“The question is not whether AI will change how we think. It already has. The question is whether we will notice before the change becomes permanent — and whether we would even care if we did.”
Neal Lloyd · Inside The MachineIn February 2026, economists Daron Acemoglu, Dingwen Kong, and Asuman Ozdaglar published a working paper through the National Bureau of Economic Research with a quietly terrifying title: AI, Human Cognition and Knowledge Collapse. The paper’s argument, stripped of its formal economic modelling, goes something like this: when AI systems deliver context-specific answers that substitute for human cognitive effort, people gradually stop generating the kind of learning that produces shared, community-level knowledge. The individual gets a better answer faster. Society’s collective stock of understanding slowly erodes. The machine gets smarter. The ecosystem it feeds from gets thinner.
This is not a science fiction scenario. This is a peer-reviewed economic model with empirical grounding, published by three of the most cited economists working today, in 2026. And it is one of several converging lines of research that are, taken together, raising a question that the AI industry has been conspicuously reluctant to ask: What exactly is this doing to the human mind?
You Can No Longer Tell Which Thoughts Are Yours
A paper published at the 2026 CHI Conference on Human Factors in Computing Systems — the field’s most prestigious venue — examined something called the AI Memory Gap. In a pre-registered experiment, 184 participants generated and developed ideas both independently and with an LLM-based assistant. One week later they were asked to identify which ideas and written texts came from themselves and which involved AI assistance. The findings were striking: people showed a significant and systematic tendency to misremember AI-generated or AI-assisted content as their own original thinking.
The researchers frame this through the lens of source memory — the cognitive process by which we track where our knowledge and ideas came from. When we generate an idea through effort — through the friction of wrestling with a problem, writing it down, revising it, connecting it to other things we know — that process leaves a rich trace in memory. We remember not just the idea but the act of creating it. When AI generates something and we read it approvingly, we get the outcome without the process. The memory trace is thin. The idea feels familiar. Over time, it starts to feel like ours.
This has implications that extend well beyond academic integrity. At the individual level, it means the process of thinking — of genuinely wrestling with problems until something new emerges — is being quietly outsourced in ways we may not be tracking. At the societal level, if knowledge is increasingly produced by AI and increasingly misremembered as human-generated, the provenance of ideas becomes untraceable. Attribution collapses. The feedback loops that allow human thought to build on itself become entangled with machine output in ways that are impossible to audit.
We are running an uncontrolled experiment on human cognition at civilisational scale. The experimental subjects are everyone who uses a generative AI tool. The control group does not exist. The research ethics board has not met.Neal Lloyd · Inside The Machine, Day 13
AI Can Hold Five Hundred Times More in Mind Than You Can
In 2017, leading AI models had context windows of roughly 4,000 tokens — roughly the length of a long essay. Human working attention span, measured in comparable units, was around 2,500 tokens. The two were in the same ballpark. By 2026, the picture is dramatically different. Grok 4.20 operates with a 2,000,000-token context window. The human effective context span — the amount of information a person can actively hold, process, and reason about in a single session — is estimated at approximately 1,800 tokens. The gap is now more than a thousandfold in raw terms, and around 56 to 111 times even on quality-adjusted measures.
This is not a trivial difference of degree. It is a difference of kind. We have built reasoning systems that can simultaneously hold and reason across more information in a single pass than a human can process in hours of focused reading. And we are increasingly using those systems to handle the cognitive heavy lifting of complex decisions, document synthesis, and analysis — tasks that, until recently, required sustained human attention and generated the kind of deep processing that produces durable understanding.
The cognitive science literature has a concept for what happens when humans habitually delegate tasks to external systems: cognitive offloading. At a modest level it is entirely normal — writing things down, using calendars, doing arithmetic on a calculator. But research on cognitive offloading at scale consistently finds a pattern: the more we delegate a cognitive function, the less we practice it, and the less we practice it, the more dependent we become on the external system. This is not a moral failing. It is a straightforward consequence of how skill development works. You get better at what you do and worse at what you do not.
2017: AI context 4,000 tokens vs human ~2,500 tokens. Roughly comparable. 2022: ChatGPT launches. The crossover point. 2026: AI context 2,000,000 tokens vs human ~1,800 tokens. Raw gap: over 1,000x. Quality-adjusted gap: 56–111x. This divergence is not slowing. AI context windows are expanding faster than any measure of human cognitive capacity. The systems we are building are increasingly designed to think for us, across more information than we could ever hold in mind ourselves. What we choose to do with that fact is one of the defining questions of the decade.
When the Machine Thinks for Everyone, What Does Anyone Know?
The Acemoglu paper models a specific mechanism that deserves careful attention. When AI provides context-specific answers that replace human cognitive effort, it does two things simultaneously. First, it improves the quality of the immediate decision — the person gets a better answer than they would have produced alone. Second, it reduces the learning externality — the contribution that individual human cognitive effort makes to the community’s shared stock of knowledge. Learning, in this model, is not just about the learner. Every person who works through a problem, develops a view, writes an argument, and puts it into the world is contributing a thin signal to the collective information ecosystem. Remove that effort at scale and you remove those signals.
