Cognitive load

The memory limit is real and the famous number attached to it is wrong. Miller (1956) measured a mean channel capacity of about 2.6 bits — roughly 6.5 categories — for the span of absolute judgment, kept that separate from the chunk limit on immediate memory, and wrote of the recurring seven: “But I suspect that it is only a pernicious, Pythagorean coincidence.” Cowan's reconsideration puts the real figure at about four chunks, “a central memory store limited to 3 to 5 meaningful items in young adults,” and only under conditions that stop people helping themselves: rehearsal blocked, grouping prevented, sensory memory wiped by a following jumble. None of that is a budget for how many things may appear on a screen; a visible list is not held in memory at all.

The menu-count rule derived from Miller has been tested and it does not hold. Larson and Czerwinski (1998) gave 19 experienced web users three hierarchies over the same 512 items: the deepest 8×8×8 tree was slowest at 58 seconds, 32×16 took 46, and the broadest 16×32 was fastest at 36. In the older menu work they review, memory span was not predictive of performance in any condition, while visual scanning was, especially in the deepest hierarchies. Breadth is paid for by looking, not by remembering, which is why cutting a list to seven items buys you a deeper tree and a slower screen.

Cognitive load theory is the instructional literature, and it is strongest where it is most concrete. Sweller, van Merriënboer and Paas (2019) report that nearly all the effects established by 1998 were reductions in extraneous load: worked examples, their strongest claim; split attention, where a diagram and the statements that explain it have to be mentally integrated unless you physically integrate them; and redundancy, where a diagram plus a full prose restatement of the same content was worse than the diagram alone. Pooling 29 systematic reviews covering 1,189 studies and 78,177 participants, Noetel and colleagues (2022) found the integration results survive best: contiguity at g = 0.74, spatial contiguity on its own at g = 0.63, and signalling — arrows, highlights, cues telling the eye where to go — at g = 0.43. Putting the label on the thing it labels, and pointing at what matters, are the best-evidenced moves in this literature. They are measured as learning outcomes, not as completion or revenue.

The prescription everyone repeats — remove unnecessary elements — measures weakest in exactly the case an app is. In the same overview, removing seductive and irrelevant detail came to g = 0.33, and it splits sharply on whether the clutter persists: g = 0.43 when the distracting material stayed on screen against g = 0.12 when it was transient. Two moderators matter more than the principle itself. Complexity: g = 0.70 for high element-interactivity material against g = 0.20 for simple material. Pacing: g = 0.41 when the material was system-paced like a lecture, against g = 0.27 when learner-paced “like on a website,” because a self-paced user manages their own load. The hypothesised expertise reversal did not survive meta-analytically — prior knowledge did not moderate the effects — and verbal redundancy came out asymmetric rather than harmful: adding text to audio helped at g = 0.29, while adding audio to text did nothing, g = −0.04.

Offloading memory onto the interface has a different problem, which is that users will do it whether or not it helps them. In three experiments by Gilbert and colleagues (2020), people could hold intentions in their own memory for full marks or set external reminders for a smaller payout per item, and they “were significantly biased towards using external reminders, even when they had a financial incentive to choose optimally.” The bias was stable when the same participants returned weeks later and was predicted by erroneous metacognitive underconfidence in their own memory; advice about which strategy would actually score better removed it, so it is not a general aversion to effort. Autofill, saved cards and recently-used lists will be taken up readily, and that is the argument for offering them. It is also the warning: once a value lives only in your interface, whether the user can still see and change it before committing is the whole of the design.

The theory's middle construct has been given up, and its instruments are unsettled. De Jong's (2010) review states that “cognitive load theory is constructed in such a way that it is hard or even impossible to falsify,” because with one kind of load that is good, one that is bad and one that is simply there, “every outcome fits within the theory post-hoc.” Sweller and colleagues then narrowed it themselves in 2019, describing germane load as redistributing working memory resources rather than adding to the total and concluding that “only intrinsic and extraneous cognitive load are distinguished as basic categories of cognitive load.” The field's workhorse measure is a single nine-point rating of perceived mental effort (Paas 1992), collected after the task: it is not concurrent, it does not measure overload, and by Sweller's own account it “cannot easily be used to differentiate between the different types.” The NASA-TLX (Hart 1986), the instrument most HCI teams reach for, is a weighted average of six subscales, two of which are not demand at all — Performance asks how successful you think you were, Frustration how irritated and stressed you felt. A TLX score that drops may mean less was asked of the user, or only that they felt better about how it went.

And the obvious way to reduce load — strip the interface back — has been measured going the wrong way. Burmistrov and colleagues (2015) ran 20 experienced users through matched flat and traditional versions of three tasks with eye tracking. Finding a target icon took 15.4 seconds in the flat set against 8.0 in the realistic one, and on screenshots of real sites both error types rose: misses from 26.0% to 35.8%, false alarms from 16.6% to 28.0%. Light and ultralight type cost no time but produced longer fixations and shorter, slower saccades, the signature of harder processing. This is one small single-lab study, n = 20, on 2014 stimuli, and not a verdict on flat design in general. It does establish the failure mode: removing the cues that mark a thing as pressable does not remove work, it moves the work from reading to guessing, and guessing shows up as errors rather than as time.