Your brain is a muscle. And AI is offering to carry all the weight.

Table Of Contents

• Introduction
• The neuroscience: synaptic pruning and negative neuroplasticity
• Case study 1: software development in the wake of AI coding agents
• Case study 2: writing skills in the wake of generative AI
• The Google effect: a precedent for what’s coming
• The broader pattern: examples across domains
• How to mitigate cognitive decline: a practical framework
• Conclusion
• References

Introduction

In 2000, Eleanor Maguire and her colleagues at University College London scanned the brains of licensed London taxi drivers — people who had spent years memorizing 25,000 streets and thousands of landmarks through a grueling training process known as “The Knowledge.” The posterior hippocampi of these drivers were measurably larger than those of the general population, and the volume correlated with years of experience (Maguire et al., 2000, PNAS).

Fast-forward to the 2020s. A longitudinal study in Scientific Reports found that habitual GPS users showed a steeper decline in hippocampal-dependent spatial memory over time than people who still navigated by landmarks and mental maps (Dahmani & Bhout, 2020). A 2024 population-based cohort study analyzing 8.9 million U.S. death records across 443 occupations found that taxi and ambulance drivers had the lowest Alzheimer’s-related mortality rates of any profession studied. Bus drivers and airline pilots, who follow fixed or automated routes, showed no such protection.

The brain regions you don’t use will atrophy. We are now entering an era where AI is volunteering to do our thinking, our coding, our writing, our reasoning. What happens to your brain when you let it?

The neuroscience: synaptic pruning and negative neuroplasticity

The principle of neuroplasticity cuts in both directions. The same mechanism that grows a taxi driver’s hippocampus also allows unused neural circuits to weaken and disappear.

When you repeatedly practice a skill — debugging code, structuring an argument, navigating a city — the synaptic connections underlying that skill strengthen. The myelin sheath around those axons thickens, making signal transmission faster and more reliable. The reverse is equally true: synapses that aren’t activated undergo long-term depression (LTD) and are eventually pruned. The brain consumes roughly 20% of the body’s energy despite being 2% of its mass and aggressively recycles circuits it doesn’t need.

Mahncke et al. (2006) described this as negative neuroplasticity — a self-reinforcing spiral where cognitive disuse leads to neural decline, which leads to further disuse. The person avoids demanding tasks because they’ve become harder, which makes the brain weaker, which makes the tasks harder still. This framework was developed for age-related decline and has since been extended to explain neurological degradation in chronic traumatic brain injury (Progress in Brain Research).

A 2024 study in Science Advances analyzed age-skill profiles across thousands of workers and found that cognitive skills begin declining as early as 30 — but the rate of decline varied substantially depending on whether people kept using those skills at work. Workers exposed to complex, challenging cognitive tasks showed far flatter decline curves than those whose jobs became routine. Skill usage was one of the strongest predictors of maintained cognitive function, independent of education level.

Your brain doesn’t care that you used to be a great programmer or a great writer. It cares about what you’re doing right now. Neural pathways that go unused aren’t preserved — they’re dismantled.

Case study 1: software development in the wake of AI coding agents

AI coding assistants have transformed software development fast. AI now generates an estimated 41% of all code — 256 billion lines in 2024 alone. By early 2025, 25% of Y Combinator’s Winter batch had codebases that were 95% AI-generated. Over 95% of developers report using AI-generated code in production.

In February 2025, Andrej Karpathy coined the term “vibe coding” — an approach where the developer describes what they want in natural language, accepts what the AI produces, and moves on without understanding it. Karpathy acknowledged this is “not too bad for throwaway weekend projects,” which implicitly concedes something about production systems.

The data is starting to come in.

One of the first randomized controlled trials of AI coding tools, by the Model Evaluation and Threat Research group (METR, 2025), studied 16 experienced open-source developers averaging over 22,000 GitHub stars per repository. Developers using AI assistants took 19% longer to complete tasks while believing they had worked 20% faster. The gap between perceived and actual performance is its own red flag: the developers’ ability to evaluate their own work had been compromised.

