Signs of AI Design Slop

Notifications You must be signed in to change notification settings

Fork 0

Star 5

BranchesTags

Open more actions menu

Folders and files

NameName

Last commit message

Last commit date

Latest commit

History

2 Commits

2 Commits

.gitignore

.gitignore

CONTRIBUTING.md

CONTRIBUTING.md

LICENSE

LICENSE

README.md

README.md

SOURCES.md

SOURCES.md

Repository files navigation

A field guide to the visual tells of AI-generated design: websites, app interfaces, images, video, logos, slide decks, social graphics, illustration, and print. Modeled on Wikipedia's Signs of AI writing, which documented the tells of machine-written prose. This project does the same for machine-made visuals.

None of these signs proves anything on its own. Almost every pattern here was invented by human designers first; AI tools amplified their frequency and uniformity until they became fingerprints. One match is noise. Ten matches is a signature. Read the how to use this guide section before accusing anyone of anything.

Contributions welcome. See CONTRIBUTING.md. Content is licensed CC BY-SA 4.0: copy it, translate it, build tools on it.

Contents

How to use this guide

Part I: Websites and app interfaces

Part II: AI-generated images

Part III: AI-generated video

Part IV: Logos and branding

Part V: Slide decks and documents

Part VI: Marketing, social, and print

Part VII: Copy embedded in design

Signs that no longer work

Why models converge

Related work

How to use this guide

Each entry lists what to look for, why models produce it, and the false positives: legitimate human work that shows the same sign. Entries carry two labels.

Reliability describes how much a single match tells you. "Strong" means few humans produce it by accident (garbled letterforms, hallucinated extra limbs, placeholder text left in a shipped deck). "Moderate" means it is meaningful mostly in combination. "Weak" means humans do it constantly and the sign only matters at high density.

Status describes whether the sign still applies to current tools. Models patch their tells. Hands were the canonical image tell in 2023 and are mostly fixed in frontier models now. Purple gradients were the canonical UI tell of 2024 and are already migrating toward cream palettes and italic serif heroes. A guide like this decays; that is why it is a repository and not an article.

Three rules of interpretation:

Stack signs, never single ones. Every entry here is defensible in isolation. A dark hero with a glow is also the house style of Linear, Raycast, and a hundred human teams that copied them years before generative tools existed. The fingerprint is the co-occurrence of many defaults with no deviation anywhere.

Judge the absence of decisions, not the presence of a style. The recurring theme across every category: AI output is the statistical median of its training data. The tell is not that Inter is a bad typeface. The tell is Inter, unchosen, next to twenty other unchosen defaults.

Do not use this list to harass people. Identifying AI involvement in a design is not evidence of low effort, deception, or bad work, and misidentifying a human designer's work as AI causes real harm. Human detection accuracy on current AI images is barely above chance. Treat every conclusion as probabilistic.

Part I: Websites and app interfaces

The output of v0, Lovable, Bolt, and LLM-generated frontends (Claude, ChatGPT, Cursor builds). This is the most heavily documented category because the generators share a narrow toolchain: Tailwind CSS, shadcn/ui, Lucide icons, Framer Motion. The defaults of that toolchain, replicated across millions of tutorial pages in training data, became the AI look.

Color and gradients

Purple-to-indigo gradient

The most cited tell in the category. Indigo-to-violet hero washes and gradient buttons, typically built from Tailwind's indigo-500 (#6366F1), violet-500 (#8B5CF6), or purple-500 (#A855F7). One 2026 crawl put blue-purple gradient prevalence at roughly 78 percent of indexed marketing sites, up from about 10 percent in 2020.

Why: Tailwind shipped indigo as its demo accent in 2019. Thousands of tutorials replicated it, saturating training data, so models learned that modern web equals indigo. Generated sites then re-entered training corpora and compounded the bias.

False positives: purple is a legitimate brand color. The tell is the unchosen gradient in the exact default hue, stacked with other defaults.

Reliability: moderate. Status: active but migrating (see cream palette below).

Neon-on-dark with glow

Cyan, violet, or pink accents on near-black backgrounds, with colored box shadows behind cards and buttons (box-shadow: 0 0 20px rgba(139,92,246,0.5)).

Why: the dark-SaaS look dominates design-inspiration training data; it is the signature output of v0 and Cursor builds.

False positives: large. This is the "Linear style" that Linear, Raycast, and Reflect built by hand from around 2019, and thousands of human teams copied it before AI tools existed. A dark glowing SaaS page alone proves nothing about authorship.

Reliability: weak alone. Status: active.

Cream and beige "tasteful" palette

Warm cream page background, often with an amber accent. The second-wave default that appeared after purple gradients became a public joke.

Why: models absorbed the anti-purple backlash and learned that cream signals a tasteful startup. It became the new reflex, which is the general life cycle of these tells.

False positives: editorial, food, and craft brands use warm neutrals deliberately.

Reliability: moderate. Status: active, rising.

Emerald fallback

When a prompt bans purple, models cascade to emerald (#10B981) as the next safe accent, sometimes ignoring explicit written guidance against green.

Why: next most probable accent in the learned hierarchy.

