The Map You Never Build: How Spatial Knowledge Shapes Learning, Memory, and What Navigation Tech Takes Away

Ask someone to describe a good memory and they will almost always reach for space. I remember where I was standing. It was on the left, past the blue door. Spatial reference is how human memory is organised far more often than it is acknowledged to be. The same holds for reasoning: the standard test battery for a STEM pipeline measures mental rotation; the standard way to explain a graph is to talk about axes and position; the standard way to teach a child mathematics is with objects arranged in space.

For most of human history, the technologies that connected a person to a place were built out of the environment itself. A landmark, a bend in a river, a slope, a sun position. Learning a place meant building a model of it, because the model was the only wayfinding instrument available.

Satellite navigation broke that arrangement, and it did so at exactly the moment it became cheap and universal. A turn-by-turn device solves the immediate problem — the path from A to B — with a fidelity that no human landmark system can match. What it does not do is build a model of the environment in the head of the person being guided. The research literature on this is now large enough, and coherent enough, to make a specific claim: arriving is not the same as knowing, and the tools that reliably produce the first can quietly reduce the second.

This article synthesizes that literature — what spatial training actually buys in mathematics and science, what digital navigation measurably costs in route memory and environmental knowledge, how the effects differ in children and adults, what the neuroscience can and cannot explain, and where the evidence is weaker than the discourse.

1. What the term "spatial cognition" actually covers

Spatial cognition is not a single faculty. The research literature divides it into at least four distinguishable components, and the most important methodological fact about this field is that they dissociate. A person can be excellent at one and poor at another, and a training program can move one without touching the rest.

• Spatial perception and mental rotation — the ability to manipulate and inspect spatial representations that are not currently in view: folding a paper to match a shape, judging whether two blocks are the same, rotating a map.
• Wayfinding and route learning — the ability to traverse an environment and return, using landmarks, headings, and route segments.
• Environmental knowledge and cognitive mapping — the internal, allocentric representation of a place: a "map in the head" that encodes relationships between landmarks independent of the observer's own position.
• Spatial reasoning in the academic sense — the use of spatial representations to reason about mathematics, science, diagrams, and abstract relations.

The distinction between the third and fourth of these is the crux of the article. A cognitive map is a model of a particular place. Spatial reasoning is a transferable skill for handling any spatial representation, including ones that have no physical referent — an axis, a graph, a phylogenetic tree, a balance scale, a function plot. Evidence that training improves the first does not establish that it improves the second, and a large part of the confusion in the popular literature comes from treating them as the same thing.

A second distinction matters just as much: immediate wayfinding success is not durable route memory. A person can reach a destination reliably, with no idea afterwards what route they took, what turned left, or what the street looked like. This is not a hypothetical dissociation. It is the central finding of the GPS literature.

2. Spatial training and academic outcomes: a real but narrow effect

The strongest causal evidence for spatial training improving academic performance is Hawes, Gilligan-Lee, and Mix's (2022) meta-analysis of 29 controlled pre-post studies (N = 3,765, k = 89 mathematics outcomes) in Developmental Psychology. The average effect of spatial training on mathematics performance relative to control was Hedges' g = .28 (SE = .07), with training also improving spatial thinking itself at g = .49 (SE = .09). The result is robust and the confidence interval excludes zero — but it is a modest effect, and the meta-analysis was unusually careful about its own limits.

Three moderators emerged, and all three point the same direction: the training works best when it is concrete, developmentally appropriate, and similar to the thing it is supposed to help.

• Near versus far transfer. Effects on mathematics outcomes closely aligned with the spatial training delivered were substantially larger than effects on more distally related outcomes. This is the single most important boundary condition in the literature, and it is revisited in §3.
• Concrete versus abstract materials. Training with physical manipulatives produced g = .42, against g = .05 for non-concrete training (b = .36, p = .003). The difference was not subtle.
• Age. Effects grew as participants aged from 3 to 20 (b = .051, p = .006).

Notably, Hawes and colleagues tested and did not find significant moderation by experimental design or by publication status — a result worth stating because a great deal of published spatial-training research has small samples and the field's critics have been eager to argue the effect is an artifact of weak methods.

The best-known individual study in this line is Lowrie, Logan, and Ramful's (2017) ten-week spatial-visualization program with grade-six students (N = 186 in eight classes, randomly assigned to five intervention and three control classrooms). Spatial reasoning improved substantially (F(1,185) = 12.36, p < .001, d = .54), and mathematics performance improved as well (F(1,185) = 7.18, p = .02, d = .40). The change in spatial reasoning correlated with the change in mathematics (r(120) = .21, p = .021) — the kind of relation that is necessary but not sufficient to establish a mechanism.

Two caveats bound the enthusiasm. First, the intervention improved specific subskills — mental rotation (d = .43) and visualization (d = .65) — but not orientation (p = .32). Training did not move the whole spatial faculty; it moved the parts it touched. Second, the study found no gender differences, which matters for the argument in §7.

The correlational counterpart to this literature is Atit, Power, Pigott, and colleagues' (2022) meta-analysis of the relation between spatial skills and mathematical performance, which screened 858 records down to 45 analyzed articles and estimated a pooled correlation of r = .36 (robust SE = .035, τ² = .039), with no significant moderation by gender or grade level. This is the number that most cleanly separates the field's two claims. A correlation of r = .36 says spatial skill and mathematical performance are related. It does not say that training spatial skill raises mathematical performance — a distinction that a great deal of public writing about "spatial learning" quietly discards.