Stanford researchers found that developers using AI assistance wrote less secure code in four out of five tasks. When samples were audited, 73% of AI-generated code contained vulnerabilities. The developers using AI also expressed higher confidence that their code was secure — the Dunning-Kruger effect running on autopilot.

A 2024 GitClear analysis of millions of lines of code found that code churn — the percentage of code discarded within two weeks of being written — was projected to double in 2024. AI-generated code is being written faster and revised far more, creating what researchers call “AI-induced technical debt.”

A 2025 study from Microsoft Research and Carnegie Mellon found that the more people relied on AI tools, the less critical thinking they engaged in. High confidence in the AI caused users to take a “mental backseat,” especially on tasks perceived as easy. Over time, this led to diminished independent problem-solving. The study also found that AI-assisted workers produced a less diverse set of solutions to the same problem — a homogenization the researchers characterized as a form of critical thinking deterioration.

When a developer lets AI write their code, specific cognitive skills go dormant. Algorithmic reasoning — decomposing problems, spotting edge cases, choosing data structures — requires sustained prefrontal cortex activation; when AI handles this, those circuits weaken. Debugging is similar: it depends on holding a mental model of a system’s state and reasoning backward from symptoms to causes. If you never built that model, you can’t debug. Programming language fluency decays without practice, exactly as human language fluency does — developers are already reporting they struggle to write functions from scratch in interviews or when AI tools go down. And experienced engineers develop an intuition for how systems should be structured, a feel for code smells and scalability traps, built through thousands of hours of writing and reviewing code. That intuition can’t be acquired through prompts.

Developers are already describing it this way: panicking during AI service outages because they’ve lost the confidence to write code without assistance. One engineer who described their navigation skills as “atrophied” after years of GPS use said the same thing was happening to their programming skills.

Case study 2: writing skills in the wake of generative AI

Since ChatGPT launched in late 2022, surveys show up to 88% of university students have used generative AI for writing assignments. A Cornell University study found a 60% drop in writing center usage between fall 2022 and spring 2024. A follow-up survey of 67 writing center administrators found nearly a third reporting declining visits, with 12 centers experiencing drops of 31% or more. The Technical University of Munich reported similar numbers.

Students aren’t going to writing centers because they’re going to ChatGPT instead. The problem is that the writing center visit — struggling through feedback, revising a draft, arguing with a tutor about thesis structure — is exactly the kind of effortful cognitive engagement that builds writing circuits. Pasting a prompt and copying the output is not.

A 2024 study by Abbas et al. found that university students who used ChatGPT frequently had significantly lower GPAs. The mechanism was twofold: AI use correlated with procrastination and with reduced deep engagement with material. Wecks et al. (2024) found that students who used generative AI for test preparation scored an average of seven points lower on exams than those who prepared without it. The AI-prepared students had encoded the material less deeply because the AI had done the cognitive work.

MIT researchers equipped 54 students with EEG headsets and monitored brain activity over four months of essay writing. Students relying heavily on ChatGPT showed reduced prefrontal cortex activation — the region associated with planning, reasoning, and original thought. When this group was asked to write without AI in a fourth session, the negative effects persisted. Their brains had adapted to a lower-effort mode of operation and couldn’t easily shift back.

Multiple studies (Marzuki et al., 2023; Kim et al., 2025) conclude that while AI can help with surface-level mechanics, it’s ineffective at developing higher-order skills like argument construction, logical coherence, and rhetorical strategy. Students who rely on AI for these never build the neural circuitry for them.

