The Evaluating Self

22Cognitive Offloading

Evan F. Risko and Sam J. Gilbert, Trends in Cognitive Sciences 20(9), pp. 676–688, 2016

Linked to 17 chapters: supports 10 · mechanism 9 · same claim 4 · pairing 4

Cognitive offloading, 2016 — two questions, one model What makes a person offload What offloading does to cognition Demand predicts it: more rotation, more items, more intentions, interruptions. But the estimate of demand decides. Equal costs, unequal tilting of the head. Low confidence sets more reminders, at any level of actual ability. Saved information is recalled less; where it was saved is recalled better. A search engine raises the estimate of what one knows without it. Devices, cameras, automation: the same trade, and the same decay of what goes unused. Offloading is chosen on an estimate of one's own capacity, and the choice alters both the capacity and the estimate — a drift toward the external wherever the tools are good.
Figure 22.0 — The two questions the review asks of one behaviour, the answers the experiments give, and the model that joins them.

A reader faced with a page of rotated text can rotate the page, tilt the head, or turn the letters upright in imagination. Someone who must take a parcel to the post office can hold the intention in mind or leave the parcel by the front door. Risko, at Waterloo, and Gilbert, at University College London, gathered the experiments on choices of this kind and gave the behaviour one name and one definition: cognitive offloading is "the use of physical action to alter the information processing requirements of a task", performed so as to reduce what the head has to do. The practice is old — counting on fingers, the abacus, the knotted quipu — and until recently experimenters treated it as contamination to be prevented rather than behaviour to be measured. The review asks two questions of it. What decides whether a person offloads or thinks unaided? And what does offloading do to the thinking that remains? The answer to the first is that the demand a task would otherwise impose predicts offloading, but only through an estimate of that demand which can be, and demonstrably is, wrong. The answer to the second is that offloading changes both the memory it replaces and the estimate that will be made next time, and that those two changes push in the same direction.

Thinking with the body

Risko and Gilbert divide offloading by where the demand is sent: onto the body, or into the world. The body-based kind is the less obvious of the two, and the review collects a repertoire of it. People use their eyes to hold places in space rather than encoding positions, mark items with a finger or a nod of the head while counting, move their hands to put a partial thought outside the head while explaining arithmetic, and gesture to simulate a spatial transformation instead of performing it internally. In each case an action appears unprompted in the middle of a mental act and buys some saving.

Offloading: an action that changes the task, not the person The task rotated text on a page; an intention to keep Internal processing mental rotation, rehearsal Physical action that alters the task the whole cost paid in the head Onto the body tilt, gesture, count on fingers Into the world write it down, save it, set a cue Both routes reach the same goal. The second pays in movement and perception what the first pays in memory, and leaves the saving in the body's posture or in the state of the world.
Figure 22.1 — The definition drawn as a mechanism. An action changes the task rather than the person, and the same goal is reached with less of the work done internally.

The case the review works hardest is external normalization: the head tilt in front of rotated text. Turning a stimulus upright in the head is mental rotation, and it costs time in proportion to the angle. Risko and colleagues, in 2014, put arrays of words on a screen at various orientations and left participants free to move. Spontaneous physical rotation became more likely as the display was turned further from upright and as the number of words in it grew, and both manipulations also raise the cost of doing the rotation internally. The same shape — more internal demand, more offloading — has since been found for prospective memory, short-term memory, gesture while speaking and gesture while thinking.

Demand and offloading can nevertheless be pulled apart, and the experiment that does it is the hinge of the paper. Dunn and Risko, in 2015, compared displays in which the words and the surrounding frame were both rotated against displays in which the words alone were rotated inside an upright frame. The two conditions produce the same rotation costs and the same response on a physiological measure of demand. Head tilting was far commoner in the first. What differs between them is belief: participants report that the rotated word-and-frame displays take longer, produce more errors and feel more effortful to read, and they act on the report rather than on the cost. The proximate cause of offloading is an evaluation of demand, not demand.