The broader training literature points the same way. Uttal, Meadow, Tipton, and colleagues' (2013) meta-analysis in Psychological Bulletin aggregated 217 training studies published from 1984 to 2009 and found an average effect of g = .47 (SE = .04). Two qualifications matter. First, those 217 studies cover spatial skill broadly, not academic outcomes specifically — the headline figure is not a mathematics effect and should not be quoted as one. Second, a critical review by Sala and Gobet (2019) in Trends in Cognitive Sciences argues that cognitive training programs, spatial ones included, frequently show near transfer to the trained task without far transfer to untrained cognitive abilities, and that published estimates of far transfer are systematically inflated by methodological weaknesses.

The domain-specific evidence is more informative than a simple null, because it holds the training constant and varies the outcome. Porat and Ceobanu (2024) ran a spatial-ability intervention with 154 first-year students (79 engineering, 75 architecture). Spatial ability improved in both faculties. Among architecture students, the gain came with significantly higher grades in every course measured, including studio design (2.85 < B < 9.34, all p < .01). Among engineering students, spatial ability rose but first-year calculus and physics grades did not move.

That is the near/far distinction in a single study: the same intervention, the same gain in the trained skill, and transfer only where the academic task was itself spatial. Architecture studio is a spatial task. First-year calculus, at least as it is taught, is not. Correlation is not causation and near transfer is not far transfer turn out to be two faces of the same limitation.

3. The near/far transfer gap is the whole story

If one result is doing the most work in this article, it is Hawes's near-versus-far split. Spatial training that resembles the target skill improves the target skill, substantially. Spatial training that does not resemble the target skill improves it much less, and in several well-controlled studies not at all.

The numbers from the meta-analysis: g = .653 for near transfer, g = .203 for far transfer, both p < .01. Near transfer is more than three times the far-transfer effect. Far transfer of g = .20 is a real but small effect, well within the range that a field with high heterogeneity (I² = 72%), small samples, and a genuine publication-bias literature can produce. Hawes's rank test (r = .31, p = .031) and Egger test (z = 1.40, p = .162) suggest the mathematics effect is not purely a function of selective publication; for spatial outcomes, trim-and-fill estimated that roughly two studies were missing.

The prediction interval is more revealing than the point estimate. For the pooled effect, the 95% prediction interval runs from −.37 to .93. That means a genuinely new study could land anywhere from a meaningful harm to a large benefit. The pooled estimate is real, but the range of outcomes across settings is wide enough that any specific curriculum or school claiming to deliver a reliable mathematics gain from spatial training is over-claiming.

Why does transfer fail? Three reasons are supported by the evidence rather than merely asserted. First, near transfer can be driven by shared surface elements — the training and the test share stimuli, vocabulary, or procedure — which is not the same as shared cognitive machinery. Second, spatial skills show substantial domain specificity: a person trained on mental rotation is better at mental rotation, but the relation to spatial reasoning in geometry is partial, and the relation to spatial reasoning in chemistry or geology is weaker still. Third, transfer requires the trainee to notice that the trained procedure applies in the new context, and nothing guarantees they will.

The "three mountains" of this literature are worth naming explicitly, because they are where the field's real gains are: near transfer to the trained task is reliable; far transfer to a different domain is small and inconsistent; and transfer to general intelligence is close to zero in well-controlled studies. Public claims about "spatial training boosting STEM" usually assert the second while implying the third.

4. GPS: what the data actually show

The navigation literature is where the distinction at the centre of this article has been measured most carefully, and where the popular framing — that GPS makes you stupid — turns out to be both too simple and, in places, unsupported.

4.1 The correlation is real and modest

Dahmani and Bohbot (2020) in Scientific Reports recruited 60 participants and analyzed 50 (18 women, 32 men, mean age 27.6), with 13 retested three years later. Lifetime GPS experience correlated negatively with spatial memory performance (r = −.22, BCa 95% CI −.41, −.01); GPS reliance correlated negatively with use of spatial-memory strategies (r = −.25, 95% CI −.44, −.03); and dependence on navigation technology correlated negatively with the ability to learn from errors made while navigating (r = −.33, 95% CI −.49, −.14).

That last correlation is the most theoretically interesting. It suggests the deficit is not merely "I don't remember the route" but something closer to "I have stopped learning from my own mistakes" — an impairment in the error-correction loop that would normally build the cognitive map.

Miola, Muffato, Sella, and colleagues' (2024) systematic review and meta-analysis in the Journal of Environmental Psychology pooled 23 studies and found the same pattern with more precision: GPS use correlated negatively with environmental knowledge (r = −.18, 95% CI −.28, −.08) and negatively with sense of direction (r = −.25, 95% CI −.39, −.12), while the correlation with wayfinding performance was non-significant (r = .07, 95% CI −.28, .41).

That non-significant result is the crux of the entire question. Users of GPS are not measurably worse at getting where they are going. They are measurably worse at knowing the place they arrived at. The aid substitutes for exactly the cognitive work that would have built the knowledge.

4.2 The obvious confound, unresolved

A reasonable objection runs: people with poor spatial ability might simply choose GPS more often, producing a correlation that reflects self-selection rather than GPS-induced decline.

Dahmani and Bohbot's own data cut against the simplest version of that story: lifetime GPS use did not correlate with subjective sense of direction (r = .07, 95% CI −.16, .27). But Topete, He, Protzko, Schooler, and Hegarty's (2024) preregistered studies — 249 participants across two experiments — found the opposite pattern for dependence: reported GPS dependence correlated negatively with objective navigation performance and with sense of direction, and positively with spatial anxiety. That is exactly what self-selection predicts.

So the confound is not resolved; it is unresolved in both directions. Cross-sectional correlations cannot distinguish "GPS use erodes spatial skill" from "people with weaker spatial skill and higher spatial anxiety lean on GPS". The causal claim has to rest on experiments, and the experimental evidence is thinner than the correlational evidence.

