NOTE-887: Risko & Gilbert — Cognitive offloading

Contribution

The review gathers work from three areas, physical normalization of rotated stimuli, prospective memory (setting reminders) and memory on the Internet (“transactive memory”), under one behaviour, and asks two questions of it: “(a) what factors influence the likelihood of individuals offloading cognition versus relying on internal processes alone, and (b) what are the cognitive consequences of this behaviour?” Its own addition is a “metacognitive framework” (Figure 3, the Key Figure) in which the decision to offload is a strategy selection driven by metacognitive evaluation, and its consequences feed back into that evaluation.

Definition. Cognitive offloading is “the use of physical action to alter the information processing requirements of a task in order to reduce cognitive demand” (abstract, text and Glossary). The authors relate it to “computational offloading [5]” (Scaife and Rogers 1996) and “epistemic actions [6]” (Kirsh and Maglio 1994, LIT-1054) with a “see also”, and do not discuss the relation further.

The two types. “Cognitive offloading can be roughly subdivided into actions that offload cognitive demands onto-the-body and into-the-world.”

• Onto the body, used “in the “here-and-now” to reduce cognitive demand”: eye fixations as spatial indexes (Ballard et al. 1997, LIT-1075), fingers and pointing to keep place, gesture, and moving the body “to simplify perceptual computations”. The worked case is external normalization: tilting the head to read rotated text instead of rotating it mentally.
• Into the world, chiefly “using it as a repository of representational information, thus eliminating the need for an internal representation”: writing down, typing, sketching. Two cases: intention offloading, “acting on the environment to create external triggers for delayed intentions” (Glossary: “Creation of a cue in the external environment to trigger a delayed intention”), and human–technology transactive memory.

Key insight

The decision to offload is driven by what people believe about the demand, not only by the demand. Spontaneous offloading rises with internal demand in every domain reviewed, but it can be pulled apart from it, it happens where it gives no benefit, and it varies with confidence independently of ability. And offloading is not free: it changes what is remembered (“a shift from remembering “what” to remembering “where.””), and it changes the beliefs that decide the next choice, so the loop can drift toward ever more external reliance.

Assumptions

• A choice between two routes. The framework “describes situations in which there are two or more ways of achieving a goal, one of which involves cognitive offloading and one of which does not.” Offloading built into the environment by design (Box 3, automation) is outside the choice.
• Choice need not be conscious. “the strategy selection phase postulated above does not necessarily imply that individuals are aware of making a choice”. Gesture is the unconscious case, choosing a GPS the deliberate one.
• Evidence is mostly from the choice/no-choice paradigm (Box 1, from Siegler and Lemaire 1997): forced-internal, forced-external and free-choice conditions crossed with a demand manipulation. The box states its limits: forced-external “cannot ensure that the information is not also stored internally”, and forcing a strategy may add the load of suppressing the preferred one.
• Effort currency left open. Box 2 sets the “cognitive impartiality” view (Soft Constraints Hypothesis: “it is not the kind of effort but rather the amount of time required that determines the preferred solution”) against the “cognitive miser” view (Minimal Memory: a bias toward minimising memory “even in the face of potentially greater perceptual-motor costs”), and does not decide between them.
• Single episodes. Every study reviewed offloads for one use within a task or a session (read this display, fulfil this intention, recall this list). Repeated use of one offloaded item appears only in the GPS example and the drift prediction.

Key results

The review reports no effect sizes, test statistics or sample sizes. Numbers below are the one percentage in the text and values read off Figure 2.

