Anti-AI-Slop Audit

When to use this

• You are reviewing a site or component and suspect it was AI-generated and needs to look like it was not.
• You just generated a UI with an LLM and want to check it against the tell list before shipping.
• You need a structured rubric to score a frontend for "does this look like a template" and specific fixes.
• You are building a design review checklist for a team that uses AI code generation.
• Do NOT use this when you need a general interface critique framework; use design-critique-rubric instead.

Mental model

AI-generated frontends have a convergent aesthetic. This is not because LLMs lack taste; it is because they reproduce the statistical center of their training data. The statistical center of web UI is: centered layouts, purple-blue gradients, uniform border-radius, fade-up-on-scroll for every element, three-card grids, emoji icons, and shadows on everything.

The tell is not any single one of these. It is the combination. A site with rounded corners is fine. A site with rounded corners AND a purple-blue gradient AND a three-card grid AND a fade-up AND default shadows AND centered everything is recognizably AI-generated to anyone who has seen more than five of them.

The fix is not replacing each cliche with a different cliche. It is making design decisions. A decision is: "this element has sharp corners because the brand is industrial" or "this card grid has 4 columns because the content has 4 items, not 3." The absence of decisions is the tell. Every element looking the same means no element was actually designed.

The audit works by counting tells. Each tell is 1 point. A score of 0-2 is fine (coincidence). 3-5 is suspicious. 6+ is obviously AI-generated. The fixes target root causes: if 6 tells fire, fixing all 6 cosmetically just produces a different-flavored template. Fix the 2-3 root causes (usually: "everything has the same treatment" and "no hierarchy") and the surface tells resolve themselves.