The Evaluating Self

86GPS and the Cognitive Map

Louisa Dahmani and Véronique D. Bohbot, Scientific Reports 10, 6310, 2020

Linked to 7 chapters: supports 4 · mechanism 4 · same claim 2 · pairing 2

GPS habits and spatial memory, 2020 — two studies Fifty drivers, tested once Thirteen, retested at three years More lifetime GPS, less use of the map strategy: r = -0.22 on the first probe. Fewer landmarks noticed and used, and a worse map drawn from memory, -0.22. Heavy users did not report a poorer sense of direction: r = 0.07. More hours since the first visit, a steeper fall on the second probe: r = -0.68. Map drawing fell with hours, -0.52, and with rising reliance, -0.64. Sense of direction did not fall with it, so the decline did not cause the use. Following instructions is stimulus-response learning; the spatial system it replaces weakens in proportion to the hours, and the weakening goes unnoticed by the person it happens to.
Figure 86.0 — The two studies, the measures each moved, and the confound the authors set out to exclude.

By 2020 a satellite navigation device had sat on the windscreen of most cars for a decade, and the experiments on it had all asked one question: how much of a route does a driver learn while the machine is talking? That answer was settled. People guided turn by turn estimate directions worse and draw the shape of a route worse than people who read a map or walked it once behind an experimenter. Louisa Dahmani and Véronique Bohbot, at McGill, asked what happens to a person's own sense of a place on a day when the device is not there at all. They recruited fifty regular drivers in Montreal, reconstructed how many hours of their lives each had spent under guidance, and put them through two virtual mazes built to separate two ways of learning where things are. Those with more lifetime GPS hours leaned less on the strategy that builds a map, drew worse maps from memory, and had noticed fewer landmarks. Three years later, in a follow-up nobody had planned, thirteen came back.

Two systems for finding the way

Anyone crossing an unfamiliar neighbourhood without help can solve the problem in either of two ways, and the two are not variants of one skill. The first learns the positions of landmarks relative to one another and assembles them into a map: the target is a little to the right of the pyramids, the pond is behind me. It depends on the hippocampus, the structure O'Keefe and Nadel proposed in 1978 as the brain's cognitive map and the one that carries episodic memory. The second learns a motor response to a cue — at the next corner, turn left — and chains the responses into a sequence. It depends on the caudate nucleus, which also carries habits such as riding a bicycle, and it lets a person drive a familiar route, in Dahmani and Bohbot's phrase, on "auto-pilot".

Two ways to solve the maze, two brain systems Left to themselves, people used one strategy or the other, and said which. Turn left in 500 metres Spatial memory strategy the relations among landmarks Stimulus-response strategy a turn triggered by a cue Hippocampus Caudate nucleus Builds a map of the surroundings that still works from a new angle. Slower to learn, flexible after. Builds a chain of moves that works only from the angle learned. Quicker to learn, rigid after. Lesion work dissociates the two, and they compete: weakening one can strengthen the other.
Figure 86.1 — The two strategies, the structures they depend on, and where a turn-by-turn instruction enters.

The separation is not a matter of emphasis. Lesion work in rodents gives a double dissociation: damage the spatial circuit and place learning fails while response learning is spared; damage the response circuit and the reverse holds. The two also compete, in that impairing the hippocampal circuit can make response learning easier rather than harder.

That sets up the paper. Following a GPS means executing step-by-step sensorimotor instructions — turn right at the next intersection, turn left in 500 metres — which is what stimulus-response learning consists of. A driver who works by instruction for years is practising one of the two systems and resting the other. Dahmani and Bohbot's hypothesis is that the resting one gets worse, and that the loss shows up on a day the driver has to find the way unaided.

The two mazes

Both tasks are virtual radial-arm mazes, built in a game engine and modelled on tasks used in rodents, each with a pit at the end of every arm and a stretch of country around it — a mountain range, trees, pyramids, a pond, boulders.

