Thesis · No. 03
Cognition

Are We Getting
Stupider?

Cognitive atrophy, the GPS effect and what happens to human intelligence when thinking becomes a service you can subscribe to.

Malta Insider
September 2026
11 min read

In 2011, a team at Columbia University ran a simple experiment. They gave participants a set of trivia questions and, before answering, told some of them that the answers would be saved to a computer — and others that they would be erased. The group who thought the answers would be saved remembered significantly less. They hadn't bothered to encode the information into memory because they knew it was stored elsewhere.

The researchers called this the Google Effect. When we know an external system holds information, we stop bothering to hold it ourselves. The brain, ever efficient, deprioritises encoding what it believes it can retrieve on demand.

That was 2011. Google only stored information. It did not think.

What happens when we outsource not just the storage of information, but the process of reasoning about it?

The brain deprioritises encoding what it can retrieve on demand. Google stored information. AI thinks. The difference is not incremental — it is categorical.

The GPS Brain

The most rigorously studied example of technology-induced cognitive atrophy is not smartphones or social media. It is GPS navigation — and what it did to the hippocampus.

London taxi drivers, famously, must memorise the entire street map of London before receiving their licence — a process called "the Knowledge" that takes, on average, four years. Studies of these drivers using neuroimaging found measurably larger posterior hippocampi compared to non-taxi-drivers. The hippocampus is the brain region responsible for spatial memory and navigation. Use it more — it grows. The effect was proportional to years of practice and absent in failed trainees, ruling out selection bias. The brain is plastic. It responds to demand.

GPS inverts this entirely. Regular GPS users show longitudinal spatial memory decline — correlated at r = −0.68 with device usage in one study, a remarkably strong relationship. They perform worse on navigation tasks requiring them to remember routes. They show declines in hippocampal-dependent spatial memory over time. Eye-tracking studies find that increased attention to GPS interfaces corresponds to less accurate spatial knowledge — even when the device is doing the work correctly.

The Enrichment Paradox — arXiv 2026
London taxi drivers develop measurably larger posterior hippocampi than bus drivers who follow fixed routes. GPS users show longitudinal spatial memory decline correlated at r = −0.68 with device usage. Students given unrestricted GPT-4 access score 17% lower on unassisted exams than controls.
Critical capability thresholds and irreversible dependency in human-AI symbiosis

This is not a metaphor. It is a measurable neurological change, documented in peer-reviewed literature across multiple decades of GPS adoption. The brain atrophied a specific capability because that capability was no longer required. Use it or lose it is not motivational language. It is neuroscience.

Now consider what AI is doing — and how much more of the cognitive stack it touches than GPS navigation.

AI Apathy — The New Cognitive Risk

Researchers have begun naming the AI-specific version of this phenomenon. AICICA — AI-chatbot-induced cognitive atrophy — describes the potential deterioration of essential cognitive abilities resulting from excessive dependence on AI tools. The terminology is new. The underlying mechanism is not: it is the same use-dependent cortical plasticity that drove the GPS effect, now applied to reasoning, writing, analysis, and decision-making.

A 2025 study of 666 participants — one of the largest to date — found a strong statistical link between frequent AI tool usage and the decline in critical thinking scores as measured by standardised tests, with cognitive offloading identified as the mediating factor. Students who outsourced writing, problem-solving, and analytical tasks to AI demonstrated measurable declines in their ability to perform those tasks independently when AI was subsequently unavailable.

666
Participants in 2025 study linking AI use to critical thinking decline
17%
Score decline for students given unrestricted GPT-4 access vs. controls on unassisted exams
79%
AI-generated code reviews accepted without modification in 2026 Amazon engineering audit

A systematic review of 67 empirical studies published between 2022 and 2025 found that in unstructured contexts — ordinary usage, not pedagogically guided — ChatGPT use triggers cognitive offloading and produces measurable declines in critical thinking. The review notes an "asymmetrical pattern": creativity and idea generation can be supported by AI, but analytical rigour, logical argumentation, and independent reasoning tend to decline. The mind reaches for the tool rather than attempting the work.

ScienceDirect · March 2026
The cognitive impact of ChatGPT in higher education: A systematic review of critical and creative thinking outcomes
67 empirical studies, 2022–2025. In unstructured contexts, asymmetrical patterns show creativity maintained but critical thinking in decline, both triggering cognitive offloading.

The Pilot Problem

Aviation discovered this first. Modern commercial aircraft can fly themselves — from takeoff to landing — with minimal human input. This capability has made air travel extraordinarily safe. It has also produced a generation of pilots whose manual flying skills have atrophied significantly compared to those trained before full autopilot dependency.

When automated systems fail — as they occasionally do — the pilots whose skills have atrophied are suddenly required to perform tasks they have not practised in years, under extreme stress, with lives in the balance. Aviation regulators now mandate minimum hours of manual flying specifically to counteract this effect. The automation is not removed. But the skill must be deliberately maintained against the pull of dependency.

The parallel to AI-assisted knowledge work is direct. A 2026 study of software engineering practices, using the Amazon infrastructure outages as a case study, found that 79% of AI-generated code reviews were accepted without modification by engineers who had grown dependent on AI-generated suggestions. When the systems failed, the engineers lacked the deep understanding needed to diagnose and fix the problems — because the AI had been handling that cognitive work, and the underlying skill had quietly eroded.

