

Between October 2025 and September 2026, the companies behind the tools said so in writing. Anthropic's artifacts skill warned against "what is often referred to as 'AI slop'": "excessive centered layouts, purple gradients, uniform rounded corners, and Inter font." Its frontend-design skill, published in November 2025, told Claude to "NEVER use generic AI-generated aesthetics like overused font families (Inter, Roboto, Arial, system fonts), cliched color schemes (particularly purple gradients on white backgrounds)" (Anthropic, frontend-design). A week later, OpenAI's coding agent was told to "avoid purple-on-white defaults" and to avoid "default stacks (Inter, Roboto, Arial, system)," three months after OpenAI's own guide had recommended Inter (OpenAI Codex). In December 2025, shadcn/ui's changelog said it plainly: "But somewhere along the way, all apps started looking the same. I guess the defaults were a little too good. My bad" (shadcn/ui 2025).

The confessions share a structure. Each treats the generic look as a prompting problem, solved by telling the model what to avoid, layered over a stack the same companies keep shipping. That is where the second wave came from.

4. The fix becomes the look

Anthropic's anti-slop prompt of October 2025 did more than ban Inter and purple. It offered replacements. It suggested "Impact choices" such as JetBrains Mono, Space Grotesk, Playfair Display, Fraunces and Bricolage Grotesque, recommended pairing principles including "serif + geometric sans," and its companion skill told the model to "pair a distinctive display font with a refined body font" and to apply "custom cursors, and grain overlays." The same prompt admitted the problem it was trying to solve: "You still tend to converge on common choices (Space Grotesk, for example) across generations" (Anthropic cookbook).

Five months later, a developer named Adrian Krebs built a detector for AI design patterns and scored about 1,600 projects submitted to Hacker News's "Show HN." Its list of "templated display fonts" included Space Grotesk, Instrument Serif, Fraunces and Bricolage Grotesque (Krebs 2026). Google Fonts' public statistics suggest several of them are growing fast. By June 2026, Anthropic's own skill described AI-generated design as clustering around new looks: a warm cream background with a high-contrast serif and a terracotta accent; a near-black background with one bright accent; a broadsheet layout of hairline rules and dense columns. By September it listed, among "the commonest tells of a generated page," "accenting just a single word or phrase in a headline, like putting one word in italic/bold or a different color," and noted that "fade-and-slide-up entrances on each section and hover transitions on every card are the generic default and read as AI-generated." It added that the terracotta accent near #D97757 is Claude's own interface color.

So the checklist's second half, the serif italics, Instrument Serif, Space Grotesk and grain, describes in large part what the anti-slop instructions recommended a year earlier. The timing fits and the fonts fit; it is still correlation, since human fashion and template marketplaces pushed the same fonts. The first half has already moved on. In Krebs's spring 2026 sample, permanent dark mode was the most common pattern, on 34.7% of pages, and a purple accent one of the rarest, on 5.4%. The checklist leads with the tell that had mostly gone.

5. An old machine

None of this needed AI. Georg Simmel described the whole cycle in a 1904 essay. Fashion, the essay says, "is the imitation of a given example," and at the same time it satisfies "the need of differentiation, the tendency towards dissimilarity, the desire for change and contrast" (Simmel 1904, p. 133). It dies by success: "As fashion spreads, it gradually goes to its doom" (pp. 138–139). The essay even explains the ugly fonts: "Judging from the ugly and repugnant things that are sometimes in vogue, it would seem as though fashion were desirous of exhibiting its power by getting us to adopt the most atrocious things for its sake alone" (p. 134). And it sees through the people who refuse fashion: their "conscious neglect of fashion represents similar imitation, but under an inverse sign" (p. 142).

Experiments have since filled in the mechanism. People abandon a taste when others adopt it, especially in domains that signal identity, such as music, hairstyles and clothes (Berger and Heath 2007). Names that rise fast die fast, because fast adoption marks them as fads (Berger and Le Mens 2009). A mathematical model shows that people trying to be different end up synchronized when they react to the same trends with the same delay (Touboul 2019). When MIT Technology Review covered that "hipster effect" paper in 2019, a man wrote to complain that the bearded, beanie-wearing hipster in its stock photo was him, used without consent; Getty's records showed it was someone else.

The trend industry has lived the cycle in public. In 2013 the forecasting group K-HOLE, which included Emily Segal, argued that real uniqueness had become nearly impossible and that blending in could be the new freedom; "normcore" was being sold as a style within months. Thirteen years later, Segal coined the word for what tasteful signifiers become once everyone uses them. Douglas had predicted the next turn too. Better tools will not end ugliness, Douglas wrote: someone will misuse them to make "a freshly ugly kind of image" (p. 316).