The Concurrent Spatial Discrimination Learning Task has twelve arms, grouped into six pairs of neighbours, one arm of each pair hiding an object. Participants meet the pairs one after another until they reach eleven correct choices out of twelve across two consecutive trials. Then the pairs are recombined: path three, shown until now beside path four, appears beside path two. The objects have not moved. Someone who learned the relation between the target and the pyramids still walks to the right arm; someone who learned a turn walks to the empty one. Errors on the probe therefore measure how little the spatial strategy was used. A second probe follows, and since people generally improve on it — the mean gain was 0.73 of a point — failing to improve measures rigidity.

What each probe takes away, and who then fails Twelve arms, six pairs Eight arms, four barred Learn six pairs to criterion Learn four hidden arms Probe: pairs recombined Probe: a wall hides the view Map user: the relation to the pyramids still holds, and few errors follow. Map user: the landmarks are gone, and the errors climb. Sequence user: the learned turn now leads to the empty path. Sequence user: the counted path is unaffected. The two probes score the same strategy in opposite directions On the twelve-arm maze a low probe score means the spatial strategy was used; on the eight-arm maze a high error count behind the wall means the same thing. A map drawn from memory and a count of the landmarks noticed are scored from the same session.
Figure 86.2 — What each probe takes away and who then fails. The two probes score the same strategy in opposite directions.

The 4-on-8 Virtual Maze has eight arms. Four are barred while the participant collects objects from the other four; then all eight open and the four already visited must be avoided. Its probe inverts the first task: a wall rises around the maze and hides everything beyond it. Now the person who had been using landmarks is the one who fails, and the errors measure landmark reliance. Afterwards each participant draws the maze from memory, scored a point for every landmark drawn and a point for every one correctly placed, and gives a structured verbal report of how the task was solved. Two raters scored the reports from nought to three, from a sequence using no landmarks up to a spatial strategy resting on several.

The GPS side was measured with a questionnaire the authors built for the study. Interviewers reconstructed each participant's lifetime hours of GPS use, and of driving, by asking how much the device was used in a typical week, how long that weekly amount had held, and how the pattern had changed over the years. Two scales went with it: seven items on how often it is reached for, and thirteen on how dependent the person feels without it. Subjective sense of direction was taken separately, on the Santa Barbara scale. Lifetime driving hours correlated with none of the navigation measures, so driving experience did not arise as a confound.

Fifty drivers

Sixty people aged nineteen to thirty-five were tested and ten set aside, six of them because they did not drive the required four days a week. The remaining fifty — eighteen women, thirty-two men, mean age 27.6 — were all local drivers, and most used a GPS at least weekly. Correlations were Pearson's, with bootstrapped bias-corrected intervals in place of adjusted p values, one-tailed where the direction had been predicted in advance.

Lifetime GPS hours ran against the spatial measures across both mazes. On the twelve-arm maze the first probe fell as lifetime hours rose, r = −0.22. On the eight-arm maze the same predictor went with a lower strategy score from the verbal reports, a worse drawn map, fewer landmarks used, and fewer landmarks noticed at all, the last the strongest at −0.26. Reliance on the device in daily life tracked the combined probe score and the landmark count. The sense-of-dependence scale, with subjective sense of direction held constant, tracked the second probe at −0.33: those who felt most dependent were the least able to correct themselves after the first probe misled them.

Lifetime hours of GPS use against each measure, fifty drivers correlation, with the bootstrapped 95 per cent interval -0.4 -0.2 0 +0.2 Twelve-arm maze, probe 1 Twelve-arm maze, both probes Eight-arm maze, strategy score Map drawn from memory Landmarks used Landmarks noticed Sense of direction -0.22 -0.20 -0.20 -0.22 -0.21 -0.26 +0.07 The teal row is the confound test: a person's own rating of their sense of direction is unrelated to their GPS hours.
Figure 86.3 — Lifetime hours of GPS use against each measure, fifty drivers, with the confound test in the last row. Correlation and bootstrapped 95 per cent interval, from Table 2 and Figs 2 to 4.

These are small effects. A correlation of −0.22 accounts for about five per cent of the differences between people, and none of the intervals sits far from zero. The authors press on one further question: whether the pattern is only about which strategy people pick. Among the seventeen who had used a spatial strategy from the start, greater lifetime GPS experience still went with noticing fewer landmarks, r = −0.37, and every other measure leaned the same way without reaching significance in so small a group. Heavy users were not merely choosing the other system; they were worse at the spatial one when they used it.

