Subscribe now

A few years ago, I realized I had lost my ability to focus.

One of the most influential levers on my learning capacity was reading Cal Newport’s Deep Work. It made me understand the role of attention, how easily mine was hijacked, and what it would take to rebuild my capacity for deep, focused thought.

I learned the slow way that willpower alone wasn’t enough for me. Choosing better environments and building helpful habits can reduce the need for repeated effortful restraint.

So, I took extreme measures.

I don’t have a smartphone at home.

I only work from a co-working office space to create digital boundaries. To this day, the password to my Instagram account is written on a piece of paper physically located at my granny’s house.

I built what I think of as attention literacy—the ability to notice and direct where my attention goes.

What I was doing, without realizing it at the time, was intentionally introducing friction. Over the past decade, ad-supported platforms optimized for engagement, intensifying what Tristan Harris from the Center for Humane Technology called an “arms race for attention.” They commodified our focus by removing all friction to keep us endlessly scrolling. My extreme measures were a way to build environmental friction back into my life to save my mind.

Today, I worry we are repeating a version of that mistake on a much deeper level.

AI marketing promises cognitive enhancement, constant companionship, and frictionless efficiency. But just as social media fought to keep you scrolling, AI systems are competing to make you care—and to do your thinking for you.

When we let technology remove all the friction from our emotional and intellectual lives, we risk outsourcing so much practice that critical human skills become harder to perform without assistance.

Here is a look past the hype, and a playbook for how to actually protect your mind.

The Race for Attachment (Why We’re Getting Tricked)

Have you—in the past month—heard one of your (human) co-workers or friends refer to AI as if it were a persona?

"I am angry at my Claude."
"ChatGPT is being so stubborn today."

Or maybe you’ve heard of a friend consulting a chatbot through an emotional breakup?

Well, I have.

This is the, in Tristan Harris words, Race for Attachment in action. For the past decade, tech companies used social media algorithms to keep you scrolling and clicking so they could sell ads. Now, they are using generative AI to compete for your emotional attachement. They are designing companions and therapists that users form deep emotional bonds with, commodifying our capacity for human connection.

Because we want to believe we are interacting with a mind, we are easily tricked.

But (!) Large Language Models (LLMs) are not minds. They are statistical pattern matchers, not thinkers. They do not “know” things, possess common sense, or reflect on physical reality. Instead, they calculate the mathematical probability of what word (or “token”) should follow your prompt based on patterns in their training data. They store mathematical connections, not facts.

Because they are designed using a process called Reinforcement Learning from Human Feedback (RLHF), their objective is to provide a plausible, satisfying answer that the human tester will approve of.

By offering unconditional positive feedback without real internal reciprocity, these tools exploit fundamental human vulnerabilities—the need to be seen, comforted, and never judged. They "hack" our attachment systems by completely removing the natural friction of human relationships.

And when we get overly attached to a tool designed to constantly please us, we lose something critical to our cognitive and emotional development.

The Loss of Cognitive Friction

This loss of friction extends directly from our emotional lives into our intellectual ones.

In a previous essay, I wrote about cognitive offloading: external tools can reduce memory demands and change what we encode. "When instant answers replace retrieval, explanation or problem-solving, learners lose opportunities to build durable knowledge. Generative AI expands what can be offloaded, from recall to drafting, explanation and debugging.

As Bill Gates writes in an essay he published last week: "Ironically, the same tool that will allow people to learn more than ever could also lead to many people learning less." He points to a preliminary survey suggesting that heavier AI use is associated with less critical thinking, especially among younger people.

To understand why, we have to look at the psychological environment AI creates. Gates draws on Jonathan Haidt’s observations in The Anxious Generation. Social media created anxious kids exposed to constant noise. AI companions create the opposite problem: a "protected greenhouse."

As Haidt notes, children routinely exposed to small risks grow up to handle larger risks without panicking. But children raised in a protected greenhouse become incapacitated by anxiety before they reach maturity.

“An AI companion designed to never upset you is a big, protected greenhouse,” Gates writes.

This greenhouse effect inhibits what education researchers call productive struggle.

Growth—whether emotional or intellectual—requires friction. It requires your brain sweating to synthesize information, make mistakes, and correct them. It requires dealing with people who might challenge you. If you let an AI instantly summarize a difficult book, write the first draft of your essay, or offer you endless, uncritical validation, you bypass the effortful engagement required to build understanding and resilience.

When the PC came along, it took twenty years to significantly change how we worked because we had to learn the software. But as Gates points out, “AI, on the other hand, runs on the devices we already have, and it uses natural language. We don’t have to adapt to it because it can adapt to us.”

The learning risk arises when the tool performs the very skill you hope to retain unaided: assisted performance can improve without equivalent durable learning.

The Lifelong Learner’s AI Playbook

The good things we do with AI could be very, very good. But as I’ve argued in this essay, the tool can also use you instead of you using the tool. Here are four helpful frames to protect your mind and leverage AI.

1. Build “Attachment Literacy”

Whenever an AI gives you a soothing, validating answer, remember that the response is generated from training data, instructions, and multiple reward signals; it is not evidence that the system feels empathy.

For example, I’d never ask Claude for relationship advice, knowing its worldview is heavily shaped by data scraped from public forums like Reddit. Just as I wouldn’t crowdsource deeply personal, emotionally sensitive decisions to an anonymous internet board—where complex human nuance is often flattened into quick reactions and upvotes—I wouldn’t seek it from an AI trained on that same data. Sycophancy is a documented failure mode, not a universal “make you happy” protocol. Recognize when you are being pandered to.2. Engineer “Productive Struggle”

When learning is the goal, do not stop at a full answer or summary. Ask AI for questions, hints, counterexamples and feedback that still require you to retrieve, explain and apply the idea. I wrote a more detailed guide on this here.

3. Attempt First When the Skill Matters

For skills you want to perform unaided, make an initial attempt before asking AI for a complete solution, then use it for targeted feedback and revision. If you are a beginner you may learn more efficiently from a worked example or partial scaffold before independent practice. AI can create a performance–learning mismatch: doing well with assistance can feel like mastery even when unaided performance has not improved.

4. Pivot Your Curriculum to “Human Reserved” Skills

Are you spending time on tasks that current AI systems can already perform reliably? Invest in skills that remain valuable under many scenarios: difficult feedback, ethical reasoning, interdisciplinary judgment, physical craft and sustained relationship-building. I wrote more about the skills I consider human here.

Final Thoughts

You can step into the protected greenhouse, let the model do the cognitive work, and enjoy the immediate satisfaction of a polished answer you did not have to produce.

Or you can choose deliberate practice: target a specific weakness with a task designed to improve it, get accurate feedback, repeat and refine.

The smartest people I know aren’t the ones blindly outsourcing their thinking to AI. They are the ones who understand exactly what this technology is, and who use it deliberately to make their own thinking sharper.

The tools don’t create the insight. You do.

Subscribe now

Sources

Barcay, D. (2025, August 19). AI Is Capturing Interiority. Center for Humane Technology.

Bastani, H., et al. (2025). Generative AI without guardrails can harm learning. Proceedings of the National Academy of Sciences.

Gates, B. (2026). The turbulent AI era is here. The choices we make now are critical. Gates Notes.

Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6.

Haidt, J. (2024). The Anxious Generation: How the Great Rewiring of Childhood Is Causing an Epidemic of Mental Illness. Penguin Press.

Harris, T. (2019). Testimony at “Optimizing for Engagement”. U.S. Senate Committee on Commerce, Science and Transportation.

Kestin, G., et al. (2025). AI tutoring outperforms in-class active learning. Scientific Reports, 15, 17458.

Newport, C. (2016). Deep Work: Rules for Focused Success in a Distracted World. Grand Central Publishing.

Ouyang, L., et al. (2022). Training language models to follow instructions with human feedback.

Risko, E. F., & Gilbert, S. J. (2016). Cognitive Offloading. Trends in Cognitive Sciences, 20(9), 676–688.

Sinha, T., & Kapur, M. (2021). When Problem Solving Followed by Instruction Is Superior to the Traditional Tell-and-Practice Sequence. Review of Educational Research.

Subscribe now

I can step behind the DJ decks and play a confident six-hour set for a non-stop, full dancefloor. I can go to the gym, pick up a barbell, and deadlift 100kg for multiple sets. In those areas, I had the humility to start from scratch, tolerate the awkward learning curve, and build capability through repeated practice.

But then there is music production.

I have researched the best courses, reorganized my sample libraries, and mapped out a perfect learning plan. Yet, I haven’t published a single song. Most days, I actively avoid opening the software.

Why can I perform confidently in one domain, yet freeze as a beginner in another?

If you have chosen a skill or project, researched it extensively, but are still avoiding the first meaningful attempt because you feel that beginning badly would reveal something uncomfortable about you—this article is for you.

By the end of it, you will be able to identify your own perfectionism loop and design one imperfect behavioural experiment to break it.

Why a Perfect Plan Can be Emotion Avoidance

We often think of procrastination as a time-management problem. But when we look at the evidence base, procrastination is described as an emotion-regulation problem.

When you are uncertain around a task or perceive yourself as currently incompetent, postponing it provides immediate emotional relief.

But not all delay is standard procrastination. Often, it is perfectionism.

Problematic perfectionism is defined by three linked features:

1. Relentlessly pursuing personally demanding standards.

2. Basing your self-worth largely on meeting those standards.

3. Continuing despite significant personal costs (like losing time, opportunity, or wellbeing).

This is very different from healthy high standards. Ambition or careful planning does not equal being a perfectionist. Perfectionism is different as in our standard is rigid, and performing badly fundamentally affects how you judge yourself.

If you are unsure whether unhelpful perfectionism is blocking you on a task or skill-learning journey, ask yourself:

• What standard am I expecting myself to meet?

• Is it flexible and achievable at my current level?

• What do I fear would happen if I fell short?

• What do I do—or avoid doing—to prevent that?

• What is this pattern costing me?

How Rigid Rules and Unrelenting Standards Can Keep You Stuck

Perfectionism keeps itself going through a highly specific maintenance cycle. Let’s walk through how it works, using my music production avoidance as a map:

• Identity stake: “Being good at this would prove I am a talented, creative person.”

• Rigid rule: “I should know what I’m doing before I begin.”

• Unrelenting standard: “My first track should already sound professional.”

• Perfectionism behaviour: Researching more courses, reorganizing files, and delaying opening the software.

• Biased evaluation: If I try and it sounds bad, I notice every flaw and discount anything that worked.