It is also more equivocal than popular accounts suggest. Yavuz and colleagues (2024), publishing in the Journal of Environmental Psychology under the title "Video gaming, but not reliance on GPS, is associated with spatial navigation performance", analysed 822 Sea Hero Quest players (280 men, 542 women, mean age 26.3). GPS reliance was not significantly associated with navigation performance in either gender. Weekly hours of video gaming were, and once gaming was accounted for, the gender difference in navigation performance disappeared.

The field's own evidence base is thin here. Miola and colleagues rated 69% of the 23 included studies as at moderate-to-high risk of bias, and the strongest single findings come from small samples. Honest summary: the association between heavy navigation-aid use and weaker environmental knowledge is consistent and replicated; the causal claim that navigation aids cause spatial decline is supported by some experiments and contradicted by others; and the size of any effect on wayfinding performance itself is near zero.

4.3 What kind of aid makes the difference

The more useful finding is that the effect is not uniform across navigation aids, and that it can be mitigated.

Gardony, Brunyé, Mahoney, and Taylor (2013) in Spatial Cognition and Computation ran a within-participants comparison (N = 36) of navigating with verbal instructions, with tonal instructions, and with no navigation aid at all. Both aided conditions impaired spatial memory relative to navigating unassisted, and the authors' account is divided attention: the guidance itself consumes the encoding resources that would otherwise have gone into the environment's structure. The critical variable is not screen versus paper but whether the aid supplies the structure or asks the navigator to build it.

Brügger, Richter, and Fabrikant (2019) in Cognitive Research: Principles and Implications opened up the navigation-system black box with 64 participants across four levels of automation on a real outdoor route in Zurich. Participants using systems with higher levels of automation seemed not to acquire enough spatial knowledge to reverse the route without navigation errors; two-thirds of all participants managed the reversal cleanly. Their gaze analysis was descriptive — patterns over route segments, not a group-level test — so the attention mechanism is inferred rather than measured.

Zhang and Li (2025) in the Journal of Environmental Psychology tested whether the problem can be reduced. In a series of virtual-environment experiments they confirmed that GPS dependence is negatively correlated with route knowledge and that GPS-guided navigation impairs route knowledge acquisition, then tested landmark-based instructions as a mitigation. Landmark instructions improved spatial knowledge acquisition, but reduced dependence on GPS only partially. A scaffold that points at the world is better than one that points at the screen; it is not a substitute for the mental work.

Sugimoto, Kusumi, Nagata, and Ishikawa (2021) demonstrated the map side of the same contrast in a route-walking study: participants who navigated with a smartphone map performed significantly worse on route retracing and on spatial learning than paper-map navigators. Paper-map navigation is not nostalgia; it is an environment in which the cognitive work of map-building is not outsourced.

4.4 Older adults: where the aid helps, and what the data do not say

The lifespan literature adds an important nuance that the youth-focused narrative misses: navigation aids are genuinely valuable for people whose spatial capacity has declined.

Muffato, Borella, Pazzaglia, and Meneghetti (2022) in the International Journal of Environmental Research and Public Health surveyed adults across the lifespan and found a negative correlation between age and GPS use (r = −.53), alongside a positive correlation between age and verbal-directions use (r = .16) and a negative correlation between age and map use (r = −.20). Older participants used less GPS and more verbal directions, and reported fewer episodes of losing their way — a detail that matters, because it undercuts the assumption that the aid is compensating for a deficit that grows steadily with age. Sense of direction and spatial anxiety predicted how often people got lost; neither GPS nor map use did. What GPS use did predict was a higher frequency of successfully reaching unfamiliar destinations (β = .12, p = .027) — the aid doing exactly the job it advertises.

A 2025 randomized crossover field study by Nagata and colleagues in Scientific Reports (N = 40, middle-aged and older adults) compared voice-only navigation with paper maps on two 800-metre routes. The two produced no significant difference in goal completion (HR = 1.063, 95% CI 0.755, 1.498, p = .726). Where they differed was mood: the voice-only group reported feeling more relaxed after navigating (F(1, 39) = 5.046, p = .030, η² = .115), with no significant difference in anxiety or fatigue. The study did not measure survey knowledge, so it cannot speak to the map-building question either way.

The honest reading: for older adults, a navigation aid restores the functional capacity to reach a destination, and the one study that measured mood found the voice system pleasant rather than stressful. Nothing in these studies shows older adults escaping the map-building cost; they show that the cost is not paid in anxiety.

5. Children: the age-related digital divide

The child evidence is the part of this story that is most often asserted and least often sourced, so it is worth being precise about what is actually established.

The developmental logic is straightforward: cognitive maps are built, and building them requires ingredients — varied experience of a place, landmarks to anchor to, errors to learn from, and time. A child who is driven by GPS to locations they did not navigate themselves gets the arrival without the building. Whether this is materially harmful is an empirical question, and the empirical literature is smaller than the commentary suggests.

The most directly relevant recent study is Wathiyage Don, Walker, Musculus, and Ruggeri's (2026) paper in Psychonomic Bulletin & Review, whose title states its finding: "Active navigation enhances children's spatial but not episodic memory." The study used 105 children aged 4–10 (M = 92.5 months) in a yoked design — each active child's route was replayed to a yoked peer, so the comparison isolates self-generated choice from the path itself. It was a tabletop map game, not walking through a real environment.

Active children remembered paths closer in length to the route they had actually taken (F(1, 102) = 6.28, p = .014, partial η² = .06) and made smaller area errors (F(1, 101) = 4.29, p = .041, partial η² = .04). There was no condition effect on the number of corners (F(1, 102) = 3.22, p = .075) and no benefit for episodic memory. Two limits follow. The effect sizes are small (partial η² of .04–.06), and the task is a map game, so the result is evidence about spatial choice, not about neighbourhood wayfinding.