6. Why the sameness happens anyway

The web was converging before any of these tools existed. A study of 227,802 homepage snapshots from 10,482 websites found that sites grew steadily more alike after about 2007, with layouts more than 30% more similar by 2019 (Goree et al. 2021). The strongest link was shared code libraries, Bootstrap above all, along with the need to work on phones, search-engine rules and slow redesigns negotiated by committees. The AI look is the 2020s version of the Bootstrap effect: the same mechanism, with a model choosing the defaults.

Generative AI adds a second layer. When writers were given story ideas from GPT-4, their individual stories were rated more creative, and the stories as a group became more similar to each other (Doshi and Hauser 2024). A 2025 meta-analysis put the loss of variety in human-AI collaboration at a large standardized effect (Holzner et al. 2025). An industry benchmark from 2026 found AI-generated interface designs much more alike than human ones, and competent: every model met more than 90% of each brief's requirements (Rusia et al. 2026). Each user is better off, which is why the sameness spreads.

Some of the sameness is useful. The authors of the web study refused to call it simply bad, because familiar layouts help people use and access a site. Strategy research makes the same point about organizations: conforming shows you belong to a category, differing reduces competition, and firms do best "as different as legitimately possible" (Deephouse 1999). A website that looked like nothing else would also be harder to use.

7. When ugly works

Aleksic's escape, deliberate ugliness, works under conditions the research has mapped. Breaking a norm can raise perceived status and competence, but only when it looks intentional and the audience knows the norm being broken (Bellezza et al. 2014). Autonomy reads as cool when the deviation is small or occasional and breaks a norm people already find unnecessary (Warren and Campbell 2014). Low production quality earned user-made videos more trust than agency videos; high quality made no difference (Hautz et al. 2014). Roughness helps peers more than companies.

And the raw look has been manufactured for years. Crystal Abidin's term for it is "calibrated amateurism": deliberate work to produce the look of an amateur (Abidin 2017). TikTok has asked brands for lo-fi "realness" since 2021. Charli XCX's Brat, the video's own example of the lineage, was designed by a professional studio over five months to feel "off-putting and kind of garish," in stretched, low-resolution Arial. Ugliness is one cheap way of saying a person made this, and it is cheap to copy, including for a model that can be asked for clashing type.

For a brand, the trend itself is the problem. Brand-asset research judges a cue by fame and uniqueness: how many people link it to the brand, and whether they link it to that brand alone. A style everyone moves to is famous and shared. In a study of UK car brands, logos scored 0.76 on ownership and fonts 0.08, the lowest of the four asset types tested (Fiocchi and Esfahani 2024). Unusual designs are liked less at first and more after repeated exposure (Landwehr et al. 2013). A brand that cycles through ugly styles never stays long enough to collect.

8. My own list

I keep a file called design-rules. Its stated purpose is how the AI I work with "avoids the visual defaults that make work read as AI-made." Its core rule is good, and I stand by it: "Design from a reason, not a reflex." The test it applies to every element is "chosen, or defaulted?"

The rest of the file is a list, and the list reads like the anti-slop canon. No italic serif word inside a sans headline. No reflexive pairing of Instrument Serif, Playfair or Fraunces over Inter. No Inter as the unexamined default. No gradient headline text. No three-card rows, no announcement pill above the hero, no purple or indigo gradient, no glassmorphism, no mesh gradient with grain, no Lucide icons as the whole icon system, no fade-up on every element. My writing rules, its sibling file, warn that AI overuses em dashes, and the essays in this series avoid them entirely. Read against the research, my file is the same instrument as Anthropic's skill, written for the same reason, with the same weakness: every item is something to avoid, and avoidance is a direction the next default can follow. Simmel would call it imitation under an inverse sign. The em dash shows how fast it happens: newer models have nearly dropped it, so a text without em dashes no longer says anything about who wrote it (Freeburg 2026).

Then there are the covers of this series. Since October 2026, every cover from the sixth essay on has used a near-black ground, pale cream line work and a single coral accent, generated by an image model from a written guide. My own rules warn against "dark mode plus one neon accent as the automatic 'premium' look." Anthropic's 2026 skill names a near-black background with one bright accent as an AI cluster, and a cream palette with a terracotta accent as another, noting that Claude's own interface color sits near #D97757. My coral, #f06c4e, sits between those two descriptions. The research notes for this essay were compiled with an Anthropic model, which is how I learned this.

I do not think this makes the covers wrong, and the essay's evidence says why. A palette held consistently across a series, attached to one method, can become an owned asset; the same palette chosen because it looks "premium" is a default. The covers will be the test of which one mine is. In The Margin of Crudeness I argued that the look of labor became free once machines could produce it. This essay is the uncomfortable sequel: the look of avoiding the machine became free too.

