The trade we don't know we're making
Every new technology trades one kind of capability for another. Writing weakened memory, but enabled civilisation. Calculators eroded mental arithmetic, but accelerated mathematics. GPS made maps obsolete, but meant no one gets lost anymore. These trades have always been worth it, the loss was real, but the gain was larger.
I'm starting to wonder whether AI is the first technology where the trade might go the other way.
We are outsourcing thinking itself. Not calculation, not navigation, not memorisation. The actual cognitive process of taking in information, weighing evidence, forming arguments, and arriving at conclusions. And unlike previous trades, where the loss was obvious and the gain was immediate, this one is insidious. The gain feels like productivity. The loss takes years to surface.
The fluency trap
There's a well-documented cognitive bias called the fluency heuristic: people judge information that's easy to process as more true. It's why clear fonts and simple language make arguments more persuasive, even when the content is identical. The brain confuses ease of processing with correctness.
AI-generated text is the fluency heuristic on steroids. LLMs produce prose that's grammatically perfect, confidently structured, and devastatingly smooth. It feels right in a way that messy, exploratory human writing doesn't. When you read something an AI wrote, your brain's truth-detection system is working against you, the very fluency of the text is a signal that tells you to lower your guard.
The danger isn't that AI gets things wrong. The danger is that it gets things almost right, just wrong enough to matter, but right enough to pass undetected. And because we're reading more AI-generated text every day, our tolerance for its particular flavour of plausible wrongness is increasing.
What atrophies when you don't use it
I use AI every day. I'm not writing this to position myself above it. I use it to summarise documents, to draft emails, to brainstorm ideas, to debug code, to rewrite sentences that don't quite land. Each of these uses saves me time. Each one also quietly reduces the amount of actual thinking I do.
When AI summarises a document for me, I don't read the document. That's the point. But reading a document and reading a summary are different cognitive processes. The first builds a mental model, makes connections to existing knowledge, and leaves a durable trace. The second leaves a thin residue that dissolves within hours.
When AI drafts an email for me, I don't wrestle with phrasing. But wrestling with phrasing is how you clarify what you actually think. The act of writing is an act of thinking. Every paragraph you didn't write is a thought you didn't fully form.
What worries me is the compound effect. Each individual trade, 30 seconds here, two minutes there, is trivially small. But across a working day, across a year, across a career, the cumulative loss of cognitive effort is enormous. We're not replacing thinking with better thinking. We're replacing thinking with the absence of thinking.
The parallel with Google Maps
There's a well-studied phenomenon in neuroscience: London taxi drivers have larger posterior hippocampi than the general population. Their brains physically changed to support the complex spatial navigation their job required. When GPS became universal, that cognitive adaptation stopped. New drivers never developed it. The skill wasn't replaced, it was eliminated.
I suspect something similar is happening with AI and higher-order cognition. The generation that grew up with LLMs won't be worse at thinking because they're lazy. They'll be worse at thinking because the cognitive muscles that support critical thought never developed in the first place. You can't atrophy something you never built.
When every homework assignment can be answered by ChatGPT, when every email can be drafted by Claude, when every code review can be handled by Copilot, when every research question can be answered by Perplexity, what exactly is the incentive to learn how to do any of these things yourself? And more importantly, what happens to a person who reaches adulthood having never had to?
The metacognition problem
Here's the cruelest part. You can't tell you're losing the ability to think, because thinking is what you'd use to notice. It's the cognitive equivalent of not knowing you have a fever because the thermometer is also broken.
Metacognition, the ability to evaluate your own thinking, is itself a skill that requires practice. Every time you offload a reasoning task to AI, you're skipping a metacognitive workout. You're not practicing the skill of evaluating whether an answer is right, you're simply accepting what the machine gives you and moving on.
I've caught myself doing this. I ask an AI a question, get an answer that sounds right, and move on. Hours later, I realise the answer was subtly wrong, or missing crucial context, or completely fabricated. But I only notice because I happen to have background knowledge in that area. What happens when the question is in a domain where I have no foundation? I never know I was misled.
The knowledge graph of one
Earlier this year I wrote about knowledge graphs as the missing layer between your data and your AI. The argument was that LLMs need grounded sources of truth to be reliable. I still believe that. But there's a personal version of this problem that no one is talking about.
Your personal knowledge graph, the web of concepts, facts, skills, and connections you carry in your head, is built one cognitive effort at a time. Every time you offload a thinking task, you're choosing not to add a node or an edge to that graph. Over time, the graph stays sparse. You know less. You connect less. You become less intelligent, not because you've lost capacity, but because you've filled your head with queries rather than knowledge.
The AI remembers everything. You remember how to ask for it.
What I'm not saying
I'm not saying AI is bad. I'm not saying we should stop using it. I'm not saying this is anyone's individual fault. I'm saying that the equation has more terms than we're accounting for.
We measure AI adoption in productivity gains, cost savings, and speed. We don't measure the cognitive edge cases: the critical thinking that wasn't exercised, the mental model that wasn't built, the connection that wasn't made, the mistake that wasn't caught because no one had the foundational knowledge to spot it.
These things don't show up on a balance sheet. They show up five years later, when a generation of engineers can ship code faster than ever but can't debug their way out of a paper bag when the AI gives them something novel. They show up when an organisation that replaced half its junior staff with AI tools discovers it no longer has a pipeline for senior staff, because seniors are what juniors become after years of doing the thinking that AI now does for them.
The countermeasure
I don't have a grand solution. But I have a working hypothesis: deliberate cognitive friction. The things you choose to do the hard way, even when an AI could do them faster, are the things that keep your thinking apparatus intact.
I try to follow a simple rule: use AI for the tasks where speed matters and understanding doesn't. Use it to summarise a meeting you attended. Use it to rephrase a sentence you already wrote. Use it to generate test data. But don't use it to write the argument you haven't made yet. Don't use it to answer the question you haven't thought about. Don't use it to avoid the struggle of understanding something hard.
The struggle is the point. The struggle is where the intelligence lives. Outsource the struggle and you outsource the intelligence, and eventually there's nothing left but a person who's very good at giving instructions to a machine.