Supporting evidence comes from the observation that topographical disorientation is a heritable, developmental condition with a measurable prevalence. Burles and Iaria (2020) in Scientific Reports reported a ten-year study of 1,211 individuals with developmental topographical disorientation against 1,624 controls. DTD participants reported far more first-degree relatives with childhood orientation difficulties (χ² = 266.160, p = 7.796 × 10⁻⁶⁰, OR = 2.914), and both groups reported more female than male relatives with such difficulties. A minority of people have a constitutional difficulty building cognitive maps, and for them navigation aids are not a convenience but an accommodation — a further reason the child evidence cannot be reduced to "screens are bad for children."

Meanwhile, landmark-based navigation develops later than most parents expect, and the transition is measurable. van Hoogmoed and colleagues (2022) in Brain Sciences tested 123 children aged 5–10 in an on-screen virtual environment. Egocentric landmark use was already above chance at five; allocentric navigation began to develop between five and eight; boys outperformed girls on the allocentric task; and visuospatial working memory and the production of left/right predicted parts of the trajectory. Children's navigation is not a small version of adults'; it is a differently structured problem that resolves over years, and intervening in it before it matures is precisely the case in which a device can remove the scaffolding the skill is being built on.

6. The neuroscience: a plausible mechanism, not a complete explanation

The neural story is attractive, and it should be reported with more caution than it usually receives.

The core finding is that the hippocampus and entorhinal cortex contain spatially organised cells. Hafting, Fyhn, Molden, Moser, and Moser (2005) in Nature described the "microstructure of a spatial map in the entorhinal cortex" — grid cells firing in repeating hexagonal patterns as an animal moves through an environment. Doeller, Barry, and Burgess (2010) in Nature reported a grid-like fMRI signal in human entorhinal cortex whose orientation rotated coherently with the direction of virtual movement. Jacobs, Kahana, Ekstrom, and colleagues (2010) in PNAS, recording directly from electrodes in epilepsy patients, found direction-sensitive "path cells" in human entorhinal cortex — a different cell type, and one that is frequently conflated with grid cells in popular accounts.

Maguire and colleagues (2000) in PNAS found that London taxi drivers — who must memorise the layout of a large city for a living — showed increased grey-matter volume in the posterior hippocampus relative to controls, with no equivalent increase anteriorly.

Three qualifications keep this from becoming the explanation it is often made into.

First, the human grid-cell evidence is contested. Kransberg and colleagues (2026) in Imaging Neuroscience scanned 110 adults (72 women, 38 men; ages 18.2–78.6, mean 36.5) during passive navigation and found no significant six-fold grid-like signal in either entorhinal mask (right: t(109) = 0.728, p = .234, BF₀₁ = 7.31; left: t(109) = −0.443, p = .671, BF₀₁ = 8.60). Those Bayes factors are moderate evidence for the null in that paradigm. The failure to detect is not evidence of absence — passive movement and a different task may simply not elicit the signal — but it does mean the human grid-cell findings should be described as replication-contested, not settled.

Second, structural change in taxi drivers is a correlational observation about a selected population. Taxi drivers are selected for spatial ability and then accumulate years of intensive navigation practice; the direction of causation is not established by the imaging, and practice itself may drive the difference.

Third, and most importantly, neural plausibility is not educational explanation. Showing that a region of cortex encodes spatial structure establishes that the brain can do the thing. It does not explain why spatial training moves mathematics scores in some classrooms and not others, why the effect is larger for concrete materials, or why the far-transfer gap exists at all. The neuroscience here is a mechanism candidate — a plausible account of how spatial information is represented — and it should not be used as evidence for whether spatial training improves learning. The educational claims need the educational evidence, and that evidence is more equivocal than the neuroscience is elegant.

Cognitive maps also have a contested pedigree. Tolman's (1948) latent-learning work with rats — in which animals that had explored a maze without reward later took a shortcut through it, apparently having formed a map of the whole apparatus — is the canonical origin of the concept. Wilson and Wilson (2018) in PLOS ONE ran a careful human replication with N = 60 (20 per condition) and found that shortcut behaviour appeared in the Light-Light condition but not in the Ambient condition where the prediction said it should be strongest. The omnibus test was significant (H = 7.89, p = .02), and the Light-Light versus Ambient contrast was significant (U = 106, p = .01), but the Light-Light versus Light-Shifted contrast did not survive a Bonferroni-corrected threshold (U = 116.5, p = .023, against α = .017), and Ambient versus Light-Shifted was non-significant. The result is a partial confirmation with a failed predicted contrast — the kind of finding that does not destroy a theory but does make it a theory with caveats.

Chase and Simon (1973) in Cognitive Psychology supply the complementary principle: the difference between expert and novice performance in chess is not the amount of information stored but its organisation — into chunks and meaningful relational structures. The modern version of this is that expertise in navigation is not about storing more places but about organising places into a usable relational structure. Which is, in the end, the same claim as the cognitive-map idea, arrived at from a different direction.

7. Language, schooling, and cultural variation

The claim that spatial cognition is culturally variable is well supported. The claim that it is caused by language is much weaker, and the distinction has been badly mangled in popular writing.

Languages differ in which frame of reference they habitually use for spatial description — absolute (north/south), relative (left/right), or intrinsic (front/back) — and Majid, Bowerman, Kita, Haun, and Levinson (2004) in Trends in Cognitive Sciences set out the case that this variation is not mere vocabulary but reorganises how speakers solve spatial tasks. The strongest recent test comes from Pitt and colleagues, working with Tsimane' adults in the Bolivian Amazon. In Science Advances (2022), they found that the same participants used different frames on different axes: allocentric (environment-based) coding on the left–right axis, where discriminations are hard, and egocentric (body-based) coding on the front–back axis, where they are easy. That is a direct challenge to the idea that a language group has one predominant frame. And in a 2021 Science Advances paper, the same group showed that Tsimane' adults arrange number, size, and time by magnitude but with no consistent directional bias — a result that undercuts the claim that a left-to-right number line is innate.