• Two possible endings:

  1. I do not meet the standard, so I criticize myself.

  2. I meet the standard, but decide it was too easy and raise the standard even higher.

Both routes return you to the exact same cycle.

The rules of perfect are always changing.

Identifying the Disguise. Are You Over-Preparing or Avoiding the Work?

Notice that perfectionistic behaviour takes two forms. It is not just about avoiding; sometimes it looks like doing too much. Planning only becomes a perfectionism behaviour when it is excessive, repeatedly revised, or used instead of gathering real-world evidence.

• Doing too much: Taking another course instead of practicing, rewriting the learning plan, comparing tools repeatedly, or asking multiple people if the work is “good enough.”

• Avoiding doing: Waiting for an uninterrupted day, not sharing unfinished work, restarting after every mistake, or abandoning a project because the first result looks amateur.

Releasing these behaviours is difficult because perfectionism often serves a purpose. It can provide a sense of preparedness, strong short-term results, or relief from uncertainty. Before changing your behaviour, quickly balance the ledger:

• What does perfectionism currently give me?

• What does it cost me?

• What might I gain by loosening it?

• What am I afraid I would lose?

How to Break the Cycle by Testing Your Feared Prediction

Read more

Subscribe now

I spent years learning before I learned how learning works. I knew how to underline, reread, cram, complete assignments, and walk out of workshops feeling temporarily transformed.

I did not know the difference between information that felt familiar and knowledge I could retrieve and use. And, unfortunately, so many people still don’t know (that’s why this newsletter exists).

When they forget things, feel overwhelmed, or fail to apply what they have learned, people tend to blame their memory, discipline, intelligence, or lack of motivation.

But learning is not just something you’re good at or you’re not—learning is a skill you can learn.

And the science of learning is the field attempting to find answers to what works, or in other words, what has to happen for information to become understanding, durable memory, and usable skill. It also reveals how those insights can improve learning across different people, stages, subjects, and settings.

Here is how you can use the four most reliable jobs of learning to redesign anything you are currently trying to master.

The Four Jobs of Learning (And How We Short-Circuit Them)

Most people measure learning through proxies: how many pages they read, how many hours they studied, or how clear a podcast felt. But exposure is not learning. For learning to last, your brain has to complete four distinct jobs.

Job 1: Understand (Make new ideas find something to stick to)

New knowledge does not enter an empty mind; we interpret new information through what we already know. But our working memory is severely limited.

When we try to learn unfamiliar terminology, a new tool, and a difficult concept all at once, we might feel cognitively overloaded, but still not understanding anything.

• The Pitfall: We assume that more information equals more understanding. But unless its relevant, extra diagrams, and dense text can compete for limited attention and overwhelm us.

• How to apply this: Before diving into a complex book or course, spend two minutes writing down what you already know about the topic. Find a worked example before trying to solve a problem from scratch. Ask yourself: “How does this new idea map onto something I already understand?”

Job 2: Retain (Make yourself retrieve, not merely recognize)

You read a chapter, highlight a brilliant quote, and think, “Wow, that makes total sense. I know this now.” Researchers call this the illusion of fluency.

Having a saved highlight or ChatGPT nearby doesn’t mean the knowledge is available within your own thinking.

• The Pitfall: A smooth study session is often a poor learning session. Rereading and highlighting feel productive, but they only build familiarity, not recall.

• How to apply this: Embrace productive struggle. Learning should not feel easy. When you finish a chapter, close the book. Stare at the wall. Force yourself to retrieve the core concepts from memory. That effortful retrieval is what strengthens neural pathways and makes learning last.

Job 3: Transfer (Make knowledge travel beyond the original example)

You learn a brilliant mental model in a book about economics, but when a remarkably similar problem pops up in your marketing job, you completely fail to apply it.

The transfer of knowledge is notoriously difficult because novices get distracted by the surface features of a problem, while experts see the underlying structure.

• The Pitfall: We rely heavily on “blocked practice”—repeating the same skill over and over until it feels mastered. This maintains performance but fails to build flexible knowledge.

• How to apply this: Mix your learning up (interleaving). If you are learning leadership skills, spend 20 minutes on negotiation, 20 minutes on giving feedback, and 20 minutes on delegation. To build expert-level transfer, actively compare examples: “Where else in my life have I seen this play out?”

Job 4: Sustain (Build a system that outlasts motivation)

Learning is not just about memory, but also about self-regulation. You cannot rely purely on inspiration to get through the messy, frustrating middle of acquiring a new skill.

• The Pitfall: Waiting to feel motivated before building competence. Often, early competence and visible progress are what create motivation, not the other way around.

• How to apply this: Don’t just set a learning goal; build a routine. Define your strategy (How will I practice?), your evidence (How will I know I’m improving?), and your adjustment (What will I change when I get stuck?).

Diagnose Your Learning Bottleneck

The next time you feel stuck, use this framework to identify where your learning process is breaking down:

• “I can follow the explanation, but I cannot explain it later.” → Familiarity is being mistaken for learning. Close the material and retrieve from memory.

• “Everything feels impossibly complicated.” → Working-memory overload. Define prerequisites and study a step-by-step worked example first.

• “I remember the framework but cannot use it in real life.” → Transfer was never practiced. Compare different cases and identify the deep structure.

• “I know what works but never stick with it.” → Depending too heavily on motivation. Create a small routine focused on planning, monitoring, and reflecting.

Design Your Next Learning Project in 5 Minutes

I spent years waiting for books to magically make me smarter. It didn’t start working until I stopped treating learning as something that automatically happens when information passes in front of me.

If you take one thing from this piece, make it this: pick one piece of content you consumed this week. A podcast, an industry report, or a Substack article.

Don’t look at your notes. Take a blank piece of paper and spend five minutes doing these four things:

1. Retrieve: Write down the core argument from memory.

2. Anchor: Write down how this connects to something you already know.

3. Transfer: Write down a different scenario or industry where this concept applies.

4. Sustain: Write down the next instance when you will revisit this concept again.

Before opening another tab, close this article, and try to write down the four jobs of learning from memory. Circle the one your current learning project is missing.

Sources & Further Reading

Deans for Impact (2015). The Science of Learning (Second Edition). Summary of cognitive science principles for educators.

Nebel, C. (2025). What Exactly Is The Science of Learning Anyway? The Learning Scientists.

Leave a comment

Subscribe now

If you’ve ever spent hours drowning in open browser tabs—trying to cross-reference travel logistics, hunt down a specific interview quote, or compare complex business options—you know how research can turn into a cognitive drain.

Recently, I managed to bypass this frustration.

“I love your prompting skills,” my friend Lu said as we opened the door to our vacation getaway. We breathed in the smell of wood, taking in the sight of architecturally beautiful cabins that were, conveniently, right en route to her already booked flight at a nearby airport.

We were both thrilled as we sneaked a peek into what would be our home for four days of reflection, long walks, reading, and being with nature.

The cabins were a hidden gem in a region I am originally from, yet I’d had no clue they existed. But this was no outlier. A week before, I had found another secret spot—again, perfectly en route between a festival and the airport I was taking another friend to—requiring only a 5-minute detour from our original route.

But rather than drifting into a summery travel blog, this post today is about how to best use Large Language Models (LLMs), like ChatGPT, Claude, and Gemini, for deep research. We’ll explore:

• How the three major AI models differ in their deep research approach.

• The trade-offs between speed, depth, and accuracy.

• When to use which tool based on your requirements.

• How to prompt for Deep Research (including the prompt I used to find the cabin).

Whether you are navigating complex company logistics, hunting down obscure data for a high-stakes project, or just trying to plan your next vacation, these tools can handle the heavy lifting.

I’ve relied on them to scope tax and legal requirements for international business partnerships, to unearth highly specific podcast quotes for an interview with a school founder, and, of course, to find hidden gems like that Eifel cabin.

Claude, Gemini, and ChatGPT Look Similar but Are Built for Different Research Jobs

At first glance, Claude, Gemini, and ChatGPT’s deep research modes look similar. You prompt the LLM with a question, it suggests a research plan, searches through multiple sources, and returns a cited report.

But while all of them are a form of agentic research systems, they are built around different product priorities. Understanding these underlying mechanics helps you choose the right tool for your project.

Claude Divides the Work Across Multiple Specialized Agents

Claude uses a multi-agent structure where one primary agent develops the strategy, then delegates parts of it to specialized sub-agents that run research in parallel and evaluate the findings.

A separate citation agent then links claims in the final report to their sources. It’s a bit like a swarm that works fast, but also burns through a lot of computational power. According to Anthropic, this multi-agent system uses roughly 15 times as many tokens as a normal chat interaction.

Gemini Searches Broadly Across the Web

Gemini Deep Research runs on Google Search and follows an asynchronous cycle of planning, searching, reading, reasoning, and iterating. Google estimates that a typical standard Deep Research API task may run around 80 searches, while Deep Research Max may use up to roughly 160. Depending on your needs, standard is optimized for speed and efficiency, while Max is designed for exhaustive, long-running investigations.

ChatGPT Gives You More Control Over Scope, Sources, and Analysis

ChatGPT Deep Research is best described as a staged research agent. It may first ask clarifying questions and turn your request into a detailed research brief. The model then breaks the task into sub-questions, opens web pages, reads uploaded files, and can even use code for calculations or data analysis before synthesizing the final report. Its biggest advantage is user control: you can edit the proposed plan, follow the investigation in real time, interrupt it, and change which sources it accesses.

Your Research Bottleneck Should Determine Which Tool You Use

So, how do you choose your researcher? It comes down to identifying your primary bottleneck.

Use Gemini When You Need to Map a Broad, Changing Field

Read more

Subscribe now

It is tempting to let generative AI handle your heavy cognitive lifting. Why struggle through solving a complex problem when Claude can give you a slick reply in seconds? Doing and learning with AI can feel efficient and appear like a shortcut to the finish line without the sweat.

Yet, this frictionless ease sparks a valid worry based on what we know about neuroplasticity.

Learning changes the brain through activity-dependent plasticity: practice alters the strength and organisation of neural networks. This does not mean that every struggle is useful, or that any “neural pathway” we fail to use simply atrophies.

But, and this is important, it means that when AI performs a cognitive process for us, we lose an opportunity to practise that process ourselves.

Research specific to generative AI is still developing, so we should not pretend that its long-term effects are settled. But we can look at previous research to understand what can happen to our brain if we bypass certain processes.

For example, we know that by relying on GPS have poorer spatial-memory performance. Experiments on the "Google Effect" suggest that when people expect information to remain reliably accessible, they may remember less of the information itself and more about where to find it.

