Modern interfaces have gotten worse. And it's being sold as progress

Bigger cards, floating panels, translucent controls, dedicated hardware buttons. Flagship interfaces keep acquiring visible changes that scatter attention, hide controls, and weaken feedback.

When a UI change makes a specific user task harder to complete, it is degradation. Calling it progress doesn't change the result.

A bigger screen doesn't give you more choice

YouTube stretched its recommendation grid across almost the entire viewport while keeping three videos per row. Thumbnails became giant posters, and the row stopped functioning as a quick overview.

Constraining the column width makes the cards smaller, fits more content into view, and makes titles and thumbnails easier to compare. The larger version spends the extra screen area enlarging the same choices.

In the Swifter study, a video player showed multiple frames around the current timeline position and let users select one directly. Participants located scenes up to 48% faster than with the comparison interfaces. The extra previews earned their space by making a specific search easier.

That is a useful test for a larger interface: what can people now find or compare more quickly?

Content density vs. visual noise

Element count explains only part of cognitive load. Two different things get bundled together:

Content density: how much useful information or action fits into a screen area.
Visual density: how many signals compete for attention—borders, colors, highlights, layers, and changes in scale.

On Best Buy's storefront, comparing RAM and prices across three MacBooks is interrupted by promotional badges, color swatches, and links to other collections. The information exists, but the screen keeps pulling attention away from the comparison.

Content densityVisual density

Large cards can feel heavier than a dense specification table. A table gives each attribute a predictable position and aligns values across products. You can scan one column instead of reconstructing the comparison from three separate compositions.

Semizer and Michel separated clutter from item count experimentally. Participants searched photographs for common objects while the researchers independently varied background clutter and the number of possible target locations. Even with the same number of locations, cluttered images took longer to search.

Removing useful information and removing visual noise are different interventions.

In macOS 27, visual clutter reaches the system chrome:

Safari: the segmented tab bar sits on a heavy grey substrate, with dark hollow slots for inactive tabs. An element visible throughout a browsing session becomes a permanent source of visual noise.
System menus: an oversized left gutter reserves room for occasional checkmarks. Labels shift right, while mixed icon and non-icon items create stepped alignment.

Instagram and Threads offer a useful counterexample. Both contain many features, but secondary tools appear progressively. Their everyday screens don't have to display a permanent inventory of everything the product can do.

Feature count and perceived density are different things.

Spatial maps vs. search lists

The old macOS Launchpad was a finite spatial map. Users arranged apps into pages and folders, then found them by location, neighbors, and color. Remembering the exact name was optional.

In macOS 26, Apple moved app browsing into Applications inside Spotlight. Search, categories, lists, and results from different sources now share one surface.

Typing a query begins with recalling something about the target. Browsing a familiar arrangement lets recognition do more of the work. A person can know “the orange icon in the bottom-left folder” without retrieving the app's name.

These approaches can coexist. The problem begins when the replacement removes the user's spatial arrangement while presenting several competing entry points. A vertical list does not preserve the map just because it contains the same apps.

Uniform icons weaken recognition

Icon standardization began in Big Sur, but the macOS 26 dark set goes further: nearly every app receives the same black background. Large color cues disappear, leaving more recognition work to the inner symbol, label, and position.

Right: a third-party reconstruction of the dark set, showing shared shapes and color areas rather than exact system rendering.

Nagy and Sanchez tested how quickly people could find a target distinguished from its neighbors only by color. When the color difference was small, search time increased as the display grew denser. With a large color difference, it stayed roughly constant.

Large color cues can carry real search value. Making every icon's background consistent removes part of that distinction.

Some symbols remain immediately recognizable. But the set has given up one useful channel of recognition to achieve a more uniform appearance.

Every boundary has a cost

A visual boundary earns its place when it explains behavior. A panel floats over scrolling content, moves independently, or appears temporarily.

When a panel stays fixed like an ordinary sidebar, wrapping it in a floating glass shell adds nested outlines, extra margins, and contrast states without adding functionality. The same content now occupies more space and asks the eye to parse more structure.

In macOS 27 Finder, the decoration also breaks spatial hierarchy:

Floating pills over dead space: navigation capsules and view pickers sit above an empty, static grey toolbar row. Controls occupy one layer while another layer underneath still consumes window height.
Plane collisions: right-side toolbar buttons spill over the preview pane. At the bottom, breadcrumb text can sit directly over file-list rows when the path bar loses its separating background.
Unstable affordances: in Settings, Back and Forward share a pill whose separator disappears on hover. The boundary between two actions changes while the user is interacting with them.

Current Finder layout in macOS 27Illustrative concept with controls anchored to the content canvas

The carousel compares the current Finder with my generated form concept. The concept explores control placement and spatial hierarchy; it is not a working interface.

The material is supposed to clarify which objects belong together and which sit above others. Here, it makes those relationships harder to read.

An interface can repeat itself without saying anything useful.

On the iOS 27 version screen, “iOS Version” appears in the navigation bar, again as the page heading, and again inside the information card. The version number adds information. The repeated headings consume attention after the context is already established.

Desktop windows waste space in a similar way:

macOS 27 Trash: an entire horizontal header row exists around a single “Empty” button.
macOS 27 Mail: three toolbar tiers precede the email. A higher-level mailbox filter appears visually nested below local message tabs, while “1 Message” and “Summarize” consume another full row.