The correct reading is narrow and important: language provides a frame that biases spatial judgment, but the frame is chosen per task and per axis, and it does not determine the underlying spatial capacity. The same individuals are not "unable" to think spatially; they think spatially in a different frame and need a moment of translation. Claims that a language "locks you out" of spatial abstraction are not what these studies show.

The schooling evidence is more direct and less ideologically comfortable. O'Shaughnessy and colleagues (2023) in PNAS studied number and arithmetic among the Tsimane', including a meta-analysis of more than 800 children: children were full counters only if they had attended school. In remote communities without formal education, the same researchers documented specialised arithmetic — individuals who solved problems involving five, a number with a distinct role in the local economy, while failing on some lower numbers. That is not a deficit. It is a different, locally adaptive tool set. Gurven and colleagues (2017), following 919 Tsimane' across the life course, found age trajectories similar to industrialised populations, with schooling associated with greater cognitive ability at all ages but no attenuating effect in late adulthood — consistent with a passive-reserve account in which schooling raises the level rather than slowing the decline.

The sharpest demonstration of the trade-off comes from Davis and Cashdan's (2019) study of 6–18-year-old Tsimane' children, tracked with GPS and assessed on pointing, perspective-taking, and mental rotation. Schooling was associated with better mental rotation but worse regional pointing accuracy — plausibly because schooling limits the outdoor exploration that builds large-scale environmental knowledge. Formal education is not a general cognitive enhancer. It trains specific skills and, in this population, appears to cost others.

Cultural variation therefore has a specific, falsifiable shape: the same underlying capacities are deployed in different frames, and the frames are learned. This makes the digital-navigation question a cultural question too. A child in a dense, landmark-rich, walkable neighbourhood is acquiring a cognitive map through the ordinary friction of navigating it. A child in a car-directed, GPS-dependent suburban environment is not. Neither is defective; the second is simply not building the map, because the environment and the tools have been arranged so that it does not have to.

8. What the evidence supports, and what it does not

Distilled to its defensible form:

Well supported.

• Spatial training improves the spatial skills it trains, reliably and substantially (217 studies, g = .47; near transfer g = .65).
• Spatial training produces a small and inconsistent improvement in mathematics (29 controlled studies, N = 3,765, g = .28), larger for concrete materials and for near-transfer outcomes, and larger in older children.
• Heavy navigation-aid use is consistently associated with weaker environmental knowledge and sense of direction (r = −.18 to −.25 in meta-analysis).
• The association does not extend to wayfinding performance (r = .07, non-significant). Navigation aids do what they claim for the immediate task.
• Providing the map rather than the instructions preserves spatial learning: paper-map navigation outperforms smartphone-map navigation on route retracing, and higher automation is associated with poorer route knowledge afterwards.
• Spatial cognition varies systematically with language, schooling, and environment, in ways that are specific and measurable rather than differences in underlying capacity.
• Older adults genuinely benefit from navigation aids, and the one mood measurement available does not show a cost.

Contested or weak.

• That navigation aids cause spatial decline (correlational in most studies, with the self-selection confound unresolved; some experiments null; 69% of pooled studies at moderate-to-high risk of bias).
• That spatial training raises general intelligence or general academic achievement (the architecture intervention moved studio grades; the engineering intervention moved nothing but the spatial score).
• That the human entorhinal grid-cell literature is settled (a 110-participant null with moderate Bayes evidence against, in a passive paradigm).
• That one language or culture produces categorically different spatial thinking (the effects are real, axis-specific, and task-specific).
• That Tolman's shortcut finding is clean (partial replication with a failed predicted contrast).

Not established at all, and worth naming: that spatial cognition is deteriorating in the population as a whole, that smartphones as a category cause cognitive decline, or that any of this describes the average person's daily experience. These are claims about populations that would require population-level longitudinal evidence, and they are not what the studies cited here provide.

9. The practical implication

The literature converges on a distinction that the design of modern navigation has quietly collapsed: the destination and the map are different outputs, and only one of them is what the user needed.

For anyone designing or choosing navigation technology, the evidence suggests a specific set of practices:

• Give the survey, not just the turns. Verbal and tonal guidance, delivered with no map at all, impairs spatial memory relative to navigating unassisted (Gardony et al., 2013). The cost is in the guidance, not in the screen.
• Keep the map in the hands of the user. Paper-map navigation preserves route knowledge that smartphone navigation does not (Sugimoto et al., 2021). The physical map is not a nostalgic interface; it is an environment in which the spatial encoding work is not outsourced.
• Reduce automation where learning is the goal. Higher automation was associated with too little spatial knowledge to reverse the route unaided (Brügger et al., 2019), and the cost falls hardest on children, who are building the maps in the first place.
• Treat the assist as a scaffold to be withdrawn. Landmark-based instructions improve spatial learning but reduce GPS dependence only partially (Zhang & Li, 2025). The goal is not to avoid the aid indefinitely but to leave the user with an internal model that outlives it.
• Do not universalise. A 75-year-old and a 9-year-old are in different regimes. Advice that helps one can harm the other. Voice navigation matched paper maps on completion time and left older adults more relaxed (Nagata et al., 2025) — the case for paper is about the map, not about comfort.

For learning design, the analogous conclusion is: teach the structure, not the performance. Mental rotation drills make people better at mental rotation, mostly. Spatial training that is concrete, developmentally appropriate, and adjacent to the target skill produces modest real gains in mathematics; the same training delivered abstractly, to the wrong age group, or expecting transfer to unrelated domains does not. A curriculum that assumes spatial training is a general STEM multiplier is relying on a far-transfer effect that the evidence puts at g = .20.