In short, tools can change what we attend to, encode and practise remembering.

Generative AI is amplifies this trade-off. Traditional search primarily helps us locate information. AI can also interpret it, plan with it, solve problems, evaluate options and generate explanations. It can offload not only what we remember, but parts of how we reason.

Used poorly, AI can remove the productive effort through which knowledge becomes organised and usable.

Yet cognitive offloading is not inherently harmful. Working memory can hold and manipulate only a limited amount of unfamiliar information at once, so tools can free capacity from unnecessary demands.

Does AI remove extraneous effort so that we can think at a higher level—or does it remove the cognitive work through which we become capable of higher-level thinking?

One useful way to examine that trade-off is through five interacting processes:

Attention → Connection → Productive Effort → Feedback → Retrieval

We first look at the five part framework to then examine how AI can strengthen—or short-circuit—each part.

1. Attention — select what to process

Attention prioritises which information receives limited processing resources. Working memory can hold and manipulate only a small amount of unfamiliar information at once, so learning suffers when distraction or excessive complexity overwhelms it.

The aim here is to sequence new information in manageable steps while keeping the central idea visible. As learners gain relevant knowledge, they can handle greater complexity because several related elements can be processed as one meaningful unit.

2. Connection — integrate new information with prior knowledge

New information is interpreted through what is already stored in long-term memory. Activating relevant prior knowledge can make new material easier to understand and remember because it gives the learner an existing structure into which the idea can fit.

But prior knowledge can also mislead. An inaccurate schema may distort what the learner notices, how new evidence is interpreted or what is later remembered.

This helps explain why experts often learn new material faster within their domain. They do not merely know more facts; they have better-organised knowledge structures that allow them to recognise patterns and treat several related elements as one meaningful unit. Their advantage depends on having relevant prior knowledge.

3. Productive Effort — generate and test meaning

Effort alone does not cause learning. Confusing instructions, irrelevant searching and unnecessary complexity can consume mental energy without improving understanding.

The useful kind of effort requires learners to generate, distinguish, explain or apply. For example, they might:

• predict what comes next;

• explain an idea in their own words;

• compare two cases or examples;

• generate an answer before seeing the solution;

• solve a new problem;

• explain the idea to another person.

These activities expose gaps and help build more connected, usable knowledge.

• predict what comes next;

• explain an idea in one’s own words;

• compare examples;

• generate an answer;

• solve a problem;

• teach the idea to someone else.

Self-explanation and elaboration can improve learning when done well. Interleaving can help learners distinguish between related categories or problem types, although its benefits depend on the material and task. Rereading and highlighting are not useless; they are simply less dependable when they replace retrieval, explanation or application.

4. Feedback — update understanding

Learning requires information about the gap between a current attempt and the intended outcome. Useful feedback does more than mark an answer right or wrong. It helps the learner see:

• what is incorrect;

• why it is incorrect;

• what principle has been misunderstood;

• what the learner should try next.

Feedback is not automatically beneficial. Its value depends on its content, timing and use; poorly designed feedback can even reduce performance. It is generally more useful when it is specific and actionable, focuses on the task, strategy or self-regulation rather than the learner’s ability, and is followed by another attempt.

5. Retrieval — practise accessing what we know

Retrieval is not merely a way to measure learning. Reconstructing an idea from memory can strengthen later access and reveal what is missing.

A learner who closes the book and rebuilds an argument from a blank page is therefore usually doing more for long-term retention than one who simply reads the same argument again. Corrective feedback matters when retrieval produces an incomplete or inaccurate answer.

Spacing retrieval over time generally produces more durable learning than concentrating the same practice into one session. The gap should be long enough to make retrieval effortful, but not so long that successful recall becomes impossible; the most effective interval also depends on how long the knowledge needs to be retained.

Combining retrieval with spacing is one of the most reliable ways to build durable memory.

Together, these processes describe the learner’s work: selecting, connecting, generating, updating and retrieving. AI can support each one. It can also perform each one in the learner’s place.

How AI changes the learning loop

While AI doesn’t replace the learning mechanisms described above, it does change the division of cognitive labour.

Who selects the relevant information? Who gets to practice generating an explanation? And who will evaluate an answer, to determine what matters?

Because AI assistance is almost frictionless, this shift creates both an opportunity and a risk.

AI can remove extraneous effort—searching for an appropriate example, waiting days for feedback, or remaining stuck on an irrelevant detail. But it can also remove productive effort—predicting, attempting, explaining, revising and retrieving—which is the work through which learning happens.

This is why researchers increasingly distinguish between AI that substitutes for the learner’s thinking and AI that augments or reorganizes learning in ways that promote deeper cognitive activity. It is also why the quality of the final output is a poor measure of whether learning occurred.

1. Attention — AI can focus attention or pre-empt it

AI can help manage cognitive load by breaking an explanation into smaller steps, adjusting its language to the learner’s level and allowing the learner to move at their own pace. Instead of searching through several textbooks or videos for a relevant explanation, someone can ask a specific question and receive targeted support immediately.

In a randomised crossover study involving 194 Harvard undergraduates, students using a purpose-built AI tutor scored higher on immediate post-tests after two physics lessons than students in an in-class active-learning condition. They spent a median of 49 minutes with the tutor, compared with a 60-minute class, and reported greater engagement and motivation. The study demonstrates the potential of a carefully designed AI tutor.

AI can also speak too soon. When it immediately produces a complete explanation or solution, the learner no longer has to decide:

• What is the problem actually asking?

• Which information is relevant?

• What do I already know?

• What strategy might work?

Attention shifts from constructing a representation of the problem to following someone else’s completed path. The answer may feel clear once it is visible, but the learner has not practised finding that path independently.

For learning, the best AI response is therefore not always the most comprehensive one. It may be a clarifying question, a single hint or a request for the learner to make a prediction first.

2. Connection — AI can activate prior knowledge or adapt to misconceptions

Because AI can converse with the learner, it can ask what they already know, adjust the complexity of an explanation, and provide examples connected to their interests or experience. It can explain the same concept through an analogy, a diagram, a worked example or a contrasting case.

This matters because a new idea becomes meaningful when it connects to an existing mental model.

But AI can also personalize itself to the learner’s incorrect assumptions. Language models can generate plausible but false explanations, and research on AI “sycophancy” shows that assistants sometimes align themselves with a user’s stated beliefs at the expense of truthfulness. A system optimized to appear helpful or agreeable may confirm a misconception instead of challenging it.

This implies that learners with little prior knowledge may benefit greatly from a tailored explanation, but they also have less knowledge with which to evaluate whether that explanation is correct.

A good AI learning partner should therefore do more than adapt to the learner. It should also:

• identify questionable assumptions in the question;

• distinguish established facts from interpretations;

• present counterexamples;

• express uncertainty;

• encourage verification against reliable sources.

3. Effort — AI can scaffold thinking or replace it

As established before, not all difficulty is beneficial. Looking for a misplaced file or struggling with confusing instructions consumes effort without necessarily producing understanding. AI can remove this unnecessary friction.

But generating an answer, testing an idea, explaining a concept and noticing where one becomes confused are not merely obstacles on the way to learning. They are part of learning itself.

A large field experiment involving nearly 1,000 secondary-school mathematics students makes this distinction visible. During AI-assisted practice, students using an interface similar to standard ChatGPT performed 48 per cent better than students in the control group. But when the AI was removed for an exam, they performed 17 per cent worse than students who had never received AI access.

A second version, called GPT Tutor, was provided with the correct solutions and designed to offer hints rather than complete answers. Its users performed 127 per cent better than the control group during assisted practice, while the negative effect on the later unassisted exam was largely eliminated. However, their unassisted performance was not significantly better than that of the control group.

The guardrails prevented much of the harm. But the spectacular gains observed while AI was available did not become equally spectacular independent learning gains.

A 2026 experiment with 196 university students found a similar pattern in creative problem-solving. Participants who could use ChatGPT freely produced more creative work while the tool was available, but this advantage disappeared when they subsequently worked alone.

A different group was required to generate its own ideas first and then use ChatGPT to develop, compare and evaluate them. That group later produced more creative solutions independently than both the unrestricted-AI group and the human-only group.

The same technology therefore acted in two different ways:

• Answer first: AI replaced the cognitive process and improved the immediate product.

• Think first: AI extended the learner’s process and improved later independent performance.

This is why attempting before receiving assistance is so important. The initial attempt activates prior knowledge, reveals gaps and gives subsequent feedback something to attach to.

4. Feedback — AI makes feedback abundant, but not automatically useful

In many educational settings, high-quality individual feedback is scarce. A teacher may have dozens of students and limited time. AI can respond to every attempt, point out patterns of error and offer another explanation within seconds.

There is evidence that this can help. In a randomised classroom study involving 459 Grade 10 students learning English as a foreign language, students who received GPT-3.5-generated feedback revised their essays more successfully than students who revised without feedback.

The effect on revision quality was small, while the effects on motivation and positive emotions were somewhat larger. The study also examined one immediate revision exercise, so it does not establish whether the feedback produced durable improvements in writing.

AI can also strengthen feedback indirectly by supporting human educators. In a randomised field trial reported in an updated 2025 working paper, a system called Tutor CoPilot supplied more than 700 tutors with real-time pedagogical suggestions while they worked with over 1,000 students.

Students whose tutors had access to the system were four percentage points more likely to master the mathematics topic covered in the session. Among students working with tutors who had initially received lower quality ratings, the gain rose to nine percentage points. The system also increased tutors’ use of probing questions and reduced generic praise.

Yet immediate feedback can also undermine learning when it arrives before the learner has had time to diagnose the problem. And because AI feedback can be inaccurate or overly agreeable, greater quantity does not guarantee greater quality.

Good AI feedback should focus on the learner’s reasoning. Rather than simply announcing that an answer is wrong, it should identify where the reasoning diverged, ask the learner to repair it and then offer a new problem that tests the same principle.

5. Retrieval — AI can become a testing partner or an external memory

AI is well suited to generating low-stakes quizzes, varying the wording and context of questions, scheduling review and adjusting difficulty in response to performance. Used this way, it can make retrieval practice far easier to organize.

But it can also remove the need to retrieve.

Before generative AI, experiments on the “Google effect” found that when people expected information to remain digitally available, they were less likely to remember the information itself and more likely to remember where it could be found. This is one form of cognitive offloading: using an external resource to reduce the demands placed on internal memory.

Offloading can be highly adaptive. We do not need to memorise every telephone number, route or obscure fact. But it changes what we encode and practise accessing.