9. Against myself

The strongest objection is that the tells matter only to people like me. No study I found tests whether ordinary visitors can recognize an AI-built website by its look. The documented trust penalty comes from disclosure: people trust content less once told that AI made it (Schilke and Reimann 2025). The people who see the tells are designers and developers, the way designers once spotted Bootstrap. For most visitors, a competent template may be exactly what a competent site should look like, and the web study's usability argument says familiarity has value.

Lists also work. A rubric brought novices' design critique close to experts' in a 2016 study (Yuan et al. 2016), and a checklist is how a team shares judgment it cannot otherwise transmit. The problem the research points to is the moment the list replaces the reason, and that is a matter of use, not a reason to throw the list away.

And the evidence on ugliness is thinner than the confidence on both sides. Nobody has tested clashing typography against polished design on recall, trust or authenticity. A font designed to be hard to read did not help people remember what they read (Taylor et al. 2020), which argues against one popular defense of ugly type, and is all I found.

10. At the desk, and how this could be wrong

For my own work, the conclusion is to keep the core rule and demote the list. "Chosen, or defaulted?" is the right question because it has no content: it cannot be copied into a style. The list is useful as a scan for reflexes, and dangerous as a description of good taste. The rest is what brand research has said for years. Choose a few assets that belong to the project, hold them long enough to be recognized, and aim to be as different as legitimately possible, which is less different than a trend and more stable.

The argument could be wrong in three specific ways. First, a repeat of the 2021 web study on sites built from 2023 to 2026 could find that AI-era websites are no more alike than the Bootstrap-era web, which would make the tells a story told by designers more than a measured change. Second, an experiment could find that deliberately ugly design improves brand recall or perceived authenticity over time, for brands as well as for individual creators, which would make the ugly cycle a strategy rather than a symptom. Third, if model makers find a way to produce variety that does not route through lists of things to avoid, the next default may never form.

Until then, I read both videos as accurate reports from inside the machine they describe. The checklist will be out of date within a year, and the fonts it names will be replaced by whatever the next round of advice recommends. The only thing on it that cannot become a style is the question it implies: did someone choose this, or did it arrive by default?