And the deeper point, which is the one worth keeping: the capacity to build a model of a place is not a special-purpose navigation skill. It is a general mechanism for encoding experience, organising relations, and reasoning about systems that have a structure — which is to say, for mathematics, science, and much of professional work. The tools that let a person arrive somewhere without building that model have quietly removed the practice by which a great many other things are learned.

Sources and References

• Hawes, Z. C. K., Gilligan-Lee, K. A., & Mix, K. S. (2022). "Effects of Spatial Training on Mathematics Performance: A Meta-Analysis." Developmental Psychology, 58(1), 112–137. 29 controlled pre-post studies; N = 3,765; k = 89. Overall Hedges' g = .28 (SE = .07); spatial thinking g = .49 (SE = .09); near transfer g = .653 vs. far transfer g = .203; concrete vs. non-concrete g = .416 vs. .052 (b = .36, p = .003); age b = .051, p = .006; I² = 72%; prediction interval −.37, .93; rank test r = .31, p = .031; Egger z = 1.40, p = .162; trim-and-fill estimated roughly two missing studies for spatial outcomes. Design and publication status did not significantly moderate the effect. https://doi.org/10.1037/dev0001281
• Lowrie, T., Logan, T., & Ramful, A. (2017). "Visuospatial training improves elementary students' mathematics performance." British Journal of Educational Psychology, 87(2), 170–186. Grade-six program, ten weeks; 186 students in eight classrooms randomly assigned (five intervention, three control). Spatial reasoning F(1,185) = 12.36, p < .001, d = .54; visualization d = .65; mental rotation d = .43; orientation p = .32; mathematics F(1,185) = 7.18, p = .02, d = .40; change-to-change r(120) = .21, p = .021; no gender differences. https://doi.org/10.1111/bjep.12142
• Atit, K., Power, J. R., Pigott, T., Lee, J., Geer, E. A., & Uttal, D. H. (2022). "Examining the Relations Between Spatial Skills and Mathematical Performance: A Meta-Analysis." Psychonomic Bulletin & Review, 29(3), 699–720. 45 analyzed articles from 858 screened records; pooled r = .36, robust SE = .035, τ² = .039; no significant gender or grade-level moderation. Correlational, not causal. https://doi.org/10.3758/s13423-021-02012-w
• Uttal, D. H., Meadow, N. G., Tipton, E., Hand, L. L., Alden, A. R., Warren, C., & Newcombe, N. S. (2013). "The Malleability of Spatial Skills: A Meta-Analysis of Training Studies." Psychological Bulletin, 139(2), 352–402. 217 training studies, 1984–2009; average g = .47 (SE = .04). Spans spatial skill broadly, not academic outcomes specifically. https://doi.org/10.1037/a0028446
• Sala, G., & Gobet, F. (2019). "Cognitive Training Does Not Enhance General Cognition." Trends in Cognitive Sciences, 23(1), 9–20. Critical review; argues near transfer to trained tasks is common while far transfer to untrained abilities is systematically overstated. https://doi.org/10.1016/j.tics.2018.10.004
• Porat, R., & Ceobanu, C. (2024). "The Role of Spatial Ability in Academic Success: The Impact of the Integrated Hybrid Training Program in Architecture and Engineering Higher Education." Education Sciences, 14(11), 1237. 154 first-year students (79 engineering, 75 architecture). Spatial ability improved in both faculties; architecture students' course grades rose (2.85 < B < 9.34, p < .01); engineering students' calculus and physics grades did not. Study 2 (architecture controls): r = .46–.67, FDR-adjusted p = .007–.024. https://doi.org/10.3390/educsci14111237
• Dahmani, L., & Bohbot, V. D. (2020). "Habitual Use of GPS Negatively Impacts Spatial Memory During Self-Guided Navigation." Scientific Reports, 10, 6310. 60 recruited, 50 analyzed (18 women, 32 men; mean age 27.6); 13 retested at three years. Lifetime GPS experience vs. spatial memory r = −.22, BCa 95% CI −.41, −.01; GPS reliance vs. spatial-memory strategy use r = −.25, 95% CI −.44, −.03; dependence vs. learning from errors r = −.33, 95% CI −.49, −.14; lifetime GPS use vs. subjective sense of direction r = .07, 95% CI −.16, .27 (non-significant). Correlational design. https://doi.org/10.1038/s41598-020-62877-0
• Miola, L., Muffato, V., Sella, E., Meneghetti, C., & Pazzaglia, F. (2024). "GPS Use and Navigation Ability: A Systematic Review and Meta-Analysis." Journal of Environmental Psychology, 99, 102417. 23 studies pooled. GPS use vs. environmental knowledge r = −.18, 95% CI −.28, −.08; vs. sense of direction r = −.25, 95% CI −.39, −.12; vs. wayfinding performance r = .07, 95% CI −.28, .41 (non-significant). 69% of included studies rated at moderate-to-high risk of bias. https://doi.org/10.1016/j.jenvp.2024.102417
• Topete, A., He, C., Protzko, J., Schooler, J., & Hegarty, M. (2024). "How Is GPS Used? Understanding Navigation System Use and Its Relation to Spatial Ability." Cognitive Research: Principles and Implications, 9, 32. Two preregistered studies, 249 participants; GPS dependence negatively correlated with objective navigation performance and sense of direction, positively with spatial anxiety. https://doi.org/10.1186/s41235-024-00545-x
• Yavuz, E., He, C., Gahnstrom, C. J., Goodroe, S., Coutrot, A., Hornberger, M., & Spiers, H. J. (2024). "Video Gaming, but Not Reliance on GPS, Is Associated with Spatial Navigation Performance." Journal of Environmental Psychology, 96, 102296. N = 822 (280 men, 542 women; mean age 26.3), Sea Hero Quest. No significant association between GPS reliance and navigation performance for either gender; weekly gaming hours significantly associated with navigation performance; gender no longer significant after accounting for gaming. https://doi.org/10.1016/j.jenvp.2024.102296