Putting cognition into the world

The second route uses the environment as a store, which removes the need for an internal representation: a note written down, a file typed into a computer, a sketch, an object moved.

Delayed intentions are where this has been studied most carefully. Prospective memory is the memory for things to be done later, its failures make up most of what people report as everyday forgetting, and its difficulty is that nothing in the environment necessarily arrives to trigger the act, which must therefore be self-initiated. People have always answered that difficulty with external triggers: a knot in a handkerchief, a note on the door, and now a phone that can cue by time, by place or by person. Gilbert calls this intention offloading and built a laboratory task for it. Participants drag numbered circles in sequence to the bottom of a screen, having been told that one or more particular circles go somewhere else instead; they may hold those exceptions in mind, or drag the circles toward their eventual destination at the start of the trial and let the screen remember. Dragging the circle to the edge is the laboratory version of leaving the parcel by the door.

Offloading in that task rises with the number of intentions to be held and with interruption, both of which lower accuracy when reminders are forbidden. It also tracks confidence. In a study in which participants performed the task with and without the ability to set reminders and predicted their own performance, those with lower confidence in their memory set more reminders, with objective ability held constant — objective ability predicted reminder use as well, and independently. The confidence involved is not specific to memory. When the same people judged which of two displays was brighter, in a procedure that held their accuracy constant, those less confident in their perceptual judgments set more reminders in the memory task. Something domain-general about how a person rates their own mind is doing part of the work.

Demand, and the offloading it produces how often people offload at two items nothing is gained by writing, and on 40% of trials they write anyway internal demand, measured where offloading is prevented Schematic. The same rise appears for rotated text, delayed intentions and words held in mind. The line does not start at zero: offloading persists where the unaided system is already perfect.
Figure 22.2 — Schematic. Across the domains the review collects, the rate of spontaneous offloading rises with the demand measured when offloading is prevented; and offloading persists at loads where the unaided system is already perfect.

That last point is where the account stops being a description of sensible behaviour. Risko and Dunn, in 2015, let participants write down words they had to hold for a few seconds. With a load of two items, unaided performance was already at ceiling, and they wrote the items down on about 40% of trials. They also judged, wrongly, that writing them down would improve their scores. Reminders are likewise set in conditions where they produce no gain in accuracy, and heads are tilted where tilting saves nothing. The review offers three readings — a benefit too small to detect, a standing bias against mental as opposed to physical effort, or a mistaken belief about what offloading will buy — and takes the demonstrated error in the belief as support for the third. The second reading is an old dispute the paper sets out rather than settles: whether the system is impartial between kinds of effort and simply takes the quicker route, or whether it is a miser that will pay in perception and movement to spend less in memory.

Remembering where instead of what

A transactive memory system is one in which what is known is distributed across several people, so that the group holds more than any member. The review extends the idea to systems made of a person and a machine, and reports what has been found about the consequences, which is the half of the literature that has reached the public.

The best-known experiment had people type trivia statements into a computer, half of them told the file would be saved and half that it would be erased. Those who expected erasure remembered more. The finding has a second face: when a first list was saved, memory for a second list improved, which its authors attributed to the saved list no longer interfering. Offloading into a store is not a free pass, because the requirement changes rather than disappearing — a shift, in the review's phrase, from remembering what to remembering where. Saving a file improved memory for where the information could be found at the expense of memory for what it said, and after difficult trivia questions, words to do with internet search engines produced more interference in a Stroop task, as though the thought of where to look had been primed by the failure to recall.

From remembering what to remembering where before after In the head the content itself offloading In the head where to find it In the world the content itself What it buys the intention is carried out more often where it was put is recalled better the list handed over stops interfering What it costs the content itself is recalled less well confidence in unaided knowledge rises the unused skill decays: routes, scenes The memory requirement is not removed but changed, and the two columns arrive together: the same act that makes the intention safer makes its content harder to recover unaided.
Figure 22.3 — What the shift into an external store costs and what it buys. The content leaves the head and a pointer to it stays behind.