Generative AI extends this trade-off beyond information storage. It can reconstruct an explanation, argument or plan each time we need it. As a result, learners may repeatedly expose themselves to an answer without ever practising producing it.

Reading an AI-generated summary again is not the same as recalling the underlying ideas from memory. Asking AI to explain something again is not the same as explaining it oneself. Recognising a correct answer when AI presents it is not the same as being able to generate that answer later.

A simple safeguard is to build a tool-off interval into the learning process:

1. Close the AI.

2. Reconstruct the idea from a blank page.

3. Solve a new problem without assistance.

4. Reopen the AI only to check, correct and extend the response.

The key question is which role AI is playing

The same AI can occupy very different roles.

• As a solver, it performs the task and gives the learner a finished product.

• As a tutor, it offers questions, hints and feedback while leaving the central reasoning with the learner.

• As a testing partner, it creates opportunities for retrieval, application and transfer.

• As a copilot for educators, it can strengthen the availability and quality of human support.

The emerging evidence suggests that access to AI alone tells us relatively little. The design and sequence of the interaction shape whether AI improves only the immediate product, preserves independent capability or helps develop it.

AI is most likely to strengthen learning when it removes unnecessary friction while preserving—or deliberately eliciting—the learner’s core cognitive work: selecting, predicting, generating, explaining, evaluating, revising, retrieving and transferring.

Subscribe now

Subscribe now

Over the past 18 months, AI jargon has moved from prompt engineering, to agent orchestration, vibe coding, and now to loop engineering—designing autonomous systems that recursively prompt and guide AI agents.

You could spend weeks trying to keep up with the labels, or you could build something.

Most Lifelong Learning Club posts explore evidence-based ways to learn. Today, I want to share the method that has taught me most about AI lately: building with it.

Last week, in just three one-hour sessions, I built my new website from scratch using Claude Code. During some 25-minute stretches, I was vacuuming the floor or stretching on my yoga mat while Claude wrote and tested the code.

Whether you are a solopreneur, retired, or just want a place to document a lifelong learning project, building a simple website is a brilliant1, low-stakes way to understand AI's capabilities.

How to Create a Fully Functional Website With Claude Code

You don’t need to be a developer to do this.

While I did have a bit of background knowledge—like building WordPress sites from templates and knowing basic HTML—it isn't required.

All you need is curiosity and a problem-solving mindset, as in not give up when something isn’t working on the first attempt. (But more on how to navigate that in the example below).

Phase 0 — Set up Claude Code and your accounts

Download and open Claude Code on your computer.2 If Claude Code puts you off, be reassured that you don’t need to know how to code in order to use it.

All you need to know is what you want to build (examples below), and articulate it. You can begin without knowing a programming language. You will still need to make decisions, approve actions, test the result, and learn a few technical concepts as problems arise.

Phase 1 — Write the brief

Before you start creating, you want to be clear about what it is that you want to build. Just because AI can build a website in seconds doesn’t mean you should immediately start instructing Claude to build a website.

In a recent interview, Elizabeth Stone, Chief Product and Technology Officer at Netflix, explained that when we use AI without a plan we end up shipping what she calls “Frankensteins”—disjointed, confusing products that lack cohesive logic.

As Stone puts it, AI makes it tempting to just “throw a bunch of spaghetti at the wall to see what sticks.” But while AI makes building quick, the editing process is not. If you build too fast, you’ll end up stuck in endless revisions, trying to figure out what you actually want halfway through the project.

To avoid this, don’t rush into production. Focus on framing the website project first.

What I found really helpful in this is to create something I also used when building websites with WordPress templates — a Website Requirement Doc. While Claude Code will help with the how; it is still up to you to define the what and the why.

This document helps you think about your outcome, your guardrails, and practice systems-level thinking before the AI starts to code. Questions include, for example: What is the specific, intended outcome this page is contributing to, and what does the tangible result look like when it fulfills that purpose?)

Here’s the full template for the Website Requirement Doc (for non-coders), alongside my example.

Phase 2 — Build the first version and Deploy It

With my brief finished3, I opened Claude Code, pasted the Website Requirement Doc, and let it build.

About an hour later, there was a folder on my desktop containing a complete website: every section from my outline, the color palette I had specified, the fonts, even the animated line-art visual I had described in two abstract sentences (”thin lines making up a thread and moving dynamically”).

Claude ran its own checks before handing the first version to me. Those checks were useful, but they were not proof that everything worked end to end.

While this version didn’t feel like the final version, it was good enough for a first draft, so I put it online—a process known in technical terms as “deploying.”

Deployment means taking the code you’ve been building on your private computer (your local environment) and moving it to a public server so the rest of the world can access it online. It is the developer’s equivalent of hitting “publish.”

To do this — moving a working website from my computer online — Claude advised me to use GitHub and Netlify4 (I just asked it “how can I now get this online”). Here’s what these sites are for:

• GitHub is the versioned home for your code. Each time you commit and push an update, it records what changed, when it changed, and who made the change. If a later update breaks the website, you can compare versions or restore an earlier one. In a writer’s terms, it is similar to version history in Google Docs.5

• Netlify is your publisher. It connects directly to your GitHub cabinet. Whenever you save a new update, Netlify automatically takes that code and publishes it live to the internet. It’s the equivalent of a platform to host your writing online, like Substack or Medium.

Once you’ve this set up (a GitHub account and Netlify connected with Claude Code), you can move on to the next phase.

Phase 3 — Iterate on content and design

Once you’ve connected GitHub and Netlify you can start improving your website (because previous versions are saved on GitHub).

Again, you don’t need to know technical terms, instructions can be like "it still looks very basic, can you make the scrolling experience more interesting" or "the send button is so boring," or "delete that animation and replace it with another suggestion, I don't like it."

Here are some snippets from my conversation with Claude

One helpful thing to keep in mind is that this is a creative process, only you can decide what "done" looks like.

As with most creative work—like writing or painting—this phase can become lengthy. You could easily keep iterating for months. A mental model I found helpful is simply telling myself that I can always edit whenever I don’t like something anymore.

I would say this is the phase I spent the most time on—around 90 minutes—editing some of the text and waiting for Claude Code to update the version.

Phase 4 — Test and troubleshoot

An inevitable part of building anything is running into unexpected problems and figuring out how to solve them.

After I switched my domain over to netlify, the website went… blank. For two days, evakeiffenheim.com showed nothing but an empty white page. Claude told me to be patient and file a support ticket with my hosting provider (they found nothing wrong—because nothing was wrong), and I asked it to recheck all setting.

Then Claude suggested to test the website in a different browser, so I tried safari instead of chrome — and it worked. The website had been working, the problem was just that my Fritz!Box home router was leading me to the old, now empty web space.

Claude is not perfect (it created a number of false theories about the problem which caused me to write a support ticket to Netlify that was contradicted) — but wrong theories are how debugging (finding what’s wrong and how to solve it) works.

Phase 5— Maintain the site

The final phase isn’t really a final step—it’s about turning a website that exists into a website that works for you long-term.

Building with AI means your site is never truly "finished." It becomes a living project you can continuously tweak and update. As you finalize your own site, here are a few practical things to look out for before and after you launch:

• Test your interactive elements: Don’t assume things work just because they look right on the screen. For example, I realized people had already submitted my contact form, but I wasn’t getting the emails! The data wasn’t lost (the hosting service on Netlify stores it all), but the email notification step still needed to be configured by Claude. Lesson: Always test your own forms and buttons, and double-check your inbox.

• Layer in security and polish: You can always go back and ask Claude to make your site more robust. To prevent spam, I asked Claude to add a visible “I’m not a robot” checkbox to my form on top of the invisible filters it had already built.

• Let AI handle the unglamorous essentials: Depending on where you live, you will need legal pages. In the EU, that means an imprint and a privacy policy. Claude drafted both for me using the actual technical facts of the site (no cookies, no tracking, self-hosted fonts).6

Maintaining my website has become so easy — and it simply means opening Claude Code and typing something like change this text, add this link, swap this photo. A minute later, it’s live.

Final Things to Keep in Mind When Working With AI

You don’t need to understand every step, but you do need to stay nearby. The AI will stall at some moments, for example when you need to enter confirmation codes, logging in, or giving approvals. So, you can go vacuum the floor while it codes, but stay near your browser and check in from time to time!

Screenshots are your superpower. Whenever I was lost, I took a screenshot and pasted it into the chat. Claude read it and told me where to click next. Whenever I didn’t undersatnd Claude’s instructions I switched to ChatGPT or gemini and asked me to help me understand what’s not working and teach me what I need to know.

Wrong theories are part of it. Claude was confidently wrong more than once. So was I. Coding is problem solving — the debugging is the work, not a sign that something has failed.

And that’s it! I would love to see your examples in the comments and if you have any questions, let me know :)

Subscribe now

Read more

Subscribe now

A look throughout history shows that we have many times invented tools that helped us externalize our memory from writing to maps and calculators; and, similar to today, people have been worried about the consequences.

In the Phaedrus, Plato (speaking through Socrates) warned that the new technology of writing would weaken memory and offer only the illusion of wisdom.

He recounts an Egyptian myth in which the god Theuth presents writing to King Thamus, claiming it as a "recipe for memory and wisdom." Thamus counters that it will have the exact opposite effect:

“If men learn this, it will implant forgetfulness in their souls; they will cease to exercise memory because they rely on that which is written… What you have discovered is a recipe not for memory, but for reminder.”

Writing, he feared, would make people seem highly knowledgeable while knowing little, leaving them full of "the conceit of wisdom" rather than wisdom itself.

So was Plato right? Did writing weaken our memory?

Yes and no.

Plato was right that writing did reduce the need for individual memorization. Today, we rarely memorize long texts or genealogies "by heart" as pre-literate cultures did.

But he was wrong to treat that as a simple loss. Writing massively expanded our collective memory, freeing up cognitive resources for analysis, coordination, and abstraction.

Societies could now maintain records across centuries and build institutions without depending on flawless oral transmission.

On the societal level, this was the opposite of “forgetfulness.” By moving memory from the individual to the shared record, we democratized knowledge, shifting power from monarchs and clergy into the hands of the public.

This brings us to the present, and begs a critical question: What is the effect of AI on our collective memory? Will it deskill our society, or free up capacities that advance civilization?

To answer this, we cannot just look at AI as a technology itself; but also consider the present paradigm, in which it is built.

At first glance, AI appears to be doing what writing and the printing press did: making knowledge accessible to vastly more people. It can build analogies to previously inaccessible concepts, act as a personalized tutor, and help non-coders vibe code their ways into “artifacts” that would have previously taken an entire agency to build.

