Stop Using Icons in Data Tables

We have been sold a lie that icons are the universal language of the modern interface.

In our quest for “clean” and “minimalist” UI, we have stripped away the clarity of text and replaced it with a field of cryptic glyphs. Designers often treat icons as a panacea for cognitive load. But in the context of complex, data-heavy products, the opposite is true. Icons don’t simplify; they increase cognitive load.

When we pack high-density information into a data table or a complex dashboard we are increasing the visual entropy of the entire system. Forcing the brain to decode intricate, non-universal shapes in a tiny 16-pixel footprint, creates a “cognitive tax” that users pay en masse every time they scan the table.

To understand why your data table feels “noisy” despite its clean aesthetic, we have to look past the pixels and into the psychophysics of how humans actually see and process information.

Why your table icons suck

Theres an implicit cognitive translation that must take place with any non-universal icon pairing. A star icon could mean “AI” or “Recommendation” or “Rating” or any number of concepts.

This type of non-universality is useful when drawing attention but within an already crowded interactive system you end up with more cognitive load not less.

This is due, in part, to the inherent visual intricacy of an icon. This is known as “high spatial frequency”.

Icon spatial frequency

• Icons are high spacial frequency: within a tiny 16px field you have lines, curves, junctions, that don’t follow a universal rhythm. This makes us perceive them as incredibly high value for foveal vision.
• Icons are objects, its a self-contained concept. In gestalt terms icons have a stronger figure-ground relationship because the brain sees them as a discrete object

Word spatial Frequency

• Text is lower spatial frequency: predictable and rhythmic, the brain recognizes the gestalt (the Bouma shape)
• Text is a texture, they’re a recognizable pattern with relative low visual cost despite each character having relatively high spacial frequency to the word lock up.

The Artsy Stuff

I also should mention visual entropy, the degree of randomness or disorder in a visual plane.

An icon has minute details (a tiny pencil, a curved arrow). When you shrink those details down you increase the visual energy.

This concept pairs with Anne Treisman’s “Feature Integration Theory” basically the text “binds” almost instantly but icons require milliseconds of “binding” (a user can’t avoid this pull to bind and parse meaning).

So within a data table you end up with an incredibly high level of cumulative cognitive load.

The “Objectness” of an icon is its greatest strength in a vacuum, but its greatest weakness in a system.

Conclusion: Designing for the Fovea, Not the Ego

When we choose an icon over text, we are making a trade: we trade the predictable, rhythmic “texture” of typography for a high-energy, high-spatial-frequency object that demands foveal attention.

In a data table, this trade is a losing one. We are essentially littering our users’ path with tiny visual speed bumps, forcing a cycle of “Feature Integration” and “Semantic Translation” that exhausts the brain long before the data is actually processed.

As designers, we must resist the urge to “illustrate” our way out of complexity. We need to respect the Bouma shape of a word and the efficiency of a low-entropy interface. If we want to build truly efficient tools, we must stop designing for how we want the interface to look and start designing for how the human brain is wired to see. Sometimes, the most sophisticated visual solution isn’t an icon at all, it’s just a word.

Sources & Links

There isn’t a ton of graphically oriented theory writing within Cognitive Translation & Interpretation Sciences but this analysis on medical imagery is a good example

For spatial frequency selectivity I recommend this article “Bridging Psychophysics And Interface Design”. Its damn dense but only 6 pages. Upload it to gemini and talk about it with a robot

For a visual interface evaluation using visual entropy i’ll point to this assessment of air traffic controllers attempting to parse data via a continuous animation vs a stepped animation timeline.