The AI Slop Test

How the test works

The AI Slop Test checks whether a screen, a page or an AI feature looks designed or generated. It has five rows. Each row has a few yes or no questions.

The rule: if you answer yes to any question, the work fails. Fix it, then run the row again.

Run it at two moments. Once before design critique, so the review spends its time on real decisions. Once more before release, by someone who did not build the work.

Row 1: visual

• Does the page rely on a default gradient, most often purple or blue on white, that nobody chose on purpose?
• Does it use glassy blobs, generic 3D shapes or stock "AI brain" art?
• Are the cards all the same size in a grid, whatever their content?
• Could a stranger tell this screen came from a different session than the screen next to it?

Craft alternative: a deliberate palette written down as hex codes, and real artefacts laid out on a grid that follows the content. Write the design rules before the model guesses shows how.

Row 2: content

• Does the page show placeholder data that looks real, such as invented names, balances or chart values?
• Does it state a metric with no source you could link to?
• Does it quote praise from someone who is not named, or who did not agree to be quoted?

Craft alternative: real data, or sample data labelled as sample. Every number carries a source and every quote carries a name the person agreed to.

Row 3: AI UX

Skip this row if the work has no AI feature, and record it as not tested.

• Does the feature show only the happy path?
• Is any of these states missing: empty, working, unsure, wrong, corrected, handed off to a person?
• Can the user act on an AI answer without a way to check it, edit it or undo it?
• Can an agent take an irreversible action, such as sending, paying or deleting, without an approval step?

Craft alternative: every state designed, with a human in the loop for anything that cannot be undone. Show confidence where the user makes a decision. Design the states the AI demo skips walks through each one.

Row 4: copy

• Does the copy lean on stock hype words such as "unlock" or "seamless", or open with a line about the modern world moving fast?
• Does the headline fail to name who it is for or what they get?
• Do sentences chain together with dashes, or set up a straw claim only to knock it down in the next clause?
• Does any claim lack a verb and an object you could check?

Craft alternative: a concrete claim, with a verb and an object, that a reader could test.

Row 5: process

• Did the work ship straight from the generator without a real user seeing it?
• Did it skip a critique pass by someone other than the maker?

Craft alternative: at least one real-user check and one critique pass before release.

Worked example

We ran the test on the old Produlogi homepage before our redesign. It failed rows 1, 2 and 4.

• Visual: an animated gradient background from a third-party script, and four stock-style portfolio images with alt text such as "AI product interface".
• Content: a "15+" products figure and "full client satisfaction" with no named clients, and a "2-4x Conversion" chart with no source.
• AI UX: not tested. The page had no AI feature.
• Copy: an H1, "Design. Build. Ship.", that named no category or audience, and training described as HRDC-claimable without naming the registered provider.
• Process: no record either way, so we made the check a release rule for the new site.

The full write-up, with what we changed, is in the Lab Note Our own homepage failed the slop test.

Scoring sheet

Copy this into your review notes for each screen or page.

• Screen or page name:
• Reviewer, who did not make it:
• Row 1, visual: pass or fail, with notes.
• Row 2, content: pass or fail, with notes.
• Row 3, AI UX: pass, fail or not tested, with notes.
• Row 4, copy: pass or fail, with notes.
• Row 5, process: pass or fail, with notes.
• Result: ship only with zero fails.

Where it comes from

The rows come from the voice and craft rules we use on our own site and client work. The AI UX row draws on Google's People + AI Guidebook and Microsoft's Guidelines for Human-AI Interaction. The visual row reflects a pattern Anthropic describes in its post on improving frontend design: unguided models converge on the same fonts and gradients.

If a product you are building fails the test and you want help fixing it, our consulting starts with a test run. If you want to learn to design AI features that pass it, see 1-on-1 coaching.