What atrophies is specific. The uncomfortable process of staring at a blank page and iterating through bad drafts is how the brain builds capacity for original thought — cognitively expensive, which is exactly why it’s neurologically valuable. Argument structuring requires sustained working memory and executive function; when AI provides the structure, the user’s prefrontal cortex doesn’t get recruited for it. Revision is a metacognitive skill, the ability to evaluate your own output and improve it, and outsourcing first drafts to AI usually means outsourcing revision too, since you never developed a sense of what the text was trying to do. A writer’s voice is built through thousands of decisions about word choice, rhythm, and emphasis made over years of practice. AI outputs converge on a generic “helpful assistant” register. Students who spend their formative writing years using AI may never develop a distinctive voice.

The Google effect: a precedent for what’s coming

We’ve seen a version of this before. In 2011, Betsy Sparrow and colleagues at Columbia published a study in Science showing that when people expect future access to information online, they show lower recall for the information itself but enhanced recall for where to find it. The internet, they argued, had become a form of transactive memory — an external store the brain treats as an extension of itself (Sparrow, Liu, & Wegner, 2011).

This was initially framed as a neutral reallocation of cognitive resources. A decade of follow-up research tells a less comfortable story. People who habitually offload to search engines show reduced confidence in their own memories, reduced motivation to encode new information, and a growing drive to avoid internal memory effort. The brain learns that remembering is unnecessary and stops trying.

The same logic applies to skills. AI coding agents are transactive memory for programming ability. Generative AI is transactive memory for writing, reasoning, and argumentation. The Google effect replaced “I know this” with “I know where to find this.” AI is replacing “I can do this” with “I know how to prompt a machine to do this.” The cognitive cost is likely to be far more severe.

The broader pattern: examples across domains

The same dynamic shows up across domains.

Students who relied on calculators for routine arithmetic showed weaker number sense and estimation ability (Reys and Yang, 1998). The neural circuits for numerical reasoning, centered in the intraparietal sulcus, require active engagement to maintain.

Mueller and Oppenheimer (2014), in Psychological Science, found that students who took notes by hand performed better on conceptual questions than those who typed on laptops. Handwriting forces selective summarization because the hand can’t keep up with speech — the brain has to decide what’s important. Typing removes that bottleneck.

GPS use and spatial memory we’ve already covered. A McGill University series using fMRI found that GPS-dependent individuals had lower hippocampal activity and less grey matter than spatial navigators (Bohbot et al., 2010).

Persistent autocorrect use has been linked to declining spelling ability. When the brain learns that errors will be automatically corrected, the circuits for orthographic memory receive less activation.

In clinical medicine, physicians who become reliant on diagnostic decision-support systems show reduced accuracy when those systems are unavailable — the same pattern as AI coding assistants: the tool improves average performance while degrading independent capability.

How to mitigate cognitive decline: a practical framework

AI tools, used well, can genuinely amplify human capability. The goal is to use them in ways that keep your brain engaged rather than atrophied.