Three years later

The follow-up was an afterthought. All fifty were invited back, but the sample had been mostly undergraduates and most had left the city, so thirteen returned after an average of 3.23 years. At baseline they had not differed from the rest on demographics, GPS habits, sense of direction or navigation. Over the interval their reliance scores rose by 5.5 points on average while their sense of dependence did not move, and every one of them had used a GPS during the three years.

Each person's change score was correlated with the hours logged since the first visit and with the change in reliance, holding baseline performance and baseline GPS use constant. The strongest result is on the flexibility measure: the more hours a person had spent under guidance, the further the second probe score fell, r = −0.68. Rising reliance tracked the same fall at −0.47 and went with needing more trials to reach criterion at all. On the eight-arm maze the drawn map deteriorated with hours used, r = −0.52, and with rising reliance, −0.64; landmarks used fell with hours, and landmarks noticed fell with reliance at −0.67. Nothing else moved. The neuropsychological tests were unchanged, and at a mean age of twenty-eight nobody expected age to be doing the work.

Change over three years against GPS use in the interval thirteen participants; partial correlation, baseline performance and baseline GPS held constant -1.0 -0.5 0 +0.5 +1.0 Probe two, hours used Probe two, reliance Trials to learn, reliance Map drawing, hours used Map drawing, reliance Landmarks used, hours Landmarks noticed, reliance Sense of direction, hours -0.68 -0.47 +0.62 -0.52 -0.64 -0.32 -0.67 +0.29 More trials to learn is a worse result, so the third row agrees with the rest. The teal row again crosses zero.
Figure 86.4 — Change over three years against GPS use in the interval, thirteen participants. Partial correlations, baseline performance and baseline GPS held constant, from Table 2 and Figs 5 and 6.

Thirteen people is thirteen people, and the intervals show it: the one on the strongest result runs from −0.91 to −0.10. The authors say plainly that spurious correlations are possible. What the sample does give is a dose relation inside individuals — the same people, measured twice, declining in proportion to the hours.

Which way the arrow runs

The obvious objection is that the causation runs backwards: people who cannot find their way buy a GPS and use it constantly, and the correlation would follow without the device doing anything. The Santa Barbara scale is in the battery to test this. If compensation were the story, heavy users should report a poor sense of direction. They did not. Lifetime hours and subjective sense of direction were unrelated, r = 0.07, and so were reliance and sense of direction. Only the sense-of-dependence scale went with feeling badly oriented, at −0.58, which is close to a restatement of what that scale asks, and its effects were reported with the sense-of-direction score partialled out. In the three-year sample, change in sense of direction tracked neither hours used nor rising reliance.

So the spatial measures fell while the self-assessment stayed where it was, and that is more interesting than the causal claim it supports. Risko and Gilbert had argued that the decision to hand a task to a device is made on a metacognitive judgment rather than on the true demand, and that using the device alters both the capacity and the judgment. Here the two come apart: the ability declines quietly and nothing in the driver's self-report registers it. A second oddity points the same way — participants who rated their sense of direction highly learned the twelve-arm maze faster but made more errors on the second probe, confident and rigid together.

On mechanism the authors separate their account from the leading alternative. Gardony and colleagues had shown in 2013 that a navigational aid impairs spatial memory by dividing attention, but their aid gave bearing and distance and left the route decisions to the user. A turn-by-turn instruction leaves nothing to decide, and Dahmani and Bohbot propose disengagement rather than divided attention as the route by which the loss occurs. Their second candidate is competition: years of stimulus-response learning may suppress the spatial system the way a lesion to the spatial circuit facilitates response learning. The neural evidence for either is circumstantial and the paper says so. Spatial-strategy users have more grey matter and more activity in the hippocampus, and in 2017 Hugo Spiers's group found that hippocampal activity tracked the demands of a route only when participants chose their own turns. The strongest version runs the other way: London taxi drivers, described in Peak, end up with an enlarged posterior hippocampus that shrinks again after retirement. Nobody in this study was scanned.