• Offloading tracks internal demand (Figure 2). External normalization: people “are more likely to spontaneously physically rotate as the display becomes more disoriented or as the number of items in the display increases”, and both manipulations raise rotation costs (Risko et al. 2014). Figure 2A runs from 1 letter to a paragraph; head rotation (in degrees per degree of stimulus rotation, near 0 to about 1.25) and restricted-motion rotation cost (ms/degree, near 1 to about 22) rise together. Intention offloading: people “are more likely to offload intentions when their memory load increases or when they encounter interruptions; both of these factors impair performance when offloading is prevented” (Gilbert 2015). In Figure 2B unaided accuracy falls from about 95% (1 intention, no interruption) to about 79% (3 intentions, interruption), and the share of intentions offloaded rises from about 33% to about 73%. Short-term memory (Risko and Dunn 2015, Figure 2C): no-choice accuracy falls from near 100% at 2 items to near 0 at 8–10, and the rate of writing items down rises from about 42% at 2 items to about 87% at 10. The pattern is also cited for co-speech and co-thought gesture.
• Offloading dissociates from demand. Words rotated inside a rotated frame and words rotated inside an upright frame “yield similar rotation costs and similar responses on a physiological measure of demand”, “Yet spontaneous rates of external normalization are much higher when both the words and frame are rotated” (Dunn and Risko 2016). People “incorrectly report that rotated word and frame displays are more time consuming and error-prone”. The review attributes this to “an erroneous metacognitive evaluation of demand”.
• Confidence predicts offloading apart from ability. “Individuals with lower confidence in their memory abilities were more likely to spontaneously set reminders, even after controlling for any influence of objective ability (which also predicted intention offloading [45]).” The relation is also domain-general: lower confidence in a staircased perceptual task, where accuracy was held constant, went with more reminders.
• Superfluous offloading. People set reminders, and rotate their heads, “in conditions where doing so led to no objective increase in accuracy”. In the short-term memory task they wrote items down “about 40% of the time when they had to remember only two items, a memory load at which performance was already at ceiling without offloading”, and “erroneously judged that offloading would improve their performance” there. Three explanations are offered: an undetected benefit, a bias against cognitive effort, or a mistaken metacognitive belief; the last is the one the authors favour.
• Benefits named. Offloaded representations “may be more durable and less prone to distortion than those stored internally”; saving a first word list improved memory for a second (Storm and Stone 2015), read as reduced proactive interference.
• Costs named.
  • Memory for offloaded content. People who typed trivia statements expecting them to be saved recalled them worse than people expecting them to be erased (Sparrow et al. 2011). Saving can improve recall of where information is “at the expense of remembering what it actually is”. Photographing museum objects impaired memory for them a day later even when the photographs were not expected at test (Henkel 2014), so “the de-prioritization of information that is potentially available externally might occur spontaneously”.
  • Spatial memory. A turn-by-turn navigation group drove a route better but remembered its scenes worse and drove it worse a second time without the aid (Fenech et al. 2010).
  • Automation (Box 3): complacency, “the failure to be sufficiently vigilant with respect to the performance of automated processes”, and automation bias, “the tendency to uncritically rely on the output of a automated decision aid”; and “skill decay” under long reliance. Whether freed resources help depends on where they go.
  • Metacognition. Searching the Internet inflates estimates of one’s own knowledge; expected Internet access made people answer “don’t know” more and report lower feeling-of-knowing.
  • Drift. Relying on an accurate GPS lowers spatial memory (arrow c) and confidence (arrow b), which makes relying on it again more likely (arrow a): “the model predicts a kind of self-reinforcing pattern that will produce a drift away from reliance on internal capabilities when situated in an environment with effective cognitive technologies”.
• The metacognitive model (Figure 3). Arrow a: strategy selection follows a “metacognitive evaluation of the available options”, informed by beliefs “relating to the person, task, and strategy” and by experiences such as effort, fluency, past success with internal and external storage, and the believed reliability of the external store. Arrow b: engaging in a strategy changes later metacognitive evaluations. Arrow c: offloading “can also directly impact our lower-level cognitive processes”. The model is a box diagram; it is not formalised and makes no quantitative prediction.
• Practical. Compensatory offloading should be adopted mainly by people aware of their impairment, which predicts under-use after brain injury. In education, “any benefit of offloading will be contingent on the fact that the demand being offloaded is unnecessary with respect to the learning goal”.

Bearing on the record

Computational-cost framing. Not in any formal sense. The definition is in terms of “cognitive demand”, and the paper says offloading helps us “minimize computational effort”, but it never sets a cost of offloading against a cost of doing the work internally, never counts uses, and never treats the offloaded item as a stored result of an earlier computation. The one place it names the currency, Box 2, leaves open whether the choice minimises time (Soft Constraints, Gray and colleagues, cf. LIT-1058) or cognitive effort specifically. Its account of when people offload is descriptive and metacognitive: they offload when they believe the internal route is harder or less reliable, and the belief can be wrong.

On QUESTION-095. This is not the formalisation the owner hoped for. It does not say that retrieving an offloaded result is a kind of amortized computation, and it has no account of when offloading pays. What it adds is three things the question’s unification has to accommodate.

• Decision and value come apart. People offload where it gives nothing (about 40% at a load where performance is at ceiling) and offload more or less with confidence when ability is held fixed. So an account of when offloading pays is not an account of when people offload; the question asks the first, and the review answers only the second.
• Retrieval is not free. “you no longer need to remember the content of the file, but you do need to remember where to find it.” That is the access cost the question’s memoization picture needs, named but not measured.
• The candidate’s staleness cost has a near relative. Automation bias, uncritical reliance on an external aid’s output, is the risk of reusing a result without checking it; the review does not frame it as staleness.
• The two types map unevenly onto the question. Offloading onto the body is “here-and-now” and stores nothing for later; only into-the-world offloading leaves a result in the environment, as the cockpit’s speed bugs do.

On CLAIM-303. Partly supports the direction of the claim, and narrows it in two places.

• Intention offloading is paying now (a reminder set at the start of a trial) against a later cost, and people do more of it as the later internal cost rises with load and interruptions. That fits the claim’s c_r side: the costlier reconstruction would be, the more is moved forward.
• Nothing here bears on the claim’s n. The reviewed offloading is for a single use; reuse frequency is not a variable in any study cited.
• The claim’s ledger holds internal ability fixed. The review says it is not: offloading lowers memory for the offloaded content and, by the drift prediction, internal capacity and confidence over time. Filing then raises later reconstruction cost and favours further filing. That is a dynamic the claim’s fixed-workload inequality leaves out.
• The superfluous-offloading results are the empirical counterpart of filing that does not pay, and they say the decision to file, if made the way people make offloading decisions, is driven by belief about cost rather than cost.