Changes made during a visual interruption go unnoticed
Aliases: change detection failure · flicker paradigm
What it is
Something on screen changes substantially, but if the change coincides with a visual interruption — a blink, a saccade, a blank frame, an occluding object, a page transition — people routinely fail to see it. This is change blindness.
Take "substantially" literally. In lab studies the altered element can be the main subject of a scene, a large region of colour, an entire building; observers stare for tens of seconds, cycling back and forth, and still cannot find it.
This is not the same as not looking. Change blindness happens while someone is looking directly at the screen and actively attending to it. That is what makes it counterintuitive, and what makes it dangerous in interfaces.
Why it happens
The visual system does not store a complete previous frame. It feels as though the whole scene is held in mind, but what actually persists is a handful of attended objects plus a coarse scene gist.
So why are changes usually obvious? Not because we compare against memory, but because the change itself produces a motion transient — a local, abrupt shift in pixels that pulls attention to its location automatically, without conscious effort. That is why changes normally feel conspicuous.
An interruption works by drowning out that transient:
- a blink, saccade or blank frame makes the entire display change at once, so the local signal no longer stands out;
- an occluder simply hides the transient where it cannot be seen.
With the motion signal gone, only memory comparison remains — and memory holds just the few attended objects. If the change is not on that list, it is not found.
This also explains why changes at the centre of interest are detected far more readily: those objects were already on the attended list.
Studying it
Three standard paradigms:
- Flicker paradigm — a blank frame is inserted between two alternating images; the measure is how long, or how many cycles, it takes to spot the change. Most common, since it needs no eye tracker.
- Saccade-contingent change — the display is altered during an eye movement, which requires real-time gaze triggering.
- Occlusion-contingent change — a moving object briefly covers the changing region; higher ecological validity.
Typical independent variables: type and duration of the interruption, whether the changed object belongs to the scene's centre of interest, magnitude of the change, and whether a cue is given in advance. Typical dependent variables: detection rate, time or number of cycles to detection, and time for the first fixation to land on the changed region (eye tracker required).
In interface research these paradigms serve two main purposes: assessing whether state changes and notifications are actually seen, and evaluating motion design by contrasting a transitioned version against an abrupt one.
One methodological caution: interruptions in the lab are imposed artificially, whereas interruptions in real interfaces arise naturally from scrolling, navigation and overlays. Ecological validity does not transfer cleanly, and recent work tends to use interruption conditions closer to actual use.
Where it stops holding
- The interruption does not need to be large. One scroll, one overlay, one navigation is enough. No dramatic blank screen is required, which is precisely why this occurs so often in real products.
- It does not apply when attention is already there. If the user is looking straight at the object, the change is caught immediately. The same interface therefore behaves differently across users and moments, so a single test session cannot cover it.
- Individual differences and domain expertise modulate the effect but do not eliminate it. Experienced operators miss changes too.
- This describes failing to see, not seeing without understanding. The latter is a separate problem with separate remedies.
- Lab detection rates do not transfer to interfaces. In these paradigms participants know a change exists and are actively hunting for it; real users do not know what to look for, so field detection is lower still.
Applying it
- Give changes a transition. Move things through the full animation path so the displacement itself generates a motion signal, instead of jumping from A to B.
- When a transition is impossible — a full reload, a filter switch, a background data push — mark the changed region briefly so the difference is pointed out.
- Updating a list while the user is scrolling is the classic failure case: scrolling is itself the interruption. Either defer insertion until scrolling stops, or mark the inserted rows.
- When several things change at once, do not mark them all. Marking everything is the same as marking nothing; mark only what the user's current task depends on.
- How to check: screen-record the flow and step through it frame by frame, looking for jumps that carry no cue at all. Every such jump is a candidate defect.