AI slop vs good design

The second line adds almost no information. It restates the heading at greater length, so a reader who already understood the section reads it twice before reaching the controls.

The habit has a mechanical cause. A header followed by a paragraph is the shape of nearly every document a model has seen, so an empty slot under a heading gets filled whether or not the product has anything to put there. Humanizer records the same pattern as a drafting leftover: a heading immediately followed by a one-line paragraph restating it.

In an interface the cost is larger than in prose. The restatement sits between the heading and the first control, which is where a scanning reader expects the controls to be, so a sentence that adds nothing also delays everything.

Add supporting copy only when the section needs context:

The second version provides a constraint the controls alone may not communicate. One test settles most cases: would the sentence survive being read aloud to someone who can already see the heading? Manage your notifications would not. Critical account alerts are always sent by email would, because it names a behavior the toggles below never expose.

A button label is already a sentence. Invite member names the action, the object, and the outcome in two words, and the reader has read it before reaching any paragraph about it. Explanatory copy attached to an obvious control usually exists because the section felt bare, not because anyone was confused.

The stronger version replaces explanation with information the reader cannot get elsewhere. The member count does work no description can: it tells a reader whether they need the button at all.

The paragraph sounds complete while saying very little. Every noun in it is a category rather than a thing: insights, data, decisions, potential. A reader finishes the sentence without learning what the product measures, what it can compare, or what they would do with the result.

Nielsen Norman Group's study called this register marketese, and stripping it alone measured 27% better than the promotional control. Filler carries an extra cost on product pages, because a visitor evaluating a tool is looking for a reason to disqualify it quickly. A paragraph with no checkable nouns gives them nothing to test, so they leave without a reason to stay.

Specific nouns and actions make the product legible. A dependable rewriting move is to name the smallest real thing the feature does and let the reader infer the size of the rest. Export the current view as CSV implies a product with views worth exporting. Unlock your full potential implies nothing anyone can verify.

Start with the information.

Staging comes from drafting. A writer clears their throat while looking for the sentence, and the throat-clearing survives into the finished text. Humanizer treats it as one of the five patterns strong enough to act on after a single sighting, because it has no defensible use: announcing that an explanation is coming always takes longer than the explanation's first sentence.

Interfaces speak their own dialect of the habit. In this section you can, Below you will find, and Use this page to all describe the page instead of helping the reader use it. Delete the run-up and check whether the remaining sentence still stands. It almost always does, and it is now the first thing the reader sees.

When a real contrast matters, keep it concrete:

Both halves now carry product information.

Formulaic contrast works by inventing a weaker version of the product and then defeating it. Nothing in the source claimed the feature was simple search, so the sentence answers a position no one held. Humanizer names both halves of the habit: the not-X-but-Y construction, and arguing with no one, which is the same move spread across two sentences.

Keep a contrast when readers genuinely arrive holding the wrong belief and both sides carry something they can act on. Archived projects remain searchable. Deleted projects are removed after 30 days survives because each half states a fact with a different consequence. The test is whether the negative half would be worth saying on its own.

A short final sentence is fine when it adds a fact. Repetition used only for emphasis should be removed.

A closer asks for a pause. A reader can grant one in an article. In an interface the pause lands directly above a control, and the reader rereads the paragraph to check whether the short line changed anything. It never has.

The same habit produces emphatic fragments: No setup. No configuration. No limits. Three fragments look decisive and carry less information than one sentence naming what the product does not require. Merge them and the claim becomes checkable.

The phrase could describe hundreds of products.

Three items can remain when the product genuinely has three distinct parts. The issue is rhetorical symmetry used as a substitute for substance.

Count what the third item adds. Design faster, collaborate smarter, and ship better repeats one claim at three points of a workflow, so the length came from rhythm rather than from the product. The signal strengthens when neighboring lists also run to three, especially where the underlying data has four items or two.

Interfaces institutionalize the problem through layout. A three-column grid asks for three features, and the third gets invented to fill the column. Let the content set the count, then choose a layout that holds it. Two strong features in two columns read as confidence; three with one filler read as a template.

Let the reader decide whether the capability is important.

Inflation is detectable without judging the claim. The sentence describes the product's position in the world rather than its behavior: it represents a shift, marks a moment, or belongs to a larger movement. Strip those phrases and read what remains. If nothing remains, the sentence carried no product information at all.

Interfaces inflate a second way, by promoting ordinary events into announcements. A banner reading A new era for your workspace above a settings change spends the reader's attention on framing instead of on what changed and what they should do about it.

Concrete behavior is more credible than abstract benefit language.