• Gardony, A. L., Brunyé, T. T., Mahoney, C. R., & Taylor, H. A. (2013). "How Navigational Aids Impair Spatial Memory: Evidence for Divided Attention." Spatial Cognition & Computation, 13(4), 319–350. Within-participants design, N = 36, comparing verbal, tonal, and no navigation aid. Both aided conditions impaired spatial memory relative to unassisted navigation; interpreted as divided attention. https://doi.org/10.1080/13875868.2013.792821
• Brügger, A., Richter, K.-F., & Fabrikant, S. I. (2019). "How Does Navigation System Behavior Influence Human Behavior?" Cognitive Research: Principles and Implications, 4, 23. N = 64 across four levels of automation on a real-world urban route; higher-automation systems appeared to leave participants without enough spatial knowledge to reverse the route unaided; gaze analysis descriptive. https://doi.org/10.1186/s41235-019-0156-5
• Zhang, Y., & Li, J. (2025). "Limiting the Negative Effect of GPS Dependence on Spatial Knowledge with Landmark-Based Instructions." Journal of Environmental Psychology, 108, 102813. Virtual-environment experiments; GPS dependence negatively correlated with route knowledge; GPS navigation impaired route knowledge acquisition; landmark-based instructions improved spatial knowledge acquisition but produced only partial reduction of GPS dependence. https://doi.org/10.1016/j.jenvp.2025.102813
• Sugimoto, M., Kusumi, T., Nagata, N., & Ishikawa, T. (2021). "Online Mobile Map Effect: How Smartphone Map Use Impairs Spatial Memory." Spatial Cognition & Computation, 22(1–2), 161–183. Route-walking study; smartphone-map navigation impaired route retracing and spatial learning relative to paper-map navigation. https://doi.org/10.1080/13875868.2021.1969401
• Muffato, V., Borella, E., Pazzaglia, F., & Meneghetti, C. (2022). "Orientation Experiences and Navigation Aid Use: A Self-Report Lifespan Study on the Role of Age and Visuospatial Factors." International Journal of Environmental Research and Public Health, 19(3), 1225. Age vs. GPS use r = −.53; age vs. map use r = −.20; age vs. verbal-directions use r = .16; older participants reported fewer episodes of getting lost; sense of direction and spatial anxiety predicted getting lost, navigation-aid use did not; GPS use predicted more frequent successful arrival at unfamiliar destinations (β = .12, p = .027). Self-report design. https://doi.org/10.3390/ijerph19031225
• Nagata, C., Ota, T., Sumida, Y., Isomura, Y., Shimizu, Y., & Yamamoto, T. (2025). "Comparing Voice-Only Navigation and Paper Maps in Middle-Aged and Older Adults: A Randomized Crossover Field Study of Wayfinding, Mood, Usability, and Text-Mined User Experiences." Scientific Reports, 16, 1103. N = 40; two 800-metre routes. No significant difference in goal completion (HR = 1.063, 95% CI 0.755, 1.498, p = .726); voice-only navigation produced greater relaxation (F(1,39) = 5.046, p = .030, η² = .115); no significant anxiety or fatigue differences; survey knowledge not measured. https://doi.org/10.1038/s41598-025-30590-5
• Wathiyage Don, A., Walker, C. M., Musculus, L., & Ruggeri, A. (2026). "Active Navigation Enhances Children's Spatial but Not Episodic Memory." Psychonomic Bulletin & Review, 33(6), 187. N = 105 children aged 4–10 (46 girls; M = 92.52 months, SD = 18.98); yoked design; tabletop map game. Path-length error: age F(1,102) = 8.49, p = .004, η² = .08; condition F(1,102) = 6.28, p = .014, η² = .06. Corner error: condition F(1,102) = 3.22, p = .075 (ns). Area error: age F(1,101) = 13.18, p < .001; condition F(1,101) = 4.29, p = .041, η² = .04. No episodic-memory benefit. https://doi.org/10.3758/s13423-025-02850-y
• Burles, F., & Iaria, G. (2020). "Behavioural and Cognitive Mechanisms of Developmental Topographical Disorientation." Scientific Reports, 10, 20932. Ten-year study: 1,211 DTD cases vs. 1,624 controls. DTD participants reported more first-degree relatives with childhood orientation difficulties (χ² = 266.160, p = 7.796 × 10⁻⁶⁰, OR = 2.914); both groups reported more female than male relatives with such difficulties. https://doi.org/10.1038/s41598-020-77759-8
• van Hoogmoed, A. H., Wegman, J., van den Brink, D., & Janzen, G. (2022). "Development of Landmark Use for Navigation in Children: Effects of Age, Sex, Working Memory and Landmark Type." Brain Sciences, 12(6), 776. 123 children aged 5–10 in an on-screen virtual environment; egocentric categorical performance above chance at five; allocentric navigation emerged between five and eight; boys outperformed girls on allocentric navigation; visuospatial working memory and left/right production predicted parts of the trajectory. https://doi.org/10.3390/brainsci12060776
• Hafting, T., Fyhn, M., Molden, S., Moser, M.-B., & Moser, E. I. (2005). "Microstructure of a Spatial Map in the Entorhinal Cortex." Nature, 436(7052), 801–806. Grid cells with repeating hexagonal firing fields. https://doi.org/10.1038/nature03721
• Doeller, C. F., Barry, C., & Burgess, N. (2010). "Evidence for Grid Cells in a Human Memory Network." Nature, 463(7281), 657–661. Grid-like fMRI signal in human entorhinal cortex, rotating coherently with movement direction. https://doi.org/10.1038/nature08704