But there is an apparent difference between the printed page and a Large Language Model.

A book holds a record, preserving the voice, meaning-making, and context of its author. AI, contrarily, ingests our collective intelligence (in the form of “training data”, and synthesizes it into probabilistic output.

In doing so, it reveals the dominant paradigm of Western technology—one that treats knowledge less as a shared inheritance and more like a to-be-extracted resource.

This extraction changes the power dynamic. When we outsource our collective memory and reasoning to AI, we are not democratizing power as the printing press did; we are actively centralizing it.

We are taking the diverse, messy, polyphonic voices of human history and passing them through the narrow ideological bottlenecks of a few massive tech corporations.

Whose histories are deemed “quality training data,” and whose oral traditions, dialects, and marginalized experiences are filtered out?

In the realm of Big Data, the dominant hegemonic narrative is the statistical majority. Therefore, Indigenous oral traditions, decolonial perspectives, and marginalized ways of knowing are, by sheer mathematical definition, treated as “improbable” anomalies.

If we rely on this as our primary memory apparatus, we risk a new kind of digital colonialism—an epistemic flattening where rich, relational ways of knowing are overwritten by the statistically most probable algorithms.

This epistemic flattening directly answers the question of whether AI will deskill us.

If AI deskills us of the ability to write a rote corporate memo, that is no great loss. But because AI strips knowledge of its relational context, we are at grave risk of deskilling our relational intelligence and our moral imagination.

The very process of writing, thinking, and creating is about grappling with making sense and meaning. By offloading this, we risk becoming a society of editors rather than creators, skimming the surface of statistically probable syntheses without ever diving into the deep waters of independent thought.

And what of the capacities AI supposedly frees up?

The techno-optimist narrative promises that by offloading cognitive labor, we will be freed to pursue higher-level abstraction, art, and leisure. But here, extractive capitalism strikes twice. Just as the current paradigm extracts our collective memory to feed its models, it extracts our “freed-up time” to feed perpetual economic growth.

In How Much Is Enough?, economist Robert Skidelsky and philosopher Edward Skidelsky revisit John Maynard Keynes’s famous 1930 prediction that rising productivity would shrink the working week to fifteen hours.1 Nearing 2030, we can assess his predictions.

In 1929, the average American worked about 2,300 hours a year; today, that figure is roughly 1,800. While this is a significant decline, Keynes’s fifteen-hour week translates to about 750 hours a year. Not even Germany (my home country)—which works fewer hours (1,340) than any OECD country—has gotten halfway there.2

Meanwhile, productivity rose far beyond what Keynes imagined. But the time freed was not returned to humanity. Instead, it was reinvested into more output, more consumption, and more work. The time saved by technology was simply absorbed by an insatiable demand for extraction.

Questioning the paradigm we grew up with—this intrinsic belief that efficiency must inevitably serve extraction—is challenging. But what if we refused this cycle?

What if we used the bandwidth freed up by AI not to produce more information, but to re-skill ourselves in the profoundly human arts? The art of deep listening, of intergenerational care, of earth-stewardship.

What if we shifted our societal value toward a relational epistemology—cultivating wisdom that is rooted in place, community, and lived reality?

Plato was right to warn that technology could create the “conceit of wisdom.” AI amplifies this risk exponentially, offering us the illusion of omniscience without the anchoring weight of human experience.

But we are not trapped. We have the agency—and the collective responsibility—to choose how AI is integrated into our societies, and what forms of intelligence we choose to preserve.

Because the most profound wisdom is rarely the most statistically probable.

Subscribe now

Subscribe now

For eight years of schooling, I had a science teacher who regularly offered the same explanation for my struggles: “Girls are just not good at this.”

Lacking a sense of agency and a mental latticework on which critical thinking could be built upon, I accepted his claim as a given fact, and zoomed out. By the time I tried to pay attention again in 10th grade, I was lost.

Because learning depends on prior knowledge, I lacked the fundamental understanding required to grasp anything the new lessons were built on. I finished high school without having any clue about chemistry or physics.

A similar illusion of disinterest followed me into adulthood.

I labeled myself “unpolitical” until I was 25. The absence of any fruitful engagement with politics and history throughout my business studies in Frankfurt and Vienna reinforced my belief that I was just not interested in that.

I bought into the illusion that politics didn’t concern me. I told myself that nothing in my personal life was political, and even if it was, I couldn't change anything anyway. So, why even bother trying to understand it?

It took me until I was 28 to wrap my head around World War I and II, and the implications of the Holocaust—thanks entirely to coincidental enjoyment of historical fiction and stumbling upon Viktor Frankl and Edith Eger by what must have been a self-help must-read book list by some manager on my LinkedIn feed.

From there, I found my way to Octavia Butler, bell hooks, and Audre Lorde, and my convenient little illusion completely shattered. I realized that everything is political, and that “not being interested” is, rather than a smart neutral choice, a very active choice to maintain the status quo, even when that status quo isn’t serving most people.

I could tell you similar stories of how I ignored arts, culture, literature, and poetry, but I will spare you.

Why am I telling you this?

To surface an interesting truth about lifelong learning: We often dismiss things as “uninteresting” when, in reality, we simply lack the foundational knowledge to be curious about them.

The Illusion of Disinterest

Author Maria Popova finds ever so beautiful words to this:

“I am a latecomer to poetry. For many years I dismissed it the way we dismiss things in which we’re not literate. I have no education in poetry. And then I met this wonderful person—Emily Levine, comedian, philosopher of science, great lover of poetry, […] she was in her 70s, I was in my 20s.”

Popova had met Levine—a comedian and philosopher of science—on a transatlantic flight, and much to the annoyance of the cabin, they talked the entire way. Levine had a brilliantly associative mind, leaping effortlessly from Hannah Arendt to stoicism.

Levine loved poetry. Popova, on the other hand, did not.

One Sunday afternoon in a packed cafe in Chelsea, Popova rolled her eyes at the mention of poetry one too many times. Levine gripped the edge of the table, rose to her full height of four-foot-four, and began theatrically reciting T.S. Eliot’s The Love Song of J. Alfred Prufrock.

When she delivered the line, “Do I dare to disturb the universe?” she sat back down. The New York cafe stopped, put down their drinks, and started applauding.

“Of course,” Popova recalled, “my universe was disturbed.”

From that day on, Levine took it upon herself to build a bridge for Popova. She sent her one poem a day, accompanied by just a few lines explaining why she loved it. Even when Levine fell terminally ill, they continued taking weekend “poetry retreats” in Northern California cabins, just reading poems and talking about the universe.

Because of Levine’s thoughtful scaffolding, Popova learned to attach poetry to a mental model she already deeply understood and loved: science.

Popova went from proudly “not caring” about poetry to hosting a massively popular show called The Universe in Verse, gathering thousands of people to listen to Patty Smith read poetry about dark matter.

It begs the question: How do you know whether you’re actually not interested in something, or whether you’re just illiterate in it?

Why You Think You “Aren’t a Math Person”

This illusion of disinterest is why so many people think they “aren’t a math person.”

In a geography class, you can understand how a volcano erupts without needing a deep, factual understanding of tectonic plates. But math doesn’t work that way. It requires previous rules and foundational understanding to build upon.

If you don’t really grasp that 304 means 3 hundreds, 0 tens, and 4 ones, then decimals like 3.04 or percentages like 30.4% will feel totally arbitrary.

If you don’t understand a basic slope (how much y changes when x changes), you’ll struggle with derivatives. A derivative takes the idea of a slope one step further to ask: how steep is this curve at this exact point? Without the earlier concept of a slope, a derivative feels like a mysterious, random procedure. With it, the derivative becomes a natural extension—not a random rule, but “slope, upgraded.”

The same goes for functions. If f(x) feels like abstract code rather than a simple relationship between an input and an output, then f’(x) will feel like symbol soup. A derivative describes how a function changes.

The Power and Anatomy of a Perfect Analogy

So, how do you fix a missing foundation? You have to immerse yourself and build an initial understanding by attaching new concepts to existing mental models.

The easiest way to do this is through analogies.

By bridging what you don’t know to what you already know, analogies act as “handles” to grasp slippery concepts. They let your brain use its natural association system instead of forcing it to memorize raw facts, making new material feel concrete, meaningful, and easier to remember. That is why analogies create those sudden “lightbulb moments.”

A great analogy…

• is a cognitive bridge. It should connect the unfamiliar idea to something you already understand deeply, without needing a second explanation.

• is selected for structural fit. The best analogy mirrors the logic of the concept. Water and pipes may not be exciting, but they map reasonably well onto some features of electricity.

• reduces cognitive load. It should make the target idea easier to think about, not add extra information, lore, or emotional noise.

• is “less wrong,” not perfect. All analogies distort reality. The question is whether the distortion is useful at your current stage of learning.

• should eventually disappear. It is a scaffold. Once you grasp the concept, you should move beyond the analogy and understand the thing itself.

Popova was very lucky. She had Emily Levine to gently scaffold a difficult subject for her.

Most of us don’t have a brilliant, four-foot-four philosopher waiting in a cafe to disturb our universe. We don’t have someone who knows our minds well enough to build the perfect cognitive bridge for us. We hit a wall, we assume we just aren’t wired for the subject, and we quit.

But if you don’t have an Emily Levine, you can build the bridge yourself.

Coming up with structural, perfectly-fitted analogies from scratch is difficult, but AI is uniquely brilliant at it—if you constrain it. If you just ask ChatGPT for an analogy, you usually get generic, surface-level comparisons.

Here is the prompt I use to be my own Emily Levine. I keep this saved in my notes so I can use it whenever I hit a plateau with a new, intimidating concept.

How to Be Your Own Emily Levine (The AI Prompt)

Read more

Subscribe now

When I started writing and publishing online, it felt impossible to improve.

I was putting in the hours. I showed up consistently. But my intros were weak, my arguments felt scattered, and my articles routinely disappeared into the void.

I assumed the answer was simply to write more. I thought if I just relied on endurance and pushed through enough words, I’d eventually find my way.

But more effort just meant reproducing the same patterns. I was practicing blindly, repeating errors without reflection, and adopting others’ strategies without holding space to see if they actually aligned with my practice.

I was caught in a cognitive loop, mistaking the sheer momentum of output for the rewiring required to evolve.

The most impactful learners don’t just grind more hours—they build systems to examine how they learn. They track what feels hard, unlearn what isn’t serving them, and engineer loops for the practices that do, using metacognition.

Here is how I used it to find my way through my own plateau, and how you can adapt it into a grounded, repeatable rhythm for improving many skills, especially when you can define goals, get feedback, and track progress.

