What Is Missing? Change Detection and the Blindness Behind It

A change detection task shows you a display, interrupts it, shows it again with something altered, and asks whether anything changed. The interruption is the entire design. Remove it and the change is trivially obvious, because movement and sudden onsets are exactly what the visual system is built to flag. Insert a blank frame of a fraction of a second and the same change can become almost impossible to find — which tells you that the question is not "can you see it" but "what did you keep".

The task, and why the gap matters

The laboratory version is precise enough to describe exactly. Lin and Luck set out the canonical procedure: a sample array of simple items — coloured squares, oriented bars — appears for around 100 milliseconds; a retention interval of roughly 900 milliseconds follows with nothing on screen; then a test array appears and the observer reports whether it matches the sample or contains one changed item.

Those numbers matter. The sample is too brief to scan item by item, and the gap is long enough that the display is gone in every sense — you are comparing the test array against a memory, not against an afterimage. Accuracy across different set sizes can then be converted into a capacity estimate, conventionally written K: roughly, the number of items held well enough to support the comparison. For simple features the estimates come out small. A display can contain a dozen squares and still yield a K of a few.

The "what is missing" format is the same logic with a friendlier face. Instead of asking whether anything changed, it removes one item and asks which one. That version is harder in one respect — you have to produce an answer rather than make a same/different judgment — and easier in another, because a removal leaves a visible hole.

Change blindness, and how much it actually shows

Outside the laboratory the same phenomenon has a better-known name. Daniel Simons and Ronald Rensink define change blindness as the striking failure to see large changes that normally would be noticed easily, and trace the modern surge of interest to the 1991 discovery that observers often failed to notice large changes to photographs made during an eye movement — including, in one demonstration, two cowboys on a bench exchanging heads.

The flicker paradigm made the effect easy to reproduce: an original and a modified scene alternate repeatedly, separated by a brief blank display, until the observer finds the change. People do eventually find most changes, but Simons and Rensink note that it can take an astonishingly long time, even for large ones.

Here is where they insert a warning that is routinely ignored in popular accounts. The counterintuitive nature of change blindness, they write, has been a source of considerable confusion about the inferences that can legitimately be drawn from it, and results have sometimes been mischaracterised, leading to extreme and relatively unsupported claims. The most common overreach is the leap from "you failed to detect this change" to "you represent almost nothing of what you see". Failing a comparison shows that what was stored did not support that particular comparison. It does not show that nothing was stored.

That distinction is worth carrying into any app or article that uses a missing-item game to make a point about your attention.

What the task is actually sensitive to

Two findings sharpen what a change-detection score reflects.

First, the task is reasonably clean as a capacity measure. Lin and Luck tested whether interference from earlier trials distorts the estimate and concluded that it does not meaningfully do so, finding that performance depended virtually exclusively on memory for the current trial with little or no impact from preceding ones.

Second, the residual effect is real but small. Joshua Hartshorne, examining the same question in PLoS One, found capacity estimates dropping by around 17% from low- to high-interference conditions in one experiment and about 12% from the first to the tenth trial in a large web-based sample, with item-specific interference decaying entirely within four or five trials. His conclusion was that interference plays a role in visual working-memory performance but does not dramatically change capacity estimates.

For a player rather than a researcher, the practical translation is: your score on the first round of a session and your score on the tenth are not perfectly comparable, but the difference is modest, and it fades quickly if the content changes. The mechanism behind that fade — and why switching content type helps — is covered in our article on interference and forgetting.

A related format asks a different question about the same display. Where a missing-item round asks what was removed, a grid and position round asks where something was, and the encoding strategies diverge accordingly.