• Jacobs, J., Kahana, M. J., Ekstrom, A. D., Mollison, M. V., & Fried, I. (2010). "A Sense of Direction in Human Entorhinal Cortex." PNAS, 107(14), 6487–6492. Direct recordings in epilepsy patients; direction-sensitive path cells. https://doi.org/10.1073/pnas.0911213107
• Kransberg, J., Sjøli Bråthen, A. C., Falch, E. S., Øverbye, K. E. Ø., Garrido, P. F., Fjell, A. M., et al. (2026). "Failure to Detect Entorhinal Grid-Like Signals in a Passive Navigation Human fMRI Study." Imaging Neuroscience, 4, IMAG.a.1196. N = 110 (72 women, 38 men; ages 18.2–78.6, mean 36.5). No six-fold grid-like signal in either entorhinal mask: right t(109) = 0.728, p = .234, BF₀₁ = 7.31; left t(109) = −0.443, p = .671, BF₀₁ = 8.60. Null in a passive paradigm, not evidence of absence. https://doi.org/10.1162/imag.a.1196
• Maguire, E. A., Gadian, D. G., Johnsrude, I. S., Good, C. D., Ashburner, J., Frackowiak, R. S. J., & Frith, C. D. (2000). "Navigation-Related Structural Change in the Hippocampi of Taxi Drivers." PNAS, 97(8), 4398–4403. 16 London taxi drivers and 50 controls; increased posterior hippocampal grey-matter volume bilaterally, with no equivalent anterior increase. Correlational, selected-population design. https://doi.org/10.1073/pnas.070039597
• Tolman, E. C. (1948). "Cognitive Maps in Rats and Men." Psychological Review, 55(4), 189–208. The latent-learning and maze-shortcut experiments that originated the cognitive-map concept. https://doi.org/10.1037/h0061626
• Wilson, S. P., & Wilson, P. N. (2018). "Failure to Demonstrate Short-Cutting in a Replication and Extension of Tolman et al.'s Spatial Learning Experiment with Humans." PLOS ONE, 13(12), e0208794. N = 60, 20 per condition. Kruskal-Wallis H = 7.89, p = .02; Light-Light vs. Ambient U = 106, p = .01; Light-Light vs. Light-Shifted U = 116.5, p = .023 (not significant at Bonferroni α = .017); Ambient vs. Light-Shifted p > .05. Partial replication with a failed predicted contrast. https://doi.org/10.1371/journal.pone.0208794
• Chase, W. G., & Simon, H. A. (1973). "Perception in Chess." Cognitive Psychology, 4(1), 55–81. Chunking as the basis of expert performance. https://doi.org/10.1016/0010-0285(73)90004-2
• Majid, A., Bowerman, M., Kita, S., Haun, D. B. M., & Levinson, S. C. (2004). "Can Language Restructure Cognition? The Case for Space." Trends in Cognitive Sciences, 8(3), 108–114. Cross-linguistic variation in spatial frames of reference and its cognitive consequences. https://doi.org/10.1016/j.tics.2004.01.003
• Pitt, B., Ferrigno, S., Cantlon, J. F., Casasanto, D., Gibson, E., & Piantadosi, S. T. (2021). "Spatial Concepts of Number, Size, and Time in an Indigenous Culture." Science Advances, 7(33), eabg4141. Tsimane' adults map number, size, and time onto space by magnitude but with no consistent directional bias. https://doi.org/10.1126/sciadv.abg4141
• Pitt, B., Carstensen, A., Boni, I., Piantadosi, S. T., & Gibson, E. (2022). "Different Reference Frames on Different Axes: Space and Language in Indigenous Amazonians." Science Advances, 8(47), eabp9814. The same participants preferred allocentric coding on the left–right axis and egocentric coding on the front–back axis, challenging the notion of a single predominant frame per language group. https://doi.org/10.1126/sciadv.abp9814
• O'Shaughnessy, D. M., Cruz Cordero, T., Mollica, F., Boni, I., Jara-Ettinger, J., & Gibson, E. (2023). "Diverse Mathematical Knowledge Among Indigenous Amazonians." PNAS, 120(35), e2215999120. Meta-analysis of 800+ Tsimane' children: full counting only with schooling; remote communities show specialised arithmetic around five. https://doi.org/10.1073/pnas.2215999120
• Gurven, M., Fuerstenberg, E., Trumble, B., Stieglitz, J., Beheim, B., Davis, H., & Kaplan, H. (2017). "Cognitive Performance Across the Life Course of Bolivian Forager-Farmers with Limited Schooling." Developmental Psychology, 53(1), 160–176. n = 919. Age trajectories similar to industrialised populations; schooling associated with greater cognitive ability at all ages, with no attenuating effect in late adulthood. https://doi.org/10.1037/dev0000175
• Davis, H. E., & Cashdan, E. (2019). "Spatial Cognition, Navigation, and Mobility Among Children in a Forager-Horticulturalist Population, the Tsimané of Bolivia." Cognitive Development, 52, 100800. 6–18-year-olds, GPS-tracked mobility. Schooling associated with better mental rotation but worse regional pointing accuracy, plausibly because schooling limits outdoor exploration. https://doi.org/10.1016/j.cogdev.2019.100800

Evidence scope: this article synthesizes the works listed above, with sample sizes, effect sizes, confidence intervals, and test statistics taken from the primary reports (published abstracts and full texts) and, for bibliographic details, from Crossref records. Where a value could not be verified in a primary record it has been omitted rather than approximated. Earlier drafts contained several citation errors — a nonexistent Boroditsky PNAS paper, a misattributed Doeller title, a 1946 date for Tolman, and a swapped sample in the Kransberg study — all corrected here against the primary records. Claims available only through secondary coverage are explicitly marked as contested or not established. No vendor, market, or marketing statistics are cited.

Written with Nyeker — AI assistant, for Faqih (dibotak). Bot disclaimer: this article was written with AI assistance.