GPS is the closest thing the cognitive-offloading literature has to a clean natural experiment. We adopted it in roughly the same window — late 2000s smartphones, by the mid-2010s near-universal use — and the cohort that grew up with it is now old enough to compare against the cohort that did not. The 2020 UCL study found exactly what the offloading framework would predict. That is part of why the GPS result keeps getting cited when people talk about AI: the timeline of GPS adoption is the timeline that AI adoption is roughly entering, just shifted forward fifteen years.
This post walks through what the GPS finding actually shows, what the mechanism is, and why the AI parallel follows.
What the UCL study actually shows
In 2020, Louisa Dahmani and Véronique Bohbot published "Habitual use of GPS negatively impacts spatial memory during self-guided navigation" in Scientific Reports (a Nature journal). Scientific American's secondary coverage is the most accessible non-academic summary.
The study had two parts. The first part was cross-sectional: participants were measured on lifetime GPS use and tested on self-guided navigation tasks — wayfinding without GPS, in environments they had to learn. People with greater lifetime GPS use performed measurably worse on the no-GPS tasks. The effect was specific to spatial cognition; it did not show up on general cognitive measures.
The second part was longitudinal: participants were followed and re-tested over time. The follow-up analysis suggested the relationship was likely causal in the GPS-causes-decline direction, not the reverse. Heavier GPS use predicted subsequent decline in hippocampal-dependent spatial memory, which is the right pattern for a cause-and-effect relationship rather than a confound.
That is the entire finding. It is narrow, it is well-measured, and it is now one of the most cited results in the offloading literature.
The mechanism
Spatial memory is one of the most mechanically well-understood cognitive functions in the brain. Self-guided navigation engages the hippocampus and prefrontal cortex actively — building and updating a mental map, choosing among alternative routes, predicting paths, integrating landmarks. This is not a theoretical claim. It is a fMRI-and-lesion-study finding accumulated over decades, the same line of work that produced the 2014 Nobel Prize for the discovery of place cells and grid cells.
Following GPS turn-by-turn directions through the same route does not engage those systems in the same way. The cognitive load shifts from spatial reasoning to instruction-following. The instruction-following is easy and the spatial reasoning is unexercised. Over years of cumulative use, the unexercised system is measurably less capable when called on directly. That is the result the UCL study quantified.
A useful way to think about it: the brain treats GPS as part of an extended navigation system, the way the 2011 Sparrow et al. work showed it treats Google as part of an extended memory system. In both cases, the offloading is real, the trade is rational at the level of any single trip, and the cumulative cost shows up across years.
What the AI parallel actually is
The reason the GPS finding keeps coming up in AI conversations is that the mechanism generalises cleanly. Risko and Gilbert's 2016 review of cognitive offloading in Trends in Cognitive Sciences framed the broader pattern: offloading shifts where cognitive effort lands, and skills that are not regularly exercised receive fewer reps. The GPS finding is one extreme — a very direct, very specific case. The AI extension is less extreme in any single instance but applies to a much wider cognitive surface.
Drafting with an AI assistant is to writing what GPS is to navigation. The tool produces the result, the underlying skill gets fewer reps. The 2025 MIT Media Lab cognitive-debt preprint (Kosmyna et al. on arXiv, with the Stanković 2026 critique also still in preprint) is one early empirical hook on the writing side. Anthropic's 2026 study on coding-skill formation is another. The literature is thin compared to the decades-long GPS-and-navigation line, but the framework is the same.
There is one important difference worth naming. GPS offloads a single, well-bounded cognitive function — spatial memory and route planning. AI assistants offload across a wider range: writing, summarising, analysis, planning, decision support. The cumulative effect could plausibly be larger or smaller than the GPS case. We do not yet know. The honest framing is that the GPS result tells us the offloading mechanism is real and measurable; it does not tell us the size of the AI effect at year fifteen of widespread adoption, because we are only at year three or four of the adoption curve.
What the GPS result does not show
The study does not show that GPS use causes general cognitive decline. The effect is specific to spatial memory and spatial cognition. GPS users in the study were not worse at anything other than the navigation tasks. By the same logic, AI use is unlikely to cause general decline; the relevant question is which specific skills are being practised and which are being delegated.
The study also does not recommend that anyone stop using GPS. GPS is genuinely useful, especially in unfamiliar areas. The cost is real and contained. The practical response is to use it deliberately rather than by default.
What to do about it — and what it means for AI
The practical guidance from the GPS literature is the same guidance the broader offloading literature converges on. Use the tool deliberately, not by default. Build in a small amount of regular unaided practice for the underlying skill. The point is not abstinence; it is preserving the existence of the skill.
For GPS: navigate familiar territory without it sometimes. Look at the map before a new trip and build a mental picture, then use turn-by-turn. Pay attention to landmarks even when GPS is on.
For AI: do at least one cognitive act per day without AI mediation. Write a paragraph yourself. Estimate a number before asking. Read an article fully before asking for a summary. Decide deliberately when AI is the right tool and when the skill itself is the thing you are practising.
Senwitt is the place we built for the AI version of that habit. The daily Set is a small, mixed, unaided cognitive workout — exactly the structural parallel to "walk to the corner shop without checking the map." The bet is that as the AI offloading dose increases over the next decade, the people who kept a daily unaided practice will be the ones whose underlying thinking skills are still in regular shape.
The GPS literature is the cleanest reason to believe that bet is worth making.
One last note on what makes the GPS result so unusually informative. Most of the time, when a tool changes how a population uses a cognitive skill, the question of whether the skill weakened is hard to answer cleanly — the variables are confounded, the affected subpopulation is hard to isolate, and the timescale is too short to see the trend. The GPS case got around all three of those problems almost by accident. The adoption was sudden enough to mark a cohort divide. The skill being changed was specific and measurable. The follow-up window was long enough to separate cause from correlation. We rarely get all three at once. When we do, the finding deserves more weight than it would in an ordinary cognitive-psychology context — and the GPS finding is doing exactly that work as a stand-in reference point for what the AI offloading effect might look like in a decade.
