How to Tell If a Website Is AI-Generated: The Tell You Can't See

The tells everyone already lists

The community canon of vibe-coded tells is well established by now, and each one exists for a reason — they are the defaults of a model averaging the web:

• The font tell — one of a handful of trendy geometric sans-serifs, everywhere, including places a designer would never put it.
• The gradient tell — a purple-to-blue hero gradient, the statistically dominant 'looks modern' choice.
• The layout tell — three equal cards inside rounded containers inside rounded sections, padding that never varies.
• The copy tell — confident, generic section headlines that could belong to any product in any industry.

The tell a screenshot can't show: contrast

These tells work because they are frequent. But they are also fixable in an afternoon, which is exactly why the next generation of AI-built sites will pass all of them. Judgment-based tells decay; measured tells do not.

Contrast fails in the least photogenic way possible: the muted gray text on a white card, the half-transparent label on a dark hero, the placeholder that looks intentional. At thumbnail size — which is how everyone reviews screenshots — a 3.9:1 pair and a 7:1 pair are indistinguishable. At measurement they are the difference between passing WCAG 2.1 AA (4.5:1 for normal text) and failing it.

The base rate makes this the highest-probability tell of all. The WebAIM Million has found low-contrast text on roughly four in five of the top million home pages every year — and that is before adding the specific failure modes of generated UI: default muted grays, opacity-based 'subtle' text, and dark mode pairs that invert a light theme's contrast margins. An AI-built site that nails every visual tell can still ship a settings page where the secondary text measures 3.6:1.

A 60-second contrast audit of any site

You do not need a full audit tool to test the invisible tell. Four steps:

• Pick the three text styles every site has: body text, secondary/muted text, and a button or link label.
• Get the exact colors — use the browser DevTools color picker, or an eyedropper on screenshots for pairs you cannot select.
• Measure each text color against the background it actually sits on (the card surface, not the page background — this is where most audits go wrong).
• Read the numbers: under 4.5:1 fails WCAG AA for normal text; under 3:1 fails even large text. APCA adds a perceptual check for dark mode and thin fonts.

From tell to fix: keep the design, move the lightness

Three measurements in and you have a verdict no screenshot can fake. The free Color Contrast Checker does the measuring and, when a pair fails, suggests the closest fix that keeps the original hue — so you can name the tell and fix it in the same minute.

Failing contrast rarely requires repainting a brand. Because WCAG contrast responds to luminance rather than hue, almost every fix is a lightness move: darken the text, lighten the card, or shift the muted tone along its own ramp until the pair clears the line. A pair like #D97706 on #FFF7ED measures about 3.0:1; the same hue family nudged to #9A5405 measures about 5.4:1. The site looks the same to the eye and passes to the auditor.

That is also the difference between noticing a vibe-coded site and shipping one nobody can call out: the visible tells are style choices, and style is allowed. Contrast is a measurement, and 4.5:1 is the line. If you are building with an agent, put the measured pairs in the project — a DESIGN.md with per-pair results means the invisible tell never makes it to production in the first place.

References

• WebAIM Million — annual accessibility analysis of the top 1,000,000 home pages
• WCAG 2.1 — 1.4.3 Contrast (Minimum)
• APCA — Accessible Perceptual Contrast Algorithm

FAQ