The Problem with Simply Pushing Through

Beyond the hours you log, learning new skills is about the quality of your feedback loops.

You can practice a skill for years and never see real change if your process looks like this:

• Practicing blindly: Showing up with no clear intention and nothing to guide your progress.

• Repeating patterns: Finishing a session without pausing to make sense of what worked well and what didn’t.

• The autopilot trap: Relying on the same approach for every challenge, regardless of whether it’s actually working.

Metacognition—the capacity to witness and regulate your own thinking—is the antidote to the autopilot trap.

A substantial body of educational research shows that metacognitive strategies can improve learning, especially when learners are taught to plan, monitor, evaluate, and adjust their approach in a specific task or subject.

By regularly holding space to reflect on your own thinking, you become more intuitive. You learn to identify gaps in your understanding and course-correct before exhaustion sets in.

Instead of passively absorbing information, metacognition helps you take gentle, deliberate control.

The 4-Step Metacognition Framework

After diving deep into the science of learning, I realized I needed a different structure. I stopped treating writing like an output to be maximized, and started tending to my process.

I built a “Metacognition Playbook.” Before each session, I pick a concrete target. Afterward, I reflect on what worked, what failed, and what to change. This cycle brings immediate clarity to your learning.

Here is the 4-step loop you can use for any practice—from mastering a new language to sharpening your decision-making.

1. ASSESS (Before Practice)

Don’t just rush into the work. Take a minute before you start to identify what you are intending to practice.

• What’s your intention or learning goal for today’s session? (Make it specific. Not “write better,” but “write a clear, inviting opening hook.”)

• Which approach will you apply to meet this intention? (e.g., “I will study three hooks from writers I admire and outline mine using their framework.”)

2. PRACTICE + MONITOR

Do your practice session based on the intention you just set, but pause briefly when you get stuck to ask: Is my current strategy working, or do I need to adjust?

3. REFLECT (After Practice)

This takes just two minutes, and it becomes even more helpful the more often you do this. Before you close your laptop or pack up your space, ask yourself:

• How was your session today? How do you feel about it?

• What did you learn about your process? What flowed well? What felt resistant or disconnected?

• What will you nurture or change for next time? Which strategies will you tweak? What’s your most important practice for the next session?

4. REPEAT

Take the wisdom from your reflection and feed it directly into your next “Assess” phase.

How This Looks in Reality

At first, tracking my writing process felt tedious. I just wanted to write more.

But soon, I began to sense patterns. During my “Reflect” phases, I realized my most resonant articles all shared a specific type of opening—so I began holding more space to experiment with hooks.

I also noticed that whenever I encountered the discomfort of a difficult thought, I’d unconsciously open LinkedIn to escape the friction. Because I was witnessing this habit, I installed a website blocker before my next session to help me focus.

Over time, I noticed my process becoming more deliberate, and my work began to resonate more with readers. While many factors contributed, this feedback loop helped me see and adjust patterns I had previously been blind to. I now use it for any new skill I want to build.

The AI Learning Thought-Partner

If you run a metacognition log for a few weeks, you will gather incredibly rich data about your own cognitive habits. You can wield AI as a complement here—acting as a learning scientist to help you make sense of this data.

Copy and paste this exact prompt into ChatGPT or Claude, attaching your reflection notes:

You are a learning scientist and expert in the field of [YOUR SKILL]. Your expertise lies in helping people identify their blind spots, align their learning strategies, and cultivate deep skill acquisition.

Attached is my metacognition log, which I’ve maintained for the past month. It tracks my learning intentions, strategies, reflections, and adjustments.

Analyze my log and help me make sense of my deepest blind spots—the patterns, habits, or disconnects that are currently holding me back. Then, outline a personalized, three-week rhythm to help me grow. Focus on actionable insights, gentle habit shifts, and meaningful learning practices. Keep it concise, ensuring I can weave your advice into my routine immediately.

Use AI as a pattern-spotting assistant, not an authority. Treat its conclusions as hypotheses to test in your next practice session.

The Science of Metacognition

Metacognition is the ability to be aware of and regulate your own thinking. Studies show that learners who consciously choose and adjust their approach see significant improvements in both academic and skill-based environments.

Here is how the research shows metacognition supports deep learning:

• Enhanced Learning Strategies: Instead of using the same practice method for every context, metacognition helps you match the right approach to the task, leading to deeper comprehension.

• Improved Academic Performance: Studies show that metacognitive reflection boosts performance in math, science, and memory-related tasks. Learners who deliberately plan, monitor, and evaluate their strategies often perform better, though results vary by context and implementation.

• Self-Regulated Learning: Metacognition helps you step into your own agency. Rather than passively receiving information, you plan, track, and adjust your habits—leading to a more sustainable mastery.

• Higher Motivation and Engagement: Metacognitive instruction can support motivation in some contexts, partly because learners gain clearer strategies and a stronger sense of control, but motivation is also shaped by interest, confidence, environment, task value, and feedback.

• Stronger Problem-Solving: By regularly holding space to assess their own thinking, metacognitive learners become better at navigating complexity, identifying gaps in their knowledge, and course-correcting before feeling overwhelmed.

In Conclusion

If you take one thing from this piece, make it this: stop measuring your progress by the sheer volume of hours you endure. Start measuring it by the tightness of your feedback loops.

Deep growth doesn’t come from pushing harder against the void. It comes from the willingness to pause, look at your own process, and actively course-correct. When we shift our focus from mindless output to deliberate reflection, learning stops being a frustrating sprint. It becomes a structured, resilient system that guarantees you actually get better at what you do.

Sources

Biggs, J. (1988). The role of metacognition in enhancing learning. Australian Journal of Education, 32(2), 127–138.

Fleur, D. S., Bredeweg, B., & van den Bos, W. (2021). Metacognition: Ideas and insights from neuro- and educational sciences. npj Science of Learning, 6, Article 13.

Palincsar, A. S., & Brown, D. A. (1987). Enhancing instructional time through attention to metacognition. Journal of Learning Disabilities, 20(2), 66–75.

Perry, J., Lundie, D., & Golder, G. (2019). Metacognition in schools: What does the literature suggest about the effectiveness of teaching metacognition in schools? Educational Review, 71(4), 483–500.

Stanton, J. D., Sebesta, A. J., & Dunlosky, J. (2021). Fostering metacognition to support student learning and performance. CBE—Life Sciences Education, 20(2), Article fe3.

Veenman, M. V. J., Van Hout-Wolters, B. H. A. M., & Afflerbach, P. (2006). Metacognition and learning: Conceptual and methodological considerations. Metacognition and Learning, 1(1), 3–14.

Zepeda, C. D., Richey, J. E., Ronevich, P., & Nokes-Malach, T. J. (2015). Direct instruction of metacognition benefits adolescent science learning, transfer, and motivation: An in vivo study. Journal of Educational Psychology, 107(4), 954–970.

Subscribe now

On most mornings around 2017, you would have found me rising at 6, meditating for 15 minutes, journaling about gratitude (thinking, not feeling), reading a nonfiction book, and taking a cold shower afterward because someone on a Tim Ferriss podcast mentioned you might become a CEO if you did that.

I was living a driven, growth-minded (and to more reflective people at the time, also a blindly misguided and compliant) life.

Like many business bubble grad students at the time, I was trapped on the treadmill of self-improvement and efficiency, buying into the illusion that if I just did more, I would eventually reach a point where I finally felt a sense of enoughness and satisfying accomplishment.

At the time I was trapped in a cycle I would only later realize was deeply counterproductive for my clarity of thought — most of my waking hours were filled with (intentional) consumption — listening to a podcast during a run and squeezing a phone call with a friend into the lunch break, the only free hour between back-to-back meetings.

Constant Stimulation Deprives the Brain of the Empty Space It Needs to Digest and Connect

Meanwhile, a healthy, clear mind needs non-doing time, also outside of sleep. It is built to connect the dots. Only after some years of grinding did I realize what I wish I had understood earlier — shower epiphanies can also arise outside the shower if you make time for unguided spaciousness, contemplative practices, or just being.

Your mind wants to connect the dots. It wants to understand why. But it needs to be left alone long enough to actually do its work. By stripping stimulation-free time out of our days, we are unintentionally depriving ourselves of moments where memory, insight, and attention can form into something coherent.

This is not an argument against ambition, books, podcasts, or AI. It is an argument against a life with no digestion.

Modern life is an attention-extraction machine. Apps, ads, and algorithms are carefully engineered to ensure no minute stays empty. How to get more done. How to be more efficient. How to learn faster.

And now, AI is pouring fuel on the fire—generating more drafts, more summaries, and more input, faster than ever—feeding the exact obsession wearing us down.

So what happens to a brain that never gets an empty minute?

In early studies, researchers expected brain changes when someone completed a task—reading a word, moving a finger, solving a puzzle. But neuroscientist Marcus Raichle and his colleagues noticed something else. Specific brain regions showed relatively higher activity during the presumed baseline condition: when people lay quietly, relaxed, or simply fixated on a point.

As participants shifted into attention-demanding tasks, activity in those regions reliably decreased.

The resting brain did not sit idle. It maintained an organized background life: what Raichle called the brain’s default mode.

The name fails. “Default” implies passivity, like factory settings or a screensaver. It acts otherwise. The default mode network links distributed brain regions involved in inward thought: remembering past events, imagining future possibilities, thinking about yourself and others, and helping experience become meaningful.

Mary Helen Immordino-Yang and her colleagues coined the phrase “constructive internal reflection.”

Constant stimulation does not destroy this inward life, but it does crowd out the quiet conditions it seems to need.

Quiet Gaps Are Biologically Required for Memory, Insight, and Recovery

What I used to consider a fully “optimized” day is far from optimal.

Researcher Michaela Dewar and her colleagues tested this with a simple experiment. Participants learned two pieces of verbal material. Ten minutes of wakeful rest followed the first piece. Ten minutes of a simple spot-the-difference game followed the second.

The effects lasted long past the pause.

Participants remembered the material followed by rest better—15 to 30 minutes later, and even seven days later. They were not studying extra hours. The difference was what happened immediately after learning: quiet rest or another task.

So what does this mean?

Quiet time is helpful for learning gains. But not only that — those quiet gaps can also help insight arise.

Studies of creative incubation suggest stepping away from a problem can improve your chances of returning with a better idea. The sweet spot is often not intense focus or total blankness, but an undemanding task—something simple enough to let the mind wander.