The closing pages turn practical. Since engagement seems to be what matters, the authors suggest that devices name landmarks in their instructions rather than only distances and turns, so that the user is looking at something. And they end where the stakes are highest, with Inuit wayfinding, which reads wind, snowdrift and the stars and for which the shortest route between two points is often not the survivable one. Where the machine's route ignores safety and the machine can fail, the decay of the older skill is not academic. It is what Ironies of Automation described in control rooms: the automation removes the daily practice that kept a skill alive, then hands the person the situation in which that skill is needed.

Since 2020

The paper has become the standard citation for the claim it makes, and the thirteen are the reason. Associations between heavy GPS use and poorer spatial ability had been reported before and have been reported since; what nobody else had was the same people tested twice with their intervening hours recorded. That design is also the part most open to attack. Thirteen participants recruited by an unplanned invitation, one-tailed intervals, a battery of correlations across two tasks and four GPS measures, and a defence of bootstrapping as an inherent correction for multiple comparisons that statisticians do not generally accept: the longitudinal result is best read as a reason to run the study properly. It has not since been run at a size that would settle it.

The surrounding evidence has grown. Sea Hero Quest, a game played by millions and analysed by Antoine Coutrot, Spiers and colleagues, showed in Nature in 2022 that wayfinding ability varies with the environment a person grew up in: people raised in cities laid out on grids do worse in irregular places. That is the same argument about habitual demand shaping a capacity, made on a sample four orders of magnitude larger. The finding that guidance quiets the hippocampus has held. The structural claim the paper stops short of has not been shown — nobody has demonstrated that years of guided driving shrink anything.

The tasks measure one thing only: learning a place seen once, which is where GPS use is heaviest and the cost should be largest, and which says nothing about a neighbourhood lived in for years. And the devices have changed since 2019, when these data were collected, in the direction the authors recommended — street-level imagery at the turn, landmarks named in the spoken instruction. Whether a driver looking at a photograph of the intersection is more engaged with the world or less is not something this study can answer. What it does deliver stands: "the greater use of GPS, the greater decline in spatial memory", and "spatial memory deficits are present when people navigate without GPS".

Claims drawn from this chapter · 6

opinionWhich way the arrow runs

Dahmani and Bohbot blame instructions that leave nothing to decide

What costs a driver their sense of a place is not that the device competes for attention but that a turn-by-turn instruction leaves no decision to make. An aid that gives bearing and distance and lets the user pick the turns should cost less.

Becausethey argue it by contrast rather than by test. Gardony's 2013 aid impaired spatial memory by dividing attention, but it left the route decisions to the user; Spiers's group found in 2017 that hippocampal activity tracked a route's demands only when participants chose their own turns. The neural evidence is circumstantial and the paper says so.
LimitsNobody here was scanned, and no condition varied the instruction's form, so the comparison with Gardony's aid is across studies. Devices have since started naming landmarks and showing the intersection, and the paper cannot say whether that leaves a driver more engaged.
PracticeChoose the form of the help, not just the amount. Ask for the material a decision needs and keep the decision; the version that hands over the answer costs the capacity.
After AIThe same tool sits on both sides of the line: asking a model to explain what it would do leaves the decision with you, and asking it to produce the thing does not.
contestedThree years later

Drivers retested after three years declined in step with their GPS hours

Thirteen of the fifty drivers came back after an average of 3.23 years, and the more hours of guided driving they had logged in between, the further their spatial memory had fallen — the same people, declining with the dose.

Becausewith baseline performance and baseline GPS use held constant, hours logged tracked the fall on the flexibility probe at r = −0.68 and the drawn map at −0.52, with the neuropsychological tests unchanged; but with thirteen people the interval on the strongest result runs from −0.91 to −0.10.
LimitsAn unplanned follow-up: most of the sample had left the city, the intervals are one-tailed, and the authors say spurious correlations are possible. It is the standard citation only because nobody else has tested the same people twice with their hours recorded.
PracticeTreat hours under guidance as the quantity that matters, not the decision to use the tool. If the relation is a dose relation, the counter-dose is scheduled unassisted hours.
After AIThree years is roughly how long an assistant has been in daily use, and nobody recorded a baseline first. The within-person comparison this paper barely managed is one no team can now run.
evidenceFifty drivers

Drivers who felt most dependent on GPS corrected themselves least

How dependent a driver felt without the device predicted rigidity, not just weakness: with self-rated sense of direction held constant, the dependence score tracked the second probe at −0.33 — the probe on which people generally recover after the first one has misled them.

