AI writing tools accumulate cognitive debt and erode thinking capacity

Executive overview

Using AI to write for you is not just outsourcing a task — it is outsourcing the cognitive strain that builds the capacity to think. An MIT study tracked 54 participants over four months and found that heavy AI-assisted writers showed 47% lower brain activity, could not recall what they had written, and produced work rated as soulless and generic. The core danger is cognitive debt: short-term efficiency gains bought at the cost of long-term mental fitness.

Writing is cognitive weightlifting — use a forklift in the gym and you get weaker.

The MIT cognitive debt study

• 54 participants tracked over four months; split into groups using heavy AI, no technology, or Google search
• Heavy AI users showed 47% reduction in brain activity (EEG)
• 83% of AI-assisted writers could recall nothing from what they had just written, vs ~10% in the no-tech group
• Neutral evaluators described AI-assisted writing as "soulless, empty, lacking individuality, typical"
• Most striking finding: when AI was removed, heavy users wrote worse than people who had never used AI at all
• This suggests not just stagnation but active degradation of writing and thinking ability

Why writing builds the brain

• Writing is culturally invented, not evolutionary — it hijacks multiple brain regions simultaneously
• Visual cortex, spatial reasoning, linguistic and auditory centres all fire in coordination
• The same spatial reasoning used to navigate physical space is recruited to hold an argument's structure in mind
• These regions wiring together is the cognitive equivalent of muscle growth
• Deep reading research (Marion Wolf) shows the same principle: grappling with difficult text strengthens neural connections that enable more sophisticated reasoning and empathy
• AI-assisted writing takes those regions offline — the connections are never formed

Cognitive fitness vs cognitive obesity

• Physical analogy: lifting weights in a gym vs having a forklift do it for you — same time spent, zero benefit
• Vacuum cleaner analogy: convenience reduced physical activity; health consequences followed decades later
• The same pattern is emerging for cognitive health: outsource thinking, reduce mental fitness
• Consuming AI-generated content compounds the problem — "ultra-processed information" alongside ultra-processed food in a cognitively sedentary environment
• Cognitive obesity may be where this trajectory leads; we are in the early stages, equivalent to the 1982 inflection point for physical obesity
• Short-form AI content mirrors TikTok scroll behaviour: no strain, no deep-reading processes developed, no benefit

Type two thinking

• Type two fun: activities unpleasant in the moment (hard mountaineering, tough workouts) but deeply satisfying and beneficial in retrospect
• Type two thinking: the blank page is hard; composing something you are proud of is deeply satisfying
• The difficulty is not a bug — it is the mechanism by which the brain gets stronger
• Prompting an AI back and forth is a narrow, tool-functional cognitive task; it does not replicate the symphony of brain regions engaged in original writing
• Future possibility: cognitive gyms — dedicated time with no AI, intentional mental exercise to offset outsourcing elsewhere

Where to draw the line

• Students: do not use AI to write. School is the intellectual gym. The whole point is the strain
• Professional writers: your thinking is the product. Outsourcing any of it defeats the purpose
• Functional/administrative writing: moving information from one format to another, summarising meeting notes — acceptable use, like an occasional fast food meal; not a diet
• Copy editing on deadline: acceptable with caution; treat it as a signal to invest in writing practice in lower-stakes situations, not a permanent crutch
• As a training tool: use AI to evaluate writing quality, then attempt to improve independently rather than accepting its rewrites

Intentionality as the defence

• We were not deliberate with smartphones or social media; we are paying the price
• Being healthy in a bad food environment requires knowledge and rules; the same applies to the information environment
• A budget that never says no is not a budget; a digital life with no constraints is not a managed one
• Rules do not have to be rigid or consuming — they just need to create actual friction at actual moments
• The goal: learn from the smartphone adoption mistake and be intentional about AI from the start

AI creativity and generalization

• Generalization beyond the distribution is the holy grail of AI research — the ability to produce genuinely new understanding, not pattern-matched output
• Current large language models embed pattern recognition and informal rules from training data; novel instances of familiar patterns can be handled, but true generalization is limited
• Apple research found that reasoning in cutting-edge LLMs does not generalise beyond rules inferred during training
• Real creativity will likely require hybrid systems: language models combined with symbolic reasoning and ontological structure
• Current models are not scaling into the generalization required for creativity as most people understand it

Productivity and organisation fundamentals

• Minimum effective dose for disorganised people: a master list — one place where everything is written down and tended daily
• Full capture (David Allen's term) vs haphazard tracking across email, chat, text, and memory; the former removes the brain's constant background anxious refresh cycle
• Once a master list exists, smarter time management organically follows — you start to see what matters, what to punt, what needs a deadline
• The inbox-to-workingmemory.txt workflow: skim all messages, type fast one-line summaries into a plain text file, archive originals, then sort summaries by type and handle in batches
• Batching by type (all scheduling together, all replies together) eliminates context-switching cost and enables better decisions — e.g. spotting that a week is overloaded before committing
• Only items requiring non-trivial future action become Trello cards; most are handled immediately in the batch pass

Career capital and internship choices

• Two questions for any career move: (1) can I build rare and valuable skills here? (2) will I be able to cash in that capital for autonomy later?
• Both must be yes; a job that builds capital but offers no flexibility on how to invest it (classic big-law partnership track) traps you
• For internships: gather information, keep the stakes low, choose whatever opens more options — the real decision is which full-time role to pursue
• Career capital traded for lifestyle control is the engine behind sustainable, autonomous work arrangements

Finding time for personal projects

• Walt Disney built an elaborate scale-model steam railway — a multi-year project — while running a major studio
• When a project is genuinely important, schedule gives; you find pockets, come home early, use evenings differently
• Four-day work week data shows most schedules have more slack than people realise — time fills to available space
• Screen time is the hidden consumer: removing default streaming and social media reveals large blocks of unscheduled time
• Boredom from removing screens drives you toward projects; it is the signal, not the problem
• Walter Isaacson wrote multiple historical bestsellers as CEO of Time Warner by simply not watching television in the evenings

AI hype and alarmist headlines

• Headlines claiming AI can now do PhD-level work or replace entry-level workers are distorted extrapolations from narrow benchmarks
• The specific case: a model fine-tuned on a curated dataset of math competition problems, written by PhD mathematicians at $100/hour, performed well on that specific type of problem — then failed on similar but different competitions
• Post-pandemic tech layoffs (600,000+ since 2022) are cyclical contractions, not AI automation; the same pattern followed the dot-com bust
• Tech CEOs have a direct financial incentive to amplify disruption narratives: investor capital requires belief in massive future upside to overlook present losses (OpenAI losing $3–4B/year)
• Rule of thumb: if the headline makes a world-shaking claim, is quoting a tech CEO, and does not match anything you observe in your own industry — ignore it
• Serious AI reporting exists; the noise drowns it out, but the noise is not hard to identify once you understand the incentive structure