The effects reach the estimate as well as the memory. Searching online about one subject led people to credit themselves with more knowledge in the head, and more brain activity, when answering questions about a different subject. People who had just used a search engine on a quiz reported higher cognitive self-esteem and predicted better performance on a further quiz they would have to take unaided. The blurring is not universal: when participants knew that a question they could not answer would be looked up, they said they did not know more often and reported a lower feeling of knowing. Access to an external store made them less willing to offer what they had.

Other technologies exact the same kind of price. Drivers who followed turn-by-turn navigation performed better on the drive, remembered less of the scenery, and did worse when asked to drive the route again without the device. Photographing an object impaired memory for it a day later, and photographing only part of it reduced the impairment. Where the offloading is designed into the equipment rather than chosen, the familiar automation costs follow: insufficient vigilance toward the automated process, uncritical acceptance of its output, and the decay of a skill that is no longer exercised. Whether a highly automated vehicle improves or degrades a driver's awareness of the road depends on where the freed attention goes.

A metacognitive model of the decision

From this Risko and Gilbert build the model the paper is written for. Where a goal can be reached in more than one way, offloading is a strategy, and it is chosen. The choice is made on a metacognitive evaluation of the options — beliefs about oneself, about the task and about the tool, together with experiences such as effort and fluency — and that evaluation, rather than the objective demand, is what the selection sees. Choosing is not the same as deliberating: gesture arrives without anyone noticing it, while typing coordinates into a navigation device instead of recalling the way has the full phenomenology of a decision, and the authors mark the range as something to be explained rather than assumed.

The decision, and the loop it closes The evaluation of both ways of doing it chooses The strategy internal or offloaded then The act tilt, write, set a cue changes The capacity it replaced memory for the content, for the route confidence in it falls, and the next choice is easier Each turn makes the next offload likelier. The capacity replaced weakens, the estimate of it falls, and the strategy that lowered it is chosen again — a drift away from the unaided mind.
Figure 22.4 — The model redrawn as a loop. The evaluation selects the strategy, the strategy weakens the capacity it replaced, the weakened capacity lowers the evaluation, and the next choice is easier to predict than the last.

The two arrows leaving the act are what make the model more than a restatement. Offloading acts on lower-level cognition, aiding and impairing memory as the saving experiments show. It also acts on metacognition: having offloaded changes what a person believes about their own capacities. Put those together and the model predicts a loop. Trusting a navigation device more than one's memory of the route is a reasonable judgment of relative accuracy; following it weakens the spatial memory and lowers confidence in it; the lowered confidence makes the device the obvious choice next time. The authors call the result a drift away from internal capability wherever the external tools are good, and point to the erosion of Inuit wayfinding as a case of it outside the laboratory. How far that drift goes over years is, they concede, unknown.

Practical implications and what is unknown

Two applications follow directly. The people who would gain most from offloading are those whose unaided abilities are impaired, and the model says such compensation will be adopted only by those whose evaluation of their abilities has kept up with the impairment. After a brain injury that mismatch is exactly what is expected, since the self-estimate was built over a lifetime and the capacity changed in an afternoon; improving insight is therefore a route to better use of external aids. In education, gesture lightens the load for children learning arithmetic, and does so more effectively than offloading onto physical manipulatives, which suggests the forms of offloading are not interchangeable. The benefit holds only where the demand removed was not itself the thing to be learned, and only where what is saved is spent on the task rather than on something else.

The review is candid about how little is settled. Almost all of it comes from adults, and how offloading strategies appear in childhood and change across a life is listed as open. So is whether the large differences between individuals are domain-specific or general, whether the various forms of offloading — pen and paper against a computer, the internet against a friend — are cognitively equivalent, what triggers offloading in the brain, whether the system is impartial between kinds of effort, and what interventions would make a person offload more or less. The long-term question is put in Carr's words, "Is Google making us stupid?", and left unanswered. The methodological obstacle is named too: studying offloading means letting participants move and handle things, which is the opposite of what laboratory procedures were built to prevent.

