Human Attention in an Algorithmic Environment

Attention has always been the binding constraint on information processing. What changed over the last two decades is who allocates it. A meaningful share of what a person attends to in a day is now selected by ranking systems optimising an objective — usually engagement — that is related to, but not identical with, what the person would have chosen deliberately.

The capacity constraint is real and small

Working memory is severely limited. Cowan (2001) revised the classic estimate downward to roughly four chunks under conditions preventing rehearsal. Cognitive load theory (Sweller, 1988) established that instructional and interface material competing for that capacity degrades performance on the primary task, regardless of how relevant the competing material is.

This matters commercially because interfaces routinely present decision-relevant information alongside a substantial volume of decision-irrelevant material. The irrelevant material is not free. It consumes the same limited resource the decision requires, which is why cluttered comparison pages produce worse decisions rather than merely slower ones.

Interruption is more costly than its duration

Research on workplace interruption found that interrupted tasks were often completed in comparable time but at the cost of higher reported stress and effort, with people compensating by working faster (Mark, Gudith and Klocke, 2008). The cost of an interruption is not the seconds it occupies; it is the reconstruction of task context afterwards, and that cost is paid in a currency dashboards do not measure.

Notification design sits directly on this finding. A notification that produces a two-second glance and a return to task has not cost two seconds. It has cost the glance plus the resumption, and where the interrupted task was a decision in progress, it may have cost the decision.

Engagement optimisation and goal conflict

A ranking system trained on engagement signals learns what holds attention. Whether the person endorses spending attention that way is not in the objective function, and there is no general reason to expect the two to coincide. Content that reliably captures attention — novelty, threat, social comparison, unresolved narrative — is a well-documented set of stimuli, and it is not the same set as content people report valuing afterwards.

Vosoughi, Roy and Aral (2018) found that false news on Twitter spread significantly faster and more broadly than true news, with novelty a plausible mechanism. The system was not designed to favour falsehood; it was designed to favour what people engage with, and novelty is engaging. This is the structural version of the problem: the objective is a proxy, and proxies come apart from the thing they proxy under optimisation pressure.

What attention scarcity does to judgement

Under attentional load, people rely more heavily on heuristic processing. The heuristic-systematic and elaboration likelihood traditions both predict this: when capacity or motivation is low, peripheral cues do the work that argument quality would otherwise do (Petty and Cacioppo, 1986).

The practical consequence is that attention scarcity is not neutral with respect to persuasion. It systematically advantages surface signals over substance. An environment engineered to fragment attention is, without anyone intending it, an environment where weak arguments perform relatively better.

What this means in practice

Two things follow for anyone designing a decision environment. First, treat attention as a budget you are spending on the person's behalf, and be able to say what each element buys them. Second, be sceptical of engagement metrics as proxies for value delivered — they measure attention captured, which is exactly the quantity that comes apart from usefulness under optimisation. The related discussion of dark patterns covers the deliberate end of this spectrum.

References

1. Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87–114.
2. Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science, 12(2), 257–285.
3. Mark, G., Gudith, D., & Klocke, U. (2008). The cost of interrupted work: More speed and stress. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems.
4. Vosoughi, S., Roy, D., & Aral, S. (2018). The spread of true and false news online. Science, 359(6380), 1146–1151.
5. Petty, R. E., & Cacioppo, J. T. (1986). Communication and Persuasion: Central and Peripheral Routes to Attitude Change. Springer.