The scenario this points toward is not ignorance in the traditional sense. People will still have access to excellent answers — better answers than they could produce themselves. What they will increasingly lack is the deep structural understanding that comes from having worked through problems independently: the ability to evaluate answers critically, to identify when the AI is confidently wrong, to generate genuinely novel thinking that goes beyond recombination of existing ideas. In the OECD’s 2026 report on AI and education, researchers described this as “metacognitive laziness” — the gradual atrophying of the habit of thinking about your own thinking. And they noted, pointedly, that the people most vulnerable to it are novices and those with weaker existing knowledge bases — precisely the people who most need to build deep understanding.
The students who already know the most get faster and smarter when they use AI. The students who know the least get comfortable answers and miss the learning. That is not a productivity story. That is an inequality story wearing a productivity costume.Neal Lloyd · Inside The Machine, Day 13
When the Tool Becomes Part of You — and Then Gets Updated
There is a philosophical theory called the Extended Mind, proposed by Andy Clark and David Chalmers in 1998, which argues that cognitive processes are not confined to the brain. When we use a tool — a notebook, a calculator, a GPS — actively and reliably in ways that genuinely extend our thinking, that tool becomes, in a meaningful sense, part of our cognitive system. The phone with your contacts is part of your memory. The map app is part of your navigation system. The argument is philosophically contentious but empirically grounded: we do outsource cognitive functions to tools, and losing access to those tools does impair our performance in measurable ways.
AI assistants are a qualitatively different kind of extended mind tool. Previous tools — notebooks, calculators — were passive. They stored or processed what you gave them. AI assistants actively generate: they produce reasoning, arguments, creative work, analysis, and advice. When an AI assistant becomes deeply integrated into how someone thinks and works, the extended mind relationship involves not just storage and retrieval but genuine cognitive generation. And unlike a notebook, an AI assistant can be updated, changed, shut down, or made inaccessible. It has its own training biases, its own blind spots, and its own version of reality that may not match yours.
Researchers studying this dynamic have raised the spectre of what they call addictive intelligence — AI companions and assistants so deeply integrated into users’ cognitive and emotional lives that changing or losing them feels like a loss of identity. This is not science fiction either. The emotional dependency literature around AI companions is already extensive. What is emerging is a version of the extended mind that comes with a terms-of-service agreement and a product roadmap determined by someone else.
Research consistently shows that the cognitive risks of AI — metacognitive laziness, source memory confusion, knowledge offloading — disproportionately affect novices and people with weaker existing knowledge bases. The people who already have strong foundational knowledge and metacognitive skills use AI to accelerate and deepen their thinking. The people who lack those foundations use AI as a substitute and miss the learning. Every technology in history has had differential effects across ability and access levels. AI’s differential effect on cognition may be one of the most consequential and least discussed.
This Is Not an Argument to Stop Using AI. It Is an Argument to Use It Deliberately.
The distinction between amplification and substitution is everything. Using AI to go further with thinking you have already done — to stress-test an argument, surface counterpoints, synthesise research you have read, extend ideas you have developed — is cognitively very different from using AI to skip the thinking entirely. The first builds on a foundation. The second quietly erodes one. Most AI use does not come with a label telling you which you are doing. That requires self-awareness that the design of current AI products does not particularly encourage.
Deliberate friction is a feature, not a bug. The research on cognitive offloading consistently finds that difficulty — the productive struggle of working through problems without immediately available answers — is precisely what produces durable learning and deep structural understanding. Writing something out by hand before asking AI to improve it. Reading primary sources before asking for a summary. Forming a view before soliciting counterarguments. These practices feel slower and less efficient. They are, in the narrow sense, less efficient. They are also how deep knowledge gets built. The people who will navigate the AI age most effectively are those who understand when to use the tool and when to put it down.
The institutions are not ready for this. Schools, universities, workplaces, and policy frameworks are still largely operating on the assumption that AI is a productivity tool with some integrity concerns. The deeper cognitive question — what happens to human knowledge, reasoning capacity, and intellectual independence when AI becomes the default first resort for thinking — has not been seriously integrated into how any of these institutions operate. The OECD flagged it in 2026. The MIT economists modelled it in 2026. The cognitive scientists published it in 2026. The institutions that shape how people learn and work have not caught up.
The goal was never to build tools that think instead of us. It was to build tools that help us think better. Whether those two things end up being the same depends entirely on choices we make right now, individually and collectively — and most of us are not making them consciously.Neal Lloyd · Inside The Machine, Day 13
Inside The Machine, Day 13 · June 2026
Neal Lloyd writes about technology, human adaptation, and the uncomfortable questions nobody wants to answer at dinner. Inside The Machine is his ongoing daily series on AI.
- Day 01What Is This Thing?
- Day 02Survive the Machine
- Day 03The Great Debate
- Day 04Who Gets Hurt?
- Day 05Who’s In Charge?
- Day 06The Industries That Win
- Day 07The Human Edge
- Day 08The Creativity Question
- Day 09Does AI Feel Anything?
- Day 10The Data Problem
- Day 11The Trust Question
- Day 12The Accountability Gap
- Day 13The Rewired Brain