References

• Abidin, C. (2017). #familygoals: Family influencers, calibrated amateurism, and justifying young digital labor. Social Media + Society, 3(2). https://doi.org/10.1177/2056305117707191
• Anthropic. (2025). web-artifacts-builder skill. https://github.com/anthropics/skills/blob/main/skills/web-artifacts-builder/SKILL.md
• Anthropic. (2025–2026). frontend-design skill (versions of November 12, 2025; June 9, 2026; September 3, 2026). https://github.com/anthropics/skills/blob/main/skills/frontend-design/SKILL.md
• Anthropic. (2025). Frontend aesthetics: A prompting guide. Claude cookbooks. https://github.com/anthropics/claude-cookbooks/blob/main/coding/prompting_for_frontend_aesthetics.ipynb
• Anthropic (Rajasekaran, P., Wei, J., Bricken, A., et al.). (2025, November 12). Improving frontend design through Skills. https://claude.com/blog/improving-frontend-design-through-skills
• Bellezza, S., Gino, F., & Keinan, A. (2014). The red sneakers effect: Inferring status and competence from signals of nonconformity. Journal of Consumer Research, 41(1), 35–54. https://doi.org/10.1086/674870
• Berger, J., & Heath, C. (2007). Where consumers diverge from others: Identity signaling and product domains. Journal of Consumer Research, 34(2), 121–134. https://doi.org/10.1086/519142
• Berger, J., & Le Mens, G. (2009). How adoption speed affects the abandonment of cultural tastes. PNAS, 106(20), 8146–8150. https://doi.org/10.1073/pnas.0812647106
• Deephouse, D. L. (1999). To be different, or to be the same? It's a question (and theory) of strategic balance. Strategic Management Journal, 20(2), 147–166. https://doi.org/10.1002/%28SICI%291097-0266%28199902%2920%3A2%3C147%3A%3AAID-SMJ11%3E3.0.CO%3B2-Q
• Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290
• Douglas, N. (2014). It's supposed to look like shit: The Internet Ugly aesthetic. Journal of Visual Culture, 13(3), 314–339. https://doi.org/10.1177/1470412914544516
• Fiocchi, G., & Esfahani, M. S. (2024). Exploring the uniqueness of distinctive brand assets within the UK automotive industry. Journal of Brand Management, 31(1), 1–15. https://doi.org/10.1057/s41262-023-00329-6
• Freeburg, E. M. (2026). The last fingerprint: How markdown training shapes LLM prose. arXiv:2603.27006. https://arxiv.org/abs/2603.27006
• Goree, S., Doosti, B., Crandall, D. J., & Su, N. M. (2021). Investigating the homogenization of web design: A mixed-methods approach. CHI '21. https://doi.org/10.1145/3411764.3445156
• Hautz, J., Füller, J., Hutter, K., & Thürridl, C. (2014). Let users generate your video ads? The impact of video source and quality on consumers' perceptions and intended behaviors. Journal of Interactive Marketing, 28(1), 1–15. https://doi.org/10.1016/j.intmar.2013.06.003
• Holzner, N., Maier, S., & Feuerriegel, S. (2025). Generative AI and creativity: A systematic literature review and meta-analysis. arXiv:2505.17241. https://arxiv.org/abs/2505.17241
• Krebs, A. (2026, April 20). Scoring Show HN submissions for AI design patterns. https://www.adriankrebs.ch/blog/design-slop/
• Landwehr, J. R., Wentzel, D., & Herrmann, A. (2013). Product design for the long run: Consumer responses to typical and atypical designs at different stages of exposure. Journal of Marketing, 77(5), 92–107. https://doi.org/10.1509/jm.11.0286
• NEMESIS. (2026, May 5). Tasteslop. Nemesis Memos. https://nemesisglobal.substack.com/p/tasteslop
• OpenAI. (2025). GPT-5 prompting guide. OpenAI Cookbook. https://github.com/openai/openai-cookbook/blob/main/examples/gpt-5/gpt-5_prompting_guide.ipynb
• OpenAI. (2025). Codex system prompt for gpt-5.1-codex-max. https://github.com/openai/codex/blob/main/codex-rs/core/gpt-5.1-codex-max_prompt.md
• Rusia, A., Bhole, A., Gupta, P., & Dey, D. (2026). Design Creativity Bench: Measuring creativity in LLM-generated UI. arXiv:2610.11539. https://arxiv.org/abs/2610.11539
• Schilke, O., & Reimann, M. (2025). The transparency dilemma: How AI disclosure erodes trust. Organizational Behavior and Human Decision Processes, 188, 104405. https://doi.org/10.1016/j.obhdp.2025.104405
• shadcn/ui. (2025, December 12). Changelog. https://ui.shadcn.com/docs/changelog
• Simmel, G. (1904). Fashion. International Quarterly, 10(1), 130–155. https://archive.org/details/internationalqua10newy
• Taylor, A., Sanson, M., Burnell, R., Wade, K. A., & Garry, M. (2020). Disfluent difficulties are not desirable difficulties: The (lack of) effect of Sans Forgetica on memory. Memory, 28(7), 850–857. https://doi.org/10.1080/09658211.2020.1758726
• Touboul, J. (2019). The hipster effect: When anti-conformists all look the same. Discrete and Continuous Dynamical Systems - B, 24(8), 4379–4415. https://doi.org/10.3934/dcdsb.2019124
• Wathan, A. (2025, August 7). [Post on X about Tailwind UI's indigo buttons]. https://x.com/adamwathan/status/1953510802159219096
• Warren, C., & Campbell, M. C. (2014). What makes things cool? How autonomy influences perceived coolness. Journal of Consumer Research, 41(2), 543–563. https://doi.org/10.1086/676680
• WebAIM. (2026). The WebAIM Million. https://webaim.org/projects/million/
• Yuan, A., Luther, K., Krause, M., Vennix, S. I., Dow, S. P., & Hartmann, B. (2016). Almost an expert: The effects of rubrics and expertise on perceived value of crowdsourced design critiques. Proceedings of CSCW '16, 1005–1017. https://doi.org/10.1145/2818048.2819953
• Zhang, J., Yu, S., Chong, D., Sicilia, A., Tomz, M. R., Manning, C. D., & Shi, W. (2025). Verbalized sampling: How to mitigate mode collapse and unlock LLM diversity. arXiv:2510.01171. https://arxiv.org/abs/2510.01171

Source note: sources were checked on 7 November 2026. Anthropic, OpenAI and shadcn/ui documents were read in their public repositories and are quoted under their open licenses; Wathan's post was read through X's public data. Douglas 2014 and Simmel 1904 were read in full, with page numbers from the original printings. Goree et al., Doshi and Hauser, Berger and Le Mens, Touboul and Fiocchi and Esfahani were read in full; Bellezza, Warren and Campbell, Hautz, Abidin, Landwehr and Holzner at abstract level. Krebs's figures come from the author's post and public dataset, which has no human-made comparison group. The Brat design details come from syndicated reprints of a New York Times report. The research notes on AI-website tells were compiled with an Anthropic model, as section 8 says. Quotations from the two videos come from auto-generated captions, lightly corrected.