False positives: green is standard for success states and for finance and health brands.

Reliability: moderate. Status: active.

Gradient text

Gradient fill clipped to headings and to big metric numbers, carrying no meaning.

Why: a common Tailwind demo pattern (bg-clip-text) that models pattern-match from dashboard templates.

Reliability: moderate. Status: active.

Timid, evenly weighted palettes

Several muted colors at equal visual weight, with no single dominant color and no sharp accent. Bold, committed palettes are statistically rare in output because they are statistically rare in the median of training data.

Reliability: weak. Status: active.

Typography

Inter everywhere

Inter (or a system sans) as the only typeface on the page. The tell is not the font; it is the font unchosen, appearing beside twenty other unchosen defaults.

Why: the most used interface typeface in training data and the configured default of the Tailwind and shadcn stack.

False positives: countless hand-built products use Inter deliberately, including Linear itself.

Reliability: weak alone. Status: active.

Geist, Space Grotesk, Instrument Serif

Geist has been called the new Inter: it is v0's default because it is Vercel's typeface, so it saturates v0 output. Space Grotesk is the reach when the model or the user tries to escape Inter, which made it a tell of its own. Instrument Serif is the newest reflex, usually italic, usually enormous, usually in a hero.

Reliability: moderate. Status: active.

Oversized italic serif display headline

A huge italic serif hero headline, or a single serif-italic accent word dropped inside a sans headline. The universal "tasteful AI startup" hero of 2025 and 2026, applied to products whose register does not call for it.

Why: models conflate distinctive with appropriate.

False positives: editorial brands legitimately want expressive serifs. The question is whether the serif is doing work or decoration.

Reliability: moderate. Status: active, rising.

Single family, single weight, flat hierarchy

One type family for the whole page, sometimes one weight for everything, size steps too close together to create hierarchy, letter spacing crushed on display type, all-caps section labels sprayed everywhere.

Why: font pairing and typographic scale require judgment the generator skips.

Reliability: moderate in combination. Status: active.

Monospace as decoration

Body copy or ordinary labels set in a code font for hacker atmosphere, making a marketing page read like a terminal.

False positives: developer-tool docs and terminal-adjacent products.

Reliability: moderate. Status: active.

The hero section

Eyebrow chip above the headline

A tiny uppercase tracked label, or a pill badge ("New", "v2.0 is here", a sparkle prefix), sitting directly above a centered oversized headline. A variant adds a decorative dot prefix and a trailing thin line.

Why: the default AI SaaS hero shape in training data. Models treat the ornament as editorial polish and emit it whether or not there is anything to announce.

Reliability: moderate. Status: active; one of the most recognizable current tells.

The full hero formula

Centered stack: badge, a full sentence set at 64 pixels or larger, one-line subhead, two calls to action (gradient primary reading "Get Started", ghost secondary), and a gradient orb or glow floating behind.

Why: the modal SaaS hero in training data. It is also a genuinely effective conversion layout, which is how it became modal.

False positives: the structure works and humans use it on purpose. The tell is the formula with zero deviation plus copy that could describe any product.

Reliability: moderate. Status: active.

Gradient orbs and blobs

Abstract purple or violet gradient blobs floating behind the hero, sometimes animated, sometimes following the cursor.

Reliability: moderate. Status: active.

Layout and structure

The three-card feature row

Exactly three equal cards below the hero, each with an icon on top, a short heading, and two lines of text. The single most consistently named tell across every source that documents this category.

Why: Tailwind tutorials taught grid layout with three columns; the median SaaS feature section in training data is a triptych; v0 emits it by default and its output becomes the next model's training data.

False positives: a triptych is a clean, responsive-friendly layout that humans also default to. Suspicion rises when the content visibly does not come in threes but was forced into three anyway.

Reliability: moderate. Status: active.

Bento grid

An asymmetric boxed grid, often dark with accent-colored tiles. It spread as the safe alternative once three-card rows were called out, and is now a tell of its own.

Reliability: moderate. Status: active.

Icon tile stacked above heading

A small rounded-square icon container above a heading: the universal AI feature-card template, emitted in this exact shape by every major generator. A variant centers one enormous line icon above the text.

Reliability: moderate. Status: active.

Numbered section markers (01 / 02 / 03)

Two-digit display numbers as section labels, or a horizontal step row with circled numbers. Editorial scaffolding one tier deeper than the eyebrow chip.

False positives: genuinely sequential processes.

Reliability: moderate. Status: active.

Repeated section kickers

The same tiny uppercase label repeated above every section heading ("FEATURES", "PRICING", "FAQ"). The model templates one section and stamps it.

Reliability: moderate. Status: active.

The cookie-cutter page order

The full skeleton in fixed order: hero, logo wall, three-card features, testimonial carousel, stats row, pricing table, FAQ, footer with a repeated call to action. One 2026 analysis found about half of indie SaaS pricing pages following an identical nine-element template.

False positives: this is also the standard human landing-page playbook. Zero deviation plus other fingerprints is the signal.

Reliability: weak alone. Status: active.

Animated stat counters and implausible numbers

A hor

[truncated for AI cost control]