Cognitive Atrophy in AI-Dependent Software Engineering — arXiv 2026
Using the 2026 Amazon outages as a case study: 79% of AI-generated reviews accepted without modification. When software ecosystems are maintained by a workforce that has forgotten how to build them, the risk of systemic collapse becomes an architectural certainty.
"Rubber Stamp" culture and the Treadwell Mandate for human oversight
When the automation failed, the engineers lacked the understanding to fix it — because the AI had been handling that thinking, and the underlying skill had quietly eroded.

The Google Effect, Supercharged

The original Google Effect altered how we encode information. AI alters something deeper: how we process it.

Google told you where to find the answer. You still had to read the page, evaluate the source, synthesise across multiple results, and form a judgement. The cognitive work of processing information remained yours. You just outsourced the retrieval.

AI performs the processing. It reads the pages. It synthesises the sources. It forms the judgement and presents it as a conclusion. The user's role collapses from thinker to prompt author and output accepter. The cognitive workout that builds analytical capacity is skipped entirely.

The Cognitive Divergence — arXiv 2026
AI goes beyond storing information — it performs the cognitive transformation of that information. The AI delegation effect is likely stronger than the Google Effect: chronic reduction of sustained attentional effort is a plausible pathway to durable capacity reduction, consistent with use-dependent cortical plasticity.
AI context windows, human attention decline, and the delegation feedback loop

The 2026 research on this is unambiguous in its direction, even where individual study sizes remain modest. The mechanism is confirmed — cognitive offloading reduces the effort invested in a cognitive task, and reduced effort means reduced practice, and reduced practice means reduced capability. The question is not whether this happens. It is how fast, and how reversible.

What Atrophies — and What Doesn't

The research does not support a simple "AI makes you dumb" narrative. The picture is more specific — and in some ways more concerning.

What AI appears to support
Idea generation — brainstorming, divergent thinking, exploring options

Surface creativity — novel combinations, aesthetic experimentation

Information retrieval speed — finding facts, summarising material

First-draft production — getting something on the page quickly
What appears to atrophy
Analytical rigour — evaluating argument quality, spotting logical flaws

Independent reasoning — reaching conclusions without external scaffolding

Deep memory encoding — retaining and connecting knowledge over time

Epistemic scepticism — the habit of questioning claims and demanding evidence

The pattern is consistent: AI handles the effortful parts of cognition — the parts that are uncomfortable, slow, and resist automation. Those effortful parts are also, not coincidentally, the parts that build intellectual capacity when practised. The comfortable, fast, generative parts remain with the human. The hard parts — the parts that make you better at thinking — are outsourced.

The Divergence Hypothesis

The most troubling scenario is not uniform cognitive decline. It is divergence.

A small group of people will use AI as a genuine cognitive amplifier — as a tool that frees mental bandwidth for higher-order thinking, that challenges their assumptions, that accelerates genuine learning. These people will become more capable, faster, with access to more breadth than any previous generation could match.

A much larger group will use AI as a cognitive outsourcing service — delegating thinking to it, accepting outputs without scrutiny, gradually losing the habit and the ability to reason independently. For this group, AI will deliver the appearance of competence while quietly eroding its foundations.

The danger is not that AI makes everyone stupider. It is that it makes the divergence in human cognitive capacity wider than it has ever been — and does so invisibly, through a mechanism that feels, at every individual moment, like efficiency.

"AI can function as a cognitive amplifier when used intentionally as an assistive partner rather than a substitute. By automating routine tasks, AI can free cognitive resources for higher-order thinking. The risk is when it replaces rather than supports active cognitive engagement."
ResearchGate consensus — Will AI dull or sharpen human intelligence? February 2026

The Irreversibility Question

The hippocampus of a taxi driver who stops driving does not remain enlarged. Studies of retired cabbies show the spatial memory advantage begins to erode within years of stopping. Use-dependent plasticity is not a one-way door. Capabilities can be recovered through deliberate practice.

But recovery requires awareness of the loss, and motivation to rebuild — two things that are hard to sustain when the AI is always there, always faster, always producing something that looks like a good answer. The incentive structure pulls toward deeper dependency, not toward the uncomfortable work of rebuilding what has been lost.

Aviation addressed this through regulatory mandate: you must log minimum manual flying hours regardless of how good the autopilot is. Education systems are beginning to address it through AI-restricted assessments. But for the billions of knowledge workers who use AI tools without any such structure — the erosion continues, unmonitored and unmeasured.

The question "are we getting stupider?" has a precise answer: in specific, measurable cognitive domains — spatial memory, critical thinking, independent analytical reasoning, epistemic scepticism — the evidence says yes, for people who delegate those functions to AI without deliberate countermeasures.

Whether that matters depends entirely on whether those domains remain important in a world where AI can perform them. That is a question the next thirty years will answer. But the atrophy begins now, in a decision made every time we reach for the tool instead of attempting the thought.

The atrophy begins in a decision made every time we reach for the tool instead of attempting the thought.

Next in Thesis · No. 04
The Luxury Trap
The subscription economy, manufactured necessity and how AI became the final link in a chain that turned everything — including thought itself — into a recurring charge.
← No. 02: Who Is the Slave, Who Is the Master? Back to Thesis