Iconography · 6 min read
Labels Are Not Training Wheels
Teams treat text labels as help for beginners, to remove once people learn the icons. Most interface icons are symbols, and symbols keep needing words.
A product team ships a new toolbar. The first release has an icon and a word under each button. Six months later someone says the words look cluttered and “everyone knows the icons by now.” The labels come off. Nobody measures what happens next, because nobody thinks anything was lost. This is a common habit: treating text labels like training wheels, a support for beginners that you take off once people have learned. Most of the time that is a mistake, and the reason sits in how signs work.
Most interface icons are symbols
The philosopher Charles Sanders Peirce sorted signs by how they connect to what they mean. An icon means by resemblance: a printer looks like a printer. An index means by a real link to its object: a progress bar moves because work is happening. A symbol means by convention: nothing about a gear says “settings.” People learn that link. They don’t see it.
Most of what we call icons on a screen are symbols in Peirce’s sense. Gear for settings, three dots for “more,” a floppy disk for save, a funnel for filter. Each one works only for people who were taught the convention, in this product or in another one.
The research has said this for a long time. Aurora Harley’s Nielsen Norman Group (NN/g) article “Icon Usability” (2014) names only a few icons with “mostly universal recognition”: home, print and the magnifying glass for search. Its rule is blunt: “a text label must be present alongside an icon to clarify its meaning in that particular context,” and “Icon labels should be visible at all times, without any interaction from the user.” Hover does not count. The article says hiding labels until hover “fails to translate well on touch devices.”
New conventions start as guesses
The clearest recent case is the sparkle icon used for AI features. In Kate Kaplan’s 2024 NN/g article on the sparkles icon, participants agreed it meant “something a little special,” then disagreed on what. In the quantitative study the top associations were “Favorite or save an item” and “Visual effects or optimization options,” at about 17% each. The article reports that “no one attributed the concept of ‘artificial intelligence’ to the icon.” Its advice: the icon “should always be accompanied by a label.”
Compare the star. The same article reports that 73% of users in NN/g’s quantitative study linked a star to saving or favoriting. That is decades of habit across many products. Even then, more than a quarter of people did not make the link. A convention can be strong and still not be universal.
So the training-wheels idea gets the timeline backwards. It assumes people learn your icons and then stop needing words. In practice the words are often how they learned the icons in the first place. Take the words away and every new user, every returning user after a long break, and every user of a new feature starts from a guess.
The practical test is not whether an icon feels familiar to the team. It is whether a new user can identify it without guessing.
The words are also commands
Labels do a second job that is easy to miss. People who control a computer by voice say what they see. The W3C’s guidance for WCAG 2.2 success criterion 2.5.3, Label in Name, puts it simply: “People who operate with voice interaction use the visible labels in their commands.” A speech-input user says “click Export” because the word Export is on the screen.
Remove the visible word and that user has to guess the hidden name behind the button. The same guidance even notes that where “a visible text label does not exist for a component, this success criterion does not apply to that component.” Taking the label off doesn’t just change the look of the screen. It takes away the word the voice user would have spoken.
WCAG does allow images as labels. The W3C’s guidance for 3.3.2, Labels or Instructions, says “Using images as labels meets the requirements of the criterion,” and adds that “care should be taken to ensure that the images are widely understood by the intended target audience.” That is a claim you can test, not an opinion you can assume. Most teams that remove labels never test it.
The case for icon-only, and where it holds
The strongest counterpoint is real. Space is limited. Harley’s article lists saving space as a genuine benefit of icons, and says icons “are fast to recognize at a glance (if well designed).” A dense editing toolbar used all day by experts can’t carry a word under forty buttons. Apple’s Human Interface Guidelines tell designers to “Include tab labels to help with navigation” and to “Use single words whenever possible,” but they don’t ban icon-only controls everywhere. Labels also cost something to translate and fit in every language.
That case holds in a narrow zone: frequent users, a stable set of controls, true icons rather than symbols, and a label one hover or one long-press away. It does not hold for primary navigation, for anything new, or for a product people use once a month. In those places the label is the cheapest piece of clarity you will ever ship. A word costs a translation. A misread symbol costs a task, a support ticket or a feature nobody finds. Harley’s article includes an example where “Not a single test participant clicked this icon,” and sums it up: “Obscure icon = wasted feature.”
My position: decide labels by sign type and by risk, not by how familiar the icons feel to the team. The team is the one group that has learned every convention in the product. They are the worst judges of what a stranger can read.
Icon lens
Roland Barthes showed that text anchors an image to one meaning. Without the word, an image floats between possible readings; with it, the reader is told which reading counts. That is the job a label does for a symbol, and it is why Appendix A.1’s working rule says symbols need a label. Digital Sanskrit takes this further. Each root carries a name as well as a form: Flow is also pravaha, Link is also bandhana. The word is part of the sign, not a caption added later. A grammar whose roots have names can teach itself to new readers. A set of drawings without names has to be memorized one glyph at a time.
Try this
**Goal:** find the symbols in one product that ship without the words they need, and decide which ones need a label first.
**Steps:**
1. Take 20 icons from one product you use or design, including its main navigation and one toolbar. 2. Mark each one I (icon: means by resemblance), X (index: tracks live state) or S (symbol: means by convention). 3. Circle every S that ships without a visible text label. Write the one or two words it needs. 4. Show five of the circled symbols, without labels, to five people who have never used the product. Ask each person to write what each one does. 5. Rank the circled symbols by how often people got them wrong and by what a wrong guess would cost.
**Check:** you have a ranked list of unlabeled symbols, and the top three either get a visible label or have test answers showing that people read them correctly without one.
**Reference:** Digital Sanskrit, Appendix A.1 (What a sign is) and Appendix A.9 (Accessibility and proof of comprehension).
Sources
- [Nielsen Norman Group, Icon Usability (Aurora Harley, July 27, 2014)](https://www.nngroup.com/articles/icon-usability/)
- [Nielsen Norman Group, The Proliferation and Problem of the Sparkles Icon (Kate Kaplan, September 20, 2024)](https://www.nngroup.com/articles/ai-sparkles-icon-problem/)
- [W3C WAI, Understanding Success Criterion 2.5.3: Label in Name](https://www.w3.org/WAI/WCAG22/Understanding/label-in-name.html)
- [W3C WAI, Understanding Success Criterion 3.3.2: Labels or Instructions](https://www.w3.org/WAI/WCAG22/Understanding/labels-or-instructions.html)
- [Apple Human Interface Guidelines, Tab bars](https://developer.apple.com/design/human-interface-guidelines/tab-bars)