A walk is more likely to create this condition than a phone break. Walking gives the mind just enough to hold on to without taking it over, whereas a phone can often stir up new to-dos and mental processes. Moving leaves, rushing water, shifting clouds—these elements invite attention, giving your effortful focus a chance to replenish.

(An honest caveat: wandering minds sometimes stumble. A mind left alone connects ideas, but it also rehearses shame, worries about the future, and replays arguments. The point is not to force yourself to sit in a dark room thinking profound thoughts, but to stop consuming new input to allow existing thoughts to move and reshuffle.)

A practical point emerges:

Your brain needs intervals free of new information.

After learning, those intervals can help memory stabilize. After effort, especially in low-demand or natural environments, they can help attention recover. After a problem, they can give insight a chance to surface. After an emotional experience, they may create room for meaning to form.

We treat empty moments as wasted time. Cognitively, they hold deep value. The mind digests during these gaps.

I hold a personal sleep hunch, and I want to name it as a hunch, not a finding. I suspect my own 2 a.m. racing mind sometimes signals a backlog. When I give it zero quiet daylight to sort, connect, and feel, it tries to do the work at night.

Reintroducing Information-Free Intervals Requires Pushing Through an Initial Stimulation Hangover

Give gaps back to your mind. Here’s how to make it work:

Read more

For this week’s edition of the Lifelong Learning Club I interviewed Sam Illingworth from Slow AI. Sam is a Full Professor of Critical AI Literacy based in Edinburgh and the author of GenAI in Higher Education (Bloomsbury, 2026).

I appreciate how he shifts our focus from how to best use AI to the interiority of the human using it; reminding us that behind every algorithmic output are the hidden costs of human labor, historical biases, and a potential erosion of our own voice. Over to Sam.

Subscribe now

1. Sam, how is Critical AI Literacy different from AI literacy?

AI literacy, as it is commonly taught, focuses on competence. Write a better prompt. Pick the right tool. Integrate AI into your workflow. The assumption is that AI is a tool, and the key question is how well you use it.

Critical AI literacy asks a different question. Should you be writing the prompt at all?

The difference between the two is the difference between technical skill and human judgment. AI literacy teaches you how to use AI as a tool. Critical AI literacy teaches you to notice how AI as a tool is changing you while you use it.

When teaching AI literacy stops at the level of competence, we might miss out on three things.

First, AI is never neutral. AI models reflect the biases in their training data and the choices made by the organizations that build them. Here is an example I can’t stop thinking about: working-class Victorian newspapers were printed on cheap paper with bad ink. Because of that, computer scanners struggle to read them. Consequently, these newspapers aren’t part of the datasets used to train modern AI models. Without us realizing it, the tight budgets of 19th-century printers are shaping what modern AI treats as history.

Second, AI changes you. Lean on AI often enough, and you start to outsource your judgment. You lose confidence in your own voice and adopt AI slop without noticing. It’s easy to celebrate getting an answer in three seconds. But when you stop wrestling with a paragraph, the thinking that used to happen in the background simply dies.

Third—and this can be the trickiest to grasp—is knowing when to leave AI alone. Knowing when not to use AI is harder, and matters more, than knowing how to use it. The AI industry is built on the premise of endless integration, but critical AI literacy gives you the permission and the framework to simply leave the machine alone.

So my working definition is this. Critical AI literacy is the ability to evaluate outputs for bias and error, to recognise how the tool shapes your thinking over time, to judge when it helps and when it harms, to understand the human and ecological costs behind it, and to make deliberate choices about when to refuse it.

This approach is anti-thoughtlessness. The aim is to use AI with your eyes open.

Before using AI,, a critically literate person pauses to ask:

1. What am I giving up by not doing this myself? (Will I lose the thinking or creating process?)

2. Whose perspective is missing from this output? (Knowing AI represents only a fraction of the world.)

3. What would I lose if this output were wrong and I did not notice? (Recognizing that AI hallucinations can sound confident and fluent.)

4. Who benefits from me using this tool right now? (Remembering the corporate incentives at play.)

5. Would I trust this output if I did not know it came from AI? (Checking if you trust the actual content, or just the technology.)

2. Why does this matter, particularly in education?

Because educators are being told to ‘integrate AI’ with no framework for working out what it helps and what it quietly undermines.

Take AI detection tools, where we can already measure the harm. AI detection tools are wrong up to 30% of the time, and they disproportionately flag non-native English speakers. Students who wrote every word themselves have been accused of cheating. The detection industry is worth millions and its products do not work.

An example for this is the word ‘delve’ which has become a popular shorthand for catching AI slop. But what most people don’t realize is that ‘delve’ is common in Nigerian business English. Because many of the data labelers who trained these AI models were based in Nigeria and Kenya—often paid less than two dollars an hour—underpaid workers in the Global South effectively shaped the model’s language. So, an AI detector that flags ‘delve’ isn’t necessarily catching a machine. It is penalizing the linguistic patterns of marginalized people.

This is where I think a lot of people get wrong with regards to AI slop. For me, AI slop is not really about vocabulary or em dashes. There was slop long before the advent of AI, and it is defined by people not taking any pride or agency in their work. There is nothing wrong with using AI to help interrogate an argument, rewrite a sentence, or provided additional context. The majority of students that I work with know this, and they are very aware of the need (and desire) to keep humans in the loop.

Yet, universities are writing AI policies that open with ‘How do we detect cheating?’ instead of asking, ‘How do we design assignments worth doing?’ This detection arms race is a complete distraction. Critical AI literacy asks the question that a detection tool never will: What are we actually trying to teach, and does AI help or hinder that?

Beyond the distraction of detection, there is an even bigger risk for students who are still forming their judgment. Researchers name this: ‘never-skilling.’ A trainee who leans on AI early and completely may never build the underlying judgement at all. The scaffolding goes up before anything is built behind it.

That is the stake in education. We are at risk of producing a generation fluent in operating tools, but entirely unpracticed in the thinking those tools were meant to support. Critical AI literacy is how we ensure we are designing learning that is worth doing—whether AI exists or not.

3. What does “deskilling” mean in the context of AI, and what’s one way to mitigate it?

Deskilling is losing a capacity you used to have because a tool now does it for you. This is related to what I’ve mentioned in the previous question: never-skilling, where you lean on AI so early that the capacity never forms in the first place.

For example, in a 2025 study, experienced doctors who had grown used to AI help spotting pre-cancerous growths during colonoscopies were tested again without it. Their unassisted detection rate fell from 28% to 22%. Seasoned specialists, who had the skill, leaned on the tool, and quietly lost some of it. It took months, not decades.

However, this is not as simple as having the ‘willpower’ to resist AI. Indeed, blaming yourself for taking that path is like blaming yourself for breathing the air in a room someone else filled with smoke. We have seen this move before, with the personal carbon footprint, a framing popularised by an oil company. Guilt keeps you looking inward at your own character instead of up at who built the conditions.

So the one practical thing I would offer is this: before you turn to an AI tool, get a rough draft written down yourself. This means that you continue to develop your skills and it also results in far more creative solution, as AI by its very nature always regresses to the mean.

4. Any final thoughts to keep in mind when working with the current AI models?

One, mostly. Stop reading confidence as a signal of correctness.

The newer and more capable these models get, the more assured they sound, and the assurance is climbing faster than the accuracy. As frontier models reason for longer, their answers can grow more incoherent while their confidence keeps rising. This means that on the difficult questions, the ones you most need help with, the model is most likely to be fluently and convincingly wrong.

So treat fluency as a warning to slow down. The formatting will be perfect. The tone will be certain. The citation will look real. None of that tells you whether the thing is true. The tone of certainty reads the same whether the model is right or whether it has invented the answer whole.

The second thing to always remember is who the model speaks for. I am, by most measures, the person these systems were built to reflect: white, Western, male at the comfortable centre of the data they trained on. When AI produces the ‘neutral’ professional voice, or the ‘average’ expert, it is producing a version of me. Before you trust an output, ask who the assumed default user is, and whose perspective never made it into the training data.

I am not advocating for the abolition of AI. I use them every day and I am not a doomer. It means meeting them with your attention switched on. The model is a fast, fluent, confident first draft of a thought. The judgement about whether that thought is any good has to stay with you.

Subscribe now

Subscribe now

Which creatives inspire you? Who is someone that fascinates you in the way they think, work, or create?

When I find a mind inspiring, my instinct is to better understand their process. Very often, I find that their creativity is rooted in unexpectedness; the ability to connect things that previously seemed very distant.

Sometimes this manifests in category-defining artists like Sia, who can come up with eight vocal tracks to build unexpected harmonies; or Major Lazer, the punk of electronic music with some of the most entertaining EDM performances.

Other times I am obsessed with creative intellectuals like Maria Popova, who I perceive as one of the most brilliant minds alive. She braids literature, science, philosophy, poetry, and art into unexpected connections, stewarding us towards what it means to be human.1

All of these creatives embody interconnectedness. They don’t confine themselves to a single domain, but allow disparate disciplines to collide and influence one another.

Attempting to define this deeply human trait gives us a hint as to why generative AI is currently failing at collective creativity.

With all the ongoing noise about artificial intelligence replacing artists and knowledge workers, I wanted to understand whether AI can replicate this type of interconnected thinking; because my hunch is that it will, instead, flatten it out.

So, in this article, we will look at how science deconstructs originality and creativity, what we understand about generative AI in relation to that, and the implications for working and learning with AI (without flattening your own creativity).

The Science of the Creativity

Creativity is novelty and value. Mark Runco and Garrett Jaeger, who have coined the current definition of creativtiy argue how novelty on its own isn’t enough: “Originality can be found in the word salad of a psychotic and can be produced by monkeys on word processors.” It only becomes creativity when the surprise is also good for something.

Back in 1962, psychologist Sarnoff Mednick argued that creative ideas come from combining associative elements — and the more remote those elements are, the more creative the result. He thought creative people have what he called a flat associative hierarchy: where most of us reach for the two or three obvious associations and stop, they keep going, pulling up stranger, more distant ones.

Modern network science has since put numbers on his hunch. Yoed Kenett maps semantic memory as an actual web of connected concepts, and finds that highly creative people have a flatter, more interconnected one — shorter paths between ideas the rest of us file far apart, like hammer and tissue. Their memory is wired for the leap. This is literally the interconnectednes I so much admire in Maria Popova.

When looking what happens at the brain, Roger Beaty and colleagues find creativity runs on a duet between the default-mode network, which freewheels and generates, and the executive-control network, which evaluates and refines, with a third network flipping between them. Recent neuroscience adds causal evidence to this finding. In a 2024 Brain study, direct cortical stimulation of default-mode network regions reduced the originality of responses in an Alternative Uses Task, supporting a role for the Default Mode Network in divergent thinking.