Each block can look reasonable in isolation. Together, they reveal a screen that has accumulated components without resolving their shared hierarchy.

The email is the reason the window exists. It should not have to wait for three rows of chrome to finish introducing it.

Hardware buttons without a clear interaction model

Action Button

The standard Action Button settings offer one global assignment. There is no simple native App → Action table.

Want a flashlight globally and a shutter button inside Camera? That requires building and maintaining conditional logic in Shortcuts.

The hardware supports the idea. The settings make the user implement the behavior model.

Camera Control

Camera Control opens Camera and takes a photo with a click. Adjusting a parameter introduces a more demanding sequence: a light double-press to open the menu, a swipe to choose a parameter, a light press to confirm, and another swipe to adjust.

All of this happens while holding the phone, regulating finger pressure, and moving along a narrow edge.

Sensitivity settings can make individual gestures easier. They leave the sequence intact. A dedicated hardware control should make the common adjustment more direct; this one puts a gesture menu on the side of the phone.

Broken hierarchy and feedback loops

Notification Center

The section heading, “While in Do Not Disturb” label, notification card, and two similar × buttons carry nearly the same visual weight.

The user has to work out which button clears a group and which dismisses the larger layer. The hierarchy exists in the software, but the visual treatment barely communicates it.

On light wallpaper, transparent cards introduce another problem: text loses separation from the background. The notification is present and still difficult to read.

macOS Volume HUD

The volume indicator moved from a large lower-center overlay to a small translucent pill beneath the notch.

The old HUD was crude but conspicuous. When the user's gaze remained near the keyboard, it appeared close enough to register. The new indicator is smaller and farther away, asking for a more deliberate glance.

Tidiness has taken priority over noticeable feedback.

iOS Brightness slider

Control Center covers the photo, text, or app whose brightness the user is trying to judge.

The slider provides clear feedback about finger position while hiding the result that matters. This creates a loop:

Open → adjust → close → check → repeat.

A compact control or preserved preview area would let the user judge the change while making it.

Transparency makes contrast unstable

Glass backgrounds change with wallpaper, video, photos, and scrolling. Text that reads clearly in one frame can lose contrast in the next.

A successful screenshot proves very little about the material's behavior across those states. Contrast needs to survive the actual backgrounds passing beneath the controls.

WCAG contrast checks remain a baseline, but checking one static color pair cannot establish readability throughout a moving translucent interface. Additional perceptual diagnostics, including APCA, can help examine the result; the rendered states still need inspection.

For #317CFF, WCAG 2 favors black text while APCA favors white. WebKit's analysis explains the difference.

Accessibility settings also need to be judged by their outcome. Reduce Transparency, Reduce Motion, and related controls address different parts of the experience. They do not provide one switch that restores a simpler geometry across every glass surface.

Performance is a separate question. Frame time, dropped frames, GPU use, and battery consumption need measurement. A polished still image cannot tell you what the device pays to produce it.

Text input changes scale and can trap the keyboard

In Mobile Safari, focusing an input with small text can trigger automatic page zoom. Fields below roughly 16px commonly cause this behavior, and the page does not always return to its previous scale afterward.

A local readability adjustment changes the geometry of the entire page.

Keyboard dismissal has another gap. UIKit and SwiftUI leave dismissal behavior to the app and its containers. If the developer provides no swipe, outside tap, or Done action, the user may have no obvious way to get the keyboard out of the way.

The field works. Typing works. Finishing the interaction remains unresolved.

How to measure UI degradation

Static screenshots show composition. They don't show whether a person can complete the task efficiently.

Check	What to measure
Finding and comparing	Time to the first correct action; errors and wrong turns; alternatives visible together; competing anchors and repeated labels
Wasted space and extra actions	Share of the window occupied by panels and empty margins; how often settings must be closed to inspect the result
Legibility across states	Worst-case rendered contrast; failing backgrounds; readability during motion
Control and system cost	Configuration and reset steps; frame time; dropped frames; GPU use and energy consumption

Test the actual job: compare laptops, find a familiar app, choose a video, adjust a photo's brightness. Ask for preference after the task. Before interaction, people are judging the picture. Afterward, they have something else to judge.

Production has accelerated. Quality control hasn't kept up

Generative tools can increase output substantially. Across three randomized field experiments involving 4,867 developers, access to an AI coding assistant increased completed tasks by about 26%. Cui et al.

Review has its own capacity limit. A 2026 preprint tracking 802 developers and 196,212 pull requests at one company found that merged PRs per engineer eventually doubled, reviewer load roughly doubled, and automated review overtook human review. Merge and revert rates remained steady. Output and review changed substantially without a corresponding increase in those failure measures.

The gains also depend on the work. In METR's early-2025 randomized study, 16 experienced developers completed 246 tasks in repositories they knew well. With AI available, they took 19% longer—even though afterward they believed it had made them faster.

Generative tools and component libraries let teams assemble screens quickly from ready-made blocks. But checking each block separately does not tell you whether the assembled interface is usable.

Headings repeat, containers accumulate, and controls obscure the content they change. Each block can look reasonable in isolation while the complete interface makes the user work harder.

Useful density means delivering more actionable information with less competing noise. Sometimes that requires showing more data, sometimes leaving space, and sometimes deleting a decorative surface that does nothing.

My conclusion: interface production has accelerated faster than quality control.