The words are symptoms rather than the defect. Each names a benefit while skipping the mechanism, so a sentence can hold four of them and still leave a reader unable to describe what the software does. Vocabulary lists also expire, since model generations favour different adjectives and a ban list assembled today will miss the next batch. Count the missing nouns rather than the offending adjectives.

The reliable rewrite is to state the behavior that produces the benefit and let the benefit go unsaid. Readers infer seamless from a description of two systems exchanging data without a copy-paste step, and they trust the inference more than the adjective.

The body should advance the user's understanding.

A status message has two jobs: say what happened, and say what to do next. Restating the heading does neither, and the temptation is strongest exactly where the product knows the most. After an import, an export, a deploy, or a sync, the system is holding counts, durations, skipped rows, and warnings, and a generated message discards all of it for a sentence confirming success.

The cost goes past wasted words. A reader told only that an upload completed has no way to know that seven records need review, so a message meant to reassure hides work they still have to do.

The second version names the condition and recovery path.

Length is not the defect. The original is unreadable because it opens with intent, moves through hedging, and reaches its only fact in the final clause, by which point the reader has stopped. W3C's succinct-text pattern addresses that shape directly: put the main point at the beginning, keep one point per sentence, and break paragraphs that run past roughly 50 words. The original runs to 48 words and carries one fact.

Legal and policy text attracts the pattern most, because caution accumulates in front of the fact. The fix is order rather than deletion. State the condition, then the consequence, then the recovery, and keep any qualification a reader needs in order to act correctly.

The descriptor beside the wordmark and the label above the title repeat identity already visible in the same view.

Another form is a URL slug restyled as metadata immediately above its own title: AESTHETIC-USABILITY-EFFECT above Aesthetic-Usability Effect. The transformed spelling adds no new identifier. A distinct, stable rule ID such as LAYR-001 can still help citation and cross-reference; that does not mean it belongs above every page headline.

Remove an identity or page-type label when its removal changes neither orientation nor meaning. Keep a label when it distinguishes a source, status, mode, or section that the surrounding interface does not make clear. Review it in context rather than banning labels or repeated brand names everywhere.

This pattern survives review because every element is defensible alone. A logo belongs in a header. A wordmark belongs beside a logo. A descriptor explains the wordmark. A page title names the page. Inspected one at a time nothing is wrong, and the reader still meets the same word four times before reaching content. That is why the check has to be positional: count what a reader can see at once, not what each element does in isolation.

Every line here is marked twice. The rocket says the same thing as "Getting started", the checkmarks repeat a list that a list already signals, and the padlock restates the word "security" beside it. The glyphs carry no information the sentence does not, and each one is announced aloud: a screen reader reads "rocket Getting started", so the ornament becomes a word in the reading order.

The sparkle is the clearest case. It was adopted as the visual shorthand for AI in the mid-2020s, so a product that opens a section with ✨ is now labelling its own copy as machine-written — usually the opposite of what the page is trying to say.

Remove a glyph when the sentence beside it already carries the meaning. Where a feature list said ✅ three times, the rewrite states what each item actually does, which is the information the checkmark was standing in for.

Emoji keep their place where they are content rather than decoration: a user's own message, a reaction, a name someone chose, a legend where the glyph is the value being explained. Status is the one structural use that can survive, and only when the emoji is not the sole carrier — a ✅ beside the word "Passed" is redundant, a ✅ instead of "Passed" fails for anyone whose reader announces it as "check mark button", and fails again on a platform whose font draws it differently.

Treat a single emoji as a weak cue. Wikipedia's own guidance on the subject makes the same reservation about every pattern it lists: they are "only potential signs of a problem, not the problem itself." What raises confidence is repetition and position — one emoji per heading across a page, or one opening every bullet, is a formatting rule rather than a choice, and a formatting rule nobody set is the definition this law works from.

The correction uses concrete behavior. The visual treatment can then follow Layr's actual design system.

The blue stripe does not represent category or state. The description adds little information.

Or keep the stripe if it has a defined meaning:

Once color carries semantics, the accent earns its place.

If 0.1 is a real version, keep it and make it useful:

The options explain the choice.

The line sounds like a conclusion, but it does not tell the reader whether this page helps prevent, find, or remove AI slop. The ID and canonical law name add a second title before the point.

The headline names the outcome. The following sentence gives the reader a criterion for reviewing the interface. LAYR-001 and The Intentionality Law remain in the citation and source representation, where they help identify the rule.

.category-row > a:hover {
  padding-left: var(--space-2);
  color: var(--color-accent-text);
}

.category-row > a:hover { color: var(--color-accent-text); }
.category-row > a:focus-visible { outline: 2px solid var(--color-focus); }

The row still responds to pointer and keyboard interaction. Its content stays in place because shifting it communicates no additional state. A motion cue with a real purpose should be judged in its own context.