1. Attempt the task yourself first. Before prompting AI to generate code or text, spend 10–20 minutes trying it yourself. Even a partial, flawed attempt activates the relevant neural circuits. This is grounded in the testing effect, one of the most robust findings in cognitive psychology: actively attempting to produce information strengthens memory far more than passive review (Roediger & Karpicke, 2006). Before prompting AI to write a React component, sketch the structure yourself, write pseudocode, decide on state management. Then use AI to refine.
2. Treat AI output as a learning exercise, not a rubber stamp. When AI generates code or text, don’t just accept it. Read it line by line and ask yourself why it chose that approach, what the alternatives are, and where it might fail. When AI writes a database query, trace it mentally with sample data, check the join logic, look for N+1 problems. Rewrite it in a different style to test whether you actually understood it.
3. Keep AI-free practice sessions. Athletes do drills without equipment to build fundamental skills. Schedule time each week — even 30 minutes — to write code or prose from scratch. Coding challenges on LeetCode or Advent of Code, handwritten journals, essays you drafted yourself.
4. Use AI as a teacher, not a ghostwriter. Instead of asking AI to write for you, ask it to explain to you. Request explanations of algorithms, critiques of your code, feedback on your writing. “Explain three ways to handle boundary conditions in binary search and the tradeoffs of each” — then implement it yourself — builds more than “write me a binary search implementation” ever could.
5. Build mental models actively. The most valuable cognitive asset a developer or writer has is their internal representation of how a system, argument, or domain works. When learning a new codebase, resist the urge to have AI summarize it. Read the code yourself. Draw architecture diagrams by hand. Trace request flows. The struggle to understand is the point.
6. Stay cognitively demanding outside work. Research on cognitive reserve shows benefits accumulate from challenging activities across all domains, not just professional ones. Learning an instrument, studying a new language, playing strategic games, physical exercise — all contribute to maintaining neural health (Vemuri et al., Mayo Clinic, 2024). The more diverse the demands, the more resilient the brain.
7. Set guardrails at the team level. Require developers to write design documents before using AI to implement. Require writers to produce outlines and first drafts before using AI for editing. Make code review a genuine learning exercise. One useful policy: AI-generated code must be accompanied by a written explanation from the developer of what it does and why. If they can’t explain it, the code doesn’t ship.

Conclusion

The short-term productivity gains from AI are real. So is the long-term cost. The developer who ships code faster with AI today may be unable to debug a critical production issue without it tomorrow. The student who produces polished essays with ChatGPT may graduate unable to construct a coherent argument in a meeting.

Cognitive abilities aren’t permanent. They’re ongoing biological processes that require continuous engagement. Every time you let an AI think for you, you’re making a withdrawal from a cognitive bank account. Small withdrawals are fine — the brain is resilient. Chronic, sustained offloading of core skills leads to measurable atrophy, the same pattern documented in every other case of neural disuse.

Use AI tools deliberately — to amplify existing skill rather than to replace the need for skill at all. Your brain is paying attention to what you ask it to do. Make sure you’re asking it to do something.

References

1. Maguire, E. A., Gadian, D. G., Johnsrude, I. S., et al. (2000). Navigation-related structural change in the hippocampi of taxi drivers. Proceedings of the National Academy of Sciences, 97(8), 4398–4403.
2. Maguire, E. A., Woollett, K., & Spiers, H. J. (2006). London taxi drivers and bus drivers: A structural MRI and neuropsychological analysis. Hippocampus, 16, 1091–1101.
3. Dahmani, L., & Bohbot, V. D. (2020). Habitual use of GPS negatively impacts spatial memory during self-guided navigation. Scientific Reports, 10, 6310.
4. Mahncke, H. W., Bronstone, A., & Merzenich, M. M. (2006). Brain plasticity and functional losses in the aged: Scientific bases for a novel intervention. Progress in Brain Research, 157, 81–109.
5. Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776–778.
6. Mueller, P. A., & Oppenheimer, D. M. (2014). The pen is mightier than the keyboard: Advantages of longhand over laptop note taking. Psychological Science, 25(6), 1159–1168.
7. Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255.
8. Gerlich, M. (2025). AI tools and cognitive offloading: Effects on learning, critical thinking, and judgement. Various preprints and institutional reports.
9. Abbas, M., Jam, F. A., & Khan, T. I. (2024). Is it harmful or helpful? Examining the causes and consequences of generative AI usage among university students. International Journal of Educational Technology in Higher Education, 21(1), 10.
10. Wecks, J., et al. (2024). Generative AI in test preparation and its effects on student performance. Working paper.
11. Christoforou, E., et al. (2024). Age and cognitive skills: Use it or lose it. Science Advances, 10.
12. METR (2025). Measuring the impact of AI coding tools on developer productivity: A randomized controlled trial.
13. Microsoft Research & Carnegie Mellon University (2025). The impact of AI tool reliance on critical thinking in knowledge work.
14. GitClear (2024). Code quality analysis: The impact of AI coding assistants on code churn and technical debt.