Becausethe second probe is scored as a gain over the first, and the mean gain was 0.73 of a point, so failing to improve measures not correcting rather than not knowing; the effect survives partialling out sense of direction.
LimitsOne correlation of −0.33 on a scale the authors built for this study, in fifty people. The related finding that confident navigators learned faster and were then more rigid is incidental and undefended.
PracticeTreat the feeling of being unable to work without the tool as a signal about correction, not only about skill. Where you feel most dependent, build in the step that makes you revise: a second pass with the first answer set aside.
After AIWhat matters in supervising a model is not raw ability but whether the first answer can be revised. Here the people most reliant on the tool were the ones an early wrong turn kept wrong.
evidenceFifty drivers

Drivers with more lifetime GPS hours navigated worse without the device

Among fifty Montreal drivers aged 19 to 35, the more hours of their lives a person had spent under satellite navigation, the worse they did on navigation tasks taken with no device at all: fewer landmarks noticed, worse maps drawn from memory, more errors where a map-like memory was needed.

Becauselifetime hours, reconstructed by interview, gave r = −0.22 on the twelve-arm probe and −0.26 on landmarks noticed, while lifetime driving hours correlated with nothing, so time at the wheel is not the explanation.
LimitsSmall effects: −0.22 accounts for about five per cent of the differences between people, and no interval sits far from zero. The design is cross-sectional, and the mazes measure learning a place seen once, not a neighbourhood lived in for years.
PracticeAsk of each thing you have handed over for years what you would score on it today, unaided. The cost of a delegation shows up on the day the tool is absent, which is not a day that arrives on its own.
After AIA delegation nobody debated, measured a decade in. The assistant now does for writing and analysis what the windscreen device did for routes, and nobody's week contains the equivalent unaided test.
evidenceWhich way the arrow runs

Heavy GPS users scored worse but felt no worse at directions

The drivers whose measured spatial memory was worst rated their own sense of direction no lower than anyone else: lifetime GPS hours and self-rated sense of direction were unrelated, r = 0.07, and over three years the measures fell while the self-rating stayed put.

Becausethe Santa Barbara sense-of-direction scale was in the battery to test the opposite account, that poor navigators buy a device and use it constantly, and it came back flat against both lifetime hours and daily reliance.
LimitsA null result on one self-report scale, in fifty people; the separate sense-of-dependence scale did track feeling badly oriented, at −0.58, which the authors treat as close to a restatement of what that scale asks.
PracticeDo not use your own sense of whether you have lost a capacity as evidence about whether you have. Put the question to a task you pass or fail without the tool present, on a schedule.
After AIRisko and Gilbert argued that a tool changes both the ability and your estimate of it. Here the two come apart, which is the worse case: the faculty that decides how closely to check is the one reporting no change.
evidenceFifty drivers

Heavy GPS users were worse at the strategy they still chose

Among the seventeen participants who used the landmark-and-map strategy from the first trial, more lifetime GPS hours still went with noticing fewer landmarks, r = −0.37. Heavy users were not simply picking the other way of finding a route; they were worse at this one.

Becausethe subgroup was defined by the verbal reports before the correlation was run, so strategy choice is held fixed by construction, and every other spatial measure in the subgroup leaned the same way.
LimitsSeventeen people, and only the landmark count reached significance; the rest of the measures point the same way without support. The subgroup is also self-selected in the sense that strategy use was itself one of the outcomes.
PracticeDo not count still choosing to do something yourself as proof that you still can. Keep the unassisted hours anyway, and look at the result rather than at the decision to attempt it.
After AIThe reassuring version of delegation is that you can always do it yourself when it matters. Here the people who did still do it themselves had got worse at it in proportion to how often they had not.

All claims · Source map