Since 2016

The account of the decision has held and has been tightened where the review could only infer. The correlation between low confidence and heavy reminder use has been turned into a causal demonstration: manipulating confidence without touching ability changes how much people offload, which is the arrow the paper drew and could not test. The bias has been measured in both directions — under-offloading by people who trust a memory that will fail them, and over-offloading of the sort the ceiling-performance result first exposed — and intention offloading has become a paradigm in its own right, extended to children, to older adults and to clinical populations, and reviewed at length by Gilbert and colleagues in 2023. The developmental question the paper listed as open has begun to be answered: the tendency to use external reminders appears early and the metacognitive steering of it matures slowly.

The Google-effects experiments have travelled least well. They arrived at the start of a decade in which social psychology's headline findings were re-examined, independent attempts to reproduce the saving-and-forgetting result have come out mixed, and the claim that search engines are primed by failures of recall is now held more loosely than its citation count suggests. What has survived is the weaker and more useful part: expecting external access changes what people encode and what they say they know.

The technologies have moved faster than the science. The model was written for notes, files, navigation devices and search engines, and applies without amendment to whatever else takes over a piece of the work, since its terms are the demand, the estimate of the demand and the loop between them. What the review asked for in 2016 — evidence on what years of offloading do to the unaided mind — is still largely missing, and the ease of offloading has increased in the meantime.

Claims drawn from this chapter · 8

opinionA metacognitive model of the decision

A tool weakens the skill it replaces and your judgement of it

Risko and Gilbert argue that handing a task to a tool changes two things: the ability it replaced, and your own estimate of that ability, which is what decides whether you use the tool next time. Use therefore drifts towards the tool.

Becausethe model's two arrows are illustrated by navigation devices and by the erosion of Inuit wayfinding, and the confidence-to-offloading arrow has since been demonstrated causally; how far the drift goes over years remains unmeasured.
LimitsThe two arrows are a model with one arrow now demonstrated; the long-run drift is the paper's central prediction and the thing nobody has measured, which is why this is an opinion rather than evidence.
PracticeKeep unaided hours on the capacities you want to retain, and schedule them rather than waiting to feel the need. The feeling is the second output of the loop, and it moves in the wrong direction.
After AIThe loop is the whole argument of this book in one diagram. What has changed is that the capacity being replaced is now judgement, and judgement is what the loop uses to decide.
evidenceThinking with the body

People offload work when it feels hard, not when it is hard

Risko and Gilbert hold that what makes people offload a task — do it with a tool or the body instead of the head — is their judgement of how hard it is, not how hard it actually is.

BecauseDunn and Risko 2015 rotated the words alone inside an upright frame and rotated words and frame together, producing the same rotation costs and the same physiological measure, and head tilting was far commoner in the condition participants believed was harder.
LimitsDemonstrated on a perceptual manipulation where belief and demand could be separated cleanly; how far the dissociation extends to complex work is not measured.
PracticeAudit what you hand off against what it actually costs you, not against how hard it feels. The feeling is the variable doing the work, and it can be wrong while the cost is unchanged.
After AIEvery decision to ask a model is made on an estimate of what the task would cost you. That estimate is the thing this literature shows to be unreliable, and it is now consulted many times a day.
evidencePutting cognition into the world

People reach for external aids as a task gets harder to hold

The harder a task would be to do in the head, the more people spontaneously push part of it out into the world — turning the page rather than rotating the shape mentally, or writing the reminder down.

BecauseRisko and colleagues 2014 found it rising with rotation angle and array size, and intention offloading rises with the number of intentions held and with interruption, both of which lower unaided accuracy.
LimitsThe relationship is between offloading and demand in laboratory tasks with clean load manipulations; in real work, availability and habit compete with demand as predictors.
PracticeExpect your own delegation to climb exactly when the work gets hard or you get interrupted — which is when the judgement you are delegating matters most. Decide the boundary in advance rather than in the moment.
After AIThe conditions that drive offloading up — high load, interruption, fatigue — are the conditions under which you are least able to check the result. The tool arrived without changing that.
contestedRemembering where instead of what

People who can search online think they know more than they do

Risko and Gilbert report that being able to look things up raises how much knowledge people credit to their own heads, though the strongest of these Google-effect studies have replicated unevenly.

