What counts as AI slop
AI slop is design and copy that comes from a model’s defaults with no decision behind it. On a website it looks the same everywhere: an indigo-to-violet gradient, a pill that says “Now in beta” above a centered headline, three cards with an icon in a rounded square, and copy that could sell any product. None of these is wrong on its own. Slop is the absence of a choice, so the detector looks for many small defaults together, not for one.
How the detector works
- 1
A real browser opens the page
Headless Chrome loads it at 1440 × 900, scrolls it once, and measures the styles it actually renders: fonts, gradients, gradient text, pill shapes, frosted glass, big shadows, uppercase labels and emoji.
- 2
A vision model describes the screenshot
DeepSeek V4 Flash looks at the first three screens and writes down what is on them, as observations, not opinions.
- 3
Jev answers 35 questions at once
Jev, the System One model from TypeSafe AI, reads the description, the measured styles and the page copy, and answers one yes/no question per tell, plus one about the whole page. Each answer is a probability. The whole call takes about a second and costs a fraction of a cent.
- 4
A text model writes the verdict
DeepSeek writes the one-line verdict from the tells Jev found. The fixes are the found tells' own, so no model invents advice. Neither changes the score.
How the score is worked out
60% of the score is Jev’s answer to one question about the whole page: was it assembled from defaults, with few decisions of its own? The other 40% is how strongly the eight strongest tells fire, each weighted by how much it gives away. An indigo gradient counts more than Inter does. Every input is a probability, so the score is one too.
| Score | Verdict | Means |
|---|---|---|
| 80–100 | Pure slop | Every default, none of the decisions. |
| 60–79 | Heavily prompted | A person was here, briefly. |
| 40–59 | Template-adjacent | Some choices, lots of presets. |
| 20–39 | Mostly human | A few defaults slipped through. |
| 0–19 | Hand-made | Somebody decided things here. |
The 34 tells it checks
The design tells come from kill-ai-slop by yetone, and the copy tells from Peter Yang’s no-ai-slop. Motion tells, like a card that scales on hover, are left out, because a screenshot cannot see them.
Color
- Indigo-to-violet gradient
- Fix: Pick one accent with a reason behind it and use it flat.
- Gradient headline text
- Fix: Set headings in one solid color and build hierarchy with size and space.
- Glow-lit premium dark mode
- Fix: Hold one flat background and separate surfaces with a hairline.
- Stock warm palette
- Fix: Use a neutral base and one restrained warm accent.
- Default rainbow of tinted boxes
- Fix: Grow any extra colors out of your one palette, or drop them.
Type
- One word in serif italic
- Fix: Keep one voice per headline; emphasise with weight or a line break.
- Display serif as the 'premium' look
- Fix: Choose a face you can defend for this product, not the default of premium.
- A kicker above every heading
- Fix: Delete kickers that repeat their heading; let scale introduce sections.
- A whole sentence at display size
- Fix: Make the big line a few words and put the rest in a normal-size subline.
- The five default typefaces
- Fix: Try a few faces with your own copy and keep the one you can explain.
- The tasteful terminal
- Fix: Keep monospace for code and use terminal styling only where it serves the product.
Copy
- The AI copywriting voice
- Fix: Write one specific sentence with a real number, noun or consequence.
- Slop vocabulary
- Fix: Replace each hype verb with what the product actually does.
- A headline that fits any product
- Fix: Say who it is for and what changes for them, in their words.
- Em dash habit
- Fix: Use a period or a comma. In short copy, use no dashes.
- Emoji on headings and bullets
- Fix: Cut decorative emoji; keep one only where it carries information.
- Faux-insight setups
- Fix: Cut the setup and make the claim on its own.
- The invented stat row
- Fix: Show one number you measured and say where it came from.
- Stock calls to action
- Fix: Name the first thing the visitor gets when they click.
Components
- The announcement pill
- Fix: Show a badge only for real, current status.
- The glowing status dot
- Fix: A flat dot plus a word, and only when something is actually live.
- Rounded-square icon tiles
- Fix: An icon must carry meaning or go. A label and one sentence is enough.
- Icons in a tint of themselves
- Fix: Let the icon inherit the text color, with no container.
- Pills and frosted glass
- Fix: Hold one small radius site-wide and use solid surfaces.
- The oversized soft shadow
- Fix: Use a tight, colorless shadow, or only a hairline.
- Cards with a colored left border
- Fix: Let a list be a list; save callouts for one or two real asides.
- AI-drawn logo or mascot
- Fix: Get a real mark from a designer, or refine one until it is crisp.
Layout
- The template hero
- Fix: Start with what only your product can show above the fold.
- The three-card feature grid
- Fix: Pick the most important thing and show it fully instead of equal boxes.
- 01 / 02 / 03 section markers
- Fix: Number only what really happens in order.
- Cards inside cards
- Fix: One surface per region; group inside it with spacing.
- Bento grid of tinted tiles
- Fix: Show the product doing one real thing instead of a mosaic of teasers.
- Logo wall and testimonial carousel
- Fix: Show one real customer with a real result, or nothing.
- Builder template left as shipped
- Fix: Replace the template's structure with one decided for your product.
Questions
- What is AI slop on a website?
- AI slop is design and copy that comes straight from a model's defaults, with no decision behind it: the indigo-to-violet gradient, the pill above the headline, three feature cards with icons in rounded squares, and a headline like 'Supercharge your workflow'. Each one is fine on its own. Together they make every AI-built page look the same.
- Does a high score mean AI built the site?
- No. It means the page uses the defaults AI tools reach for. A person who copied a template can score high, and a site built with AI but full of real design decisions can score low. The detector measures missing decisions, not which tool was used.
- How accurate is the AI slop score?
- Every tell comes back as a probability from Jev, not a yes or no, and the report shows each one. Jev reads a written description of the screenshot, so it can miss what the description misses. Treat a found tell as something to check on the page, not as proof.
- What happens to the sites I check?
- The report is kept for 90 days. The screenshot is not: the vision model sees it once and it is thrown away. Anyone with the link can open the report, and recent reports are listed on this page. Reports are not indexed by search engines. Only public pages can be checked: the detector does not log in and does not follow private addresses.
- Why only 2 scans a day?
- Each scan runs a real browser and three model calls. The limit keeps the tool free for everyone. Opening a report someone already ran today is free and does not count.
- Can I show my score on my site?
- Yes. Every report has a badge you can paste into your footer or README. It links back to your report, so visitors can see how the score was worked out.
Built on Jev
This detector is one example of what Jev does well: many small yes/no judgments, each with a probability, in one fast call. See what Jev is or what other people built with it.