Becausesearching online about one subject raised people's self-credited knowledge, and their reported sense of brain activity, on a different subject — a transfer that only an inflated self-estimate explains.
LimitsContested: these studies arrived just before the replication decade and the strong form has not held; the surviving version is that expecting access changes what people encode and what they say they know.
PracticeLog what you can do unaided, periodically and in writing. Access to a store inflates the estimate of what is in your head, and the estimate is what you will use when deciding whether you can check the tool.
After AIThe effect was measured with a search box. A system that produces finished work in your own idiom is a stronger manipulation of the same variable, and nobody has measured it over years.
contestedRemembering where instead of what

People who expect a file to be saved remember where not what

Risko and Gilbert report that when people expect information to stay available, they store where to find it instead of the content itself. Memory is not lost; what gets stored changes.

Becausein the saved-file experiments expected erasure improved recall and saving improved memory for location, which is a shift in what was stored rather than a loss of storage.
LimitsContested: independent attempts to reproduce the saving-and-forgetting result have come out mixed, and the search-engine Stroop priming that supported the strong version is now held loosely.
PracticeDecide deliberately which things you want in your head and rehearse those without the tool available, because the expectation of access is doing the encoding decision for you otherwise.
After AIWith a model the store is not even a location you could return to — the answer is regenerated, differently, each time. The weaker form of this claim is the one that transfers, and it transfers completely.
evidencePutting cognition into the world (Risko and Dunn 2015)

People write things down when they gain nothing and think it helps

Asked to remember just two items, which they could already do perfectly from memory, people still wrote them down on about 40 per cent of trials. They also wrongly believed that writing them down would raise their scores.

Becausethe same pattern appears in reminders set where they produce no accuracy gain and in head tilts that save nothing, so the bias is visible across three paradigms with different costs.
LimitsCeiling performance makes the gain of zero easy to establish and also makes the stakes trivial; the experiment shows a misjudgement, not a cost.
PracticePut a step between the impulse to delegate and the delegation — commit to your own answer first, then compare. The impulse is not tracking whether the help helps.
After AIMETR's developers were 19 per cent slower while believing themselves 20 per cent faster. The same bias, measured where there was something to lose, and the belief survived the experience.
evidenceThinking with the body

Tilting your head, gesturing and setting reminders are the same behaviour

Risko and Gilbert treat tilting your head at rotated text, gesturing through an arithmetic step and setting a phone reminder as one behaviour: physical action taken to reduce what the head has to do. All of them respond to the same manipulations.

Becauseexternal normalization, prospective memory, short-term memory and gesture all depend in the same way on load and on how well the person performs when offloading is prevented.
LimitsThe unification is by shared dependence on load and on unaided performance, which is weaker than showing one mechanism; the behaviours differ enormously in what they cost to reverse.
PracticeAudit your offloading as one category with many instances. The notebook, the search, the script and the model are the same move, which means what you know about one transfers to the others.
After AIA model is the widest external resource yet, and it takes on the deliberative work the earlier examples did not touch. The category holds; the range inside it has changed.
evidenceRemembering where instead of what

Turn-by-turn navigation makes the journey better and the driver worse

Drivers following turn-by-turn navigation drive the route better, remember less of what they passed, and do worse when they drive the same route again without it.

Becausethe navigation studies the review collects show all three, and the photography analogue points the same way — photographing an object impaired memory for it a day later, while photographing part of it reduced the impairment.
LimitsSingle-session studies on route memory, with the attention account and the encoding account both live; the photography result showing partial engagement reduces the loss suggests the effect is about attention rather than about the device.
PracticeSeparate the outcome you want from the capacity you want, and keep hours on the second. If you want to know a city, drive it without the device sometimes; the guided drives will not accumulate into knowledge.
After AIThis is the clearest consumer-scale instance of the pattern, and it happened without anyone deciding to accept the trade. The same shape is now available for writing, analysis and code.

All claims · Source map