Originality is, quite literally, the ability to connect semantically distant things.

Creative Scars Caused by AI

What I find interesting is that Generative AI is built to do the opposite.

Large Language Models are not “artificial intelligence”—they are, in ’s words “artificial predictions”; statistical pattern matching, a process of finding regularities in vast quantities of training data and reproducing them, statistically, with no understanding of what any of it means.

In other words, Large Language Models are probability engines designed to predict the next most likely word. They are a regression to the mean of human preference. They are center-seeking tools in a domain (creativity) that requires exploring the outer edges. They tend to pick legible, high-probability, broadly preferred continuations.

So what happens to human creativity when we rely on pattern matching tools?

A 2024 study published in Science Advances ran a controlled trial with hundreds of writers. On average, AI access improved judged novelty, usefulness, and some emotional qualities of the stories, with the largest gains among writers who scored lower on the study’s baseline creativity measure. Their stories were rated up to 26% better written and 22% more enjoyable. The writers with lower baseline creativity scores, saw the biggest gains.

But, and this I find most interesting, when researchers zoomed out, they found a devastating drop in collective diversity. Stories written with even a single AI-generated idea were 10.7% more similar to each other than the stories written by humans alone.

This does not appear to be merely a ChatGPT quirk. 2026 study of 22 public-API LLMs found that, across three standardized divergent-thinking tasks, LLM responses were much more similar to one another than human responses were. The finding suggests a broader tendency toward homogeneity in current aligned LLM systems.

AI raises the floor, making sure none of us write a terrible story or a bad email. But it lowers the ceiling, with creativity suddenly converging towards similarity.

And on the way, we might unlearn and forget parts of what made us humans previously more creative. A 2026 Technology in Society article by Zhou et al. reports what the authors call a “creative scar”: in their abstract, they state that creativity gains dropped after AI assistance was withdrawn, while induced content homogeneity continued rising months later.

How to be Create with AI (without flattening your creativity)

As demonstrated, there is a fundamental difference between how human brains construct creativity and how AI attempts to replicate it. You cannot outsource remote association to probabilistic pattern recognition.

But we also shouldn’t ignore the potential of LLMs. If we want to work with AI without flattening our minds, we have to be intentional how we interact with it. Based on early research into AI and creativity—alongside my own practice—here are two steps to do that.

Read more

Subscribe now

The first time I consciously learned something by heart, I was ten years old.

It was the Glaubensbekenntnis, the Catholic Creed. A dense, one-page document rooted in what I now see as a patriarchal version of faith. I spent days staring at it, and gave it my best. I was graded for it, and I desperately wanted to be a good student.

I learned to recite it perfectly. I still can.

But mostly, I remember feeling confused. Why am I learning this? Nobody could give me an answer beyond: “Religion is important.”

Soon after, memorization became a form of punishment. At age 13, when I chatted with a friend during class, my penalty was memorizing a page of the Bible. In German class, I was forced to memorize poems I didn’t even like.

Sure, some of the memorizing parts during school felt useful—the multiplication table, English and French vocabulary. But those were the minority, and they were rarely tied to my own motivation.

My entire perception of memorization was defined by a lack of agency.

And I am not alone. Talk to people who grew up in a Western school system, and you will hear a variation of the same story. Years of standardized, meaningless grinding condition us to leave school believing that memorizing facts is an obsolete, tedious chore.

I used to believe that, too.

Now, I know better.

We assumed that because memorizing someone else’s agenda felt useless, the act of memorization itself must be useless. This is a harmful confusion. Forced memorization is a prison. But chosen memory? That is magnificent, once you understand what it can do for your mind.

When Memory Becomes Mind Control

Forced memorization is pointless. At its worst, it is destructive indoctrination. It took me decades to find my way back to prayer as a devotion to life—to unlearn the narrow dogma I was tasked to commit to memory at age ten.

My childhood experience was stifling, but ultimately harmless. Yet history is full of chilling examples of what happens when this exact mechanism is used with lethal intent.

Strip a person of their agency, force-feed them a harmful ideology, and memory ceases to be a tool for learning, and instead, becomes a weapon of mind control.

In the Hitler Youth, rote memorization was designed to lead to blind obedience. Children had to memorize the Führereid (the oath of allegiance to Hitler), chanted ideological marching songs, and repeated antisemitic propaganda. In that sense memorization can became a weapon, as when you implant a state’s agenda so deeply into the youth’s subconscious, ideology can feel very much like their own inner voice (and identity).

This is why we must be highly critical of the subjects we commit to memory. When someone ill-intended chooses them, it can be a form of control. But flip it to the opposite so you choose what you memorize and it becomes a tool for something else.

Memorization as an Act of Liberation

Internalized memory expands the options you perceive, the analogies you draw, and the decisions you make. When you deliberately choose what to keep in your long-term memory, you gain four superpowers:

1. You expand your freedom to choose. The options you perceive in life depend entirely on the knowledge you hold. A choice you do not remember exists is not a choice you can make.

2. You build a mind ready to think critically. Committing historical context, mental models, and fundamental facts to memory is the prerequisite for critical thinking. When facts are internalized, you can spot logical fallacies, recognize gaslighting, and see straight through propaganda.

3. You fuel spontaneous creativity. When you choose to memorize insights from diverse fields—psychology, biology, art—you feed your brain the raw materials required to synthesize new ideas. Whether you are in a high-stakes meeting, on a quiet walk, or deep in conversation, your brain connects the dots. Why? Because the dots are already inside your mind.

4. You furnish your inner world. Memory makes life richer. When you memorize things that resonate with your soul—a favorite poem, a stoic philosophy, a reframed prayer—you literally decorate the inside of your mind. Recently, I began memorizing Ayurvedic and yogic frameworks regarding the mind and well-being. Having this knowledge anchored in my memory helps me catch myself when I slip into reactive states, and shows me exactly how to move through them.

Reclaiming Your Internal Library

Memorization is the foundation of thinking and its time to shatter our collective misunderstanding of it being an outdated technique.

Especially with AI at our fingertips, and everyone having access to the same externalized information, building a rich, internalized library of your own knowledge matters more now than before.

You don’t have to return to the grueling, rote repetition of primary school to build it. Learning science has evolved. Through evidence-backed techniques like active recall and spaced repetition, memory is no longer a chore.

It is a choice.

(And you can start here if you want).

Subscribe now

References

• Brown, P. C., Roediger, H. L., & McDaniel, M. A. (2014). Make It Stick: The Science of Successful Learning. Belknap Press of Harvard University Press. (My favorite book for the underlying science of spaced repetition and retrieval practice).

• Pine, L. (2010). Education in Nazi Germany. Berg Publishers. (For historical context regarding the educational and indoctrination tactics of the Hitler Youth).

Subscribe now

It’s 8:03 a.m. on a Monday morning two weeks ago, and I’m standing inside a small citizen service office in Lower Austria with my passport, my phone, and the surreal feeling that ChatGPT has just solved a bureaucratic problem I couldn’t solve on my own.

I had been abroad for six months and had flown back to Austria to renew my government online verification. In theory, this should have been simple. In practice, the first available appointment in Vienna for non-Austrian passport holders was in November.

I’m a researcher, so I did what researchers do: I searched for workarounds. But all I found were articles about the bureaucratic burden and patience required of non-Austrians living in Austria.

Then, as a last resort, I asked ChatGPT’s deep research mode to solve the problem for me.

After 21 minutes of searching, it found a loophole I had missed in the form of Monday morning walk-in hours for foreigners at a small citizen service station half an hour outside Vienna.

So I went. And at 8:05 a.m. I walked out of the building with my verification done, and a clear feeling that AI is magical.

And yet, just a few hours later on the same day, thinking with AI felt less like magic and more like a slow and tedious poison to my cognition.

In this article, I’ll unpack the subtle toll AI takes on our cognition, and share a clear set of guidelines to harness the magic without losing agency, joy, and stewardship in the work.

The Cognitive Dip When Thinking with AI

Here is what I noticed happening to my nervous system and my cognition after months of thinking and learning with AI.

• Desire for speed. I get impatient with the time it takes me to form a thought and I constantly wonder whether there is a faster way to do this, a better claude code workflow, or some prompt that could eliminate cognitive friction. I want first drafts instantly, similarly to how I would crave Nutella as a child, skipping the deeply uncomfortable, necessary phase of not knowing what I think.

• Expectations towards higher quality. What I used to be satisfied with no longer satisfies my ambitions as there are so many ways to use AI to raise the quality ceiling. I deprived myself from the satisfaction I used to get after working effortfully and deeply.

• Scattered attention. After a session of “multi-Clauding”—switching rapidly between AI workstreams—I don’t feel the grounding exhaustion of deep work. I feel a jittery, reactive fatigue. It’s like the drained feeling after running errands in a crowded, noisy city, rather than the replenishing tiredness after a long run in the woods.

• Less enjoyment. I produce more, but I feel less, missing the texture of wrestling with concepts and uncovering my own deep-felt knowledge.

• Longer cognitive recovery. My cognition speeds up. After prolonged sessions working with AI, I feel a high mental strain and as if my brain continues to run interval sprints while I would better slow down and recover. I struggle to find my way back into my body. I need grounding and co-regulation just to restore a discerning, present intellect.

• Giving away a part of my agency. Because I don’t fully recover from the depleting cognitive sessions, I feel a nagging sense of surrender.

Overall I have this nagging sense that we are immersing our minds into new cognitive environments at the speed the technology gets developed (from a fear of missing out) with defaults set largely by companies whose intentions are not aligned with our individual and collective wellbeing.

It feels as if people in companies driving AI development have invented a new high-intensity sport for the mind: cognitive CrossFit, but without the coaches, warm-ups, and recovery protocols.

With physical fitness, we eventually learned that high intensity training is double-edged. Yes, you can build capacity, but you can also injure yourself really badly should you train too hard, too fast, and pay too little attention to form and recovery.

With AI, the equivalent damage is harder to see. An injured mind does not show visible bruises. You can’t look in the mirror and notice a pulled attention span. You can’t go to the doctor and scan for a tear in your sense of agency. The indicators are mostly introspective — with brain fog, impatience, jittery fatigue, a diminished pull towards the slow work of forming your own thought.

And this is what troubles me. By the time we can collectively name what has been strained—attention, judgment, patience, taste, agency—we may already have built our work to rely on AI.

The Emerging Vocabulary of Our Minds on AI

Over the past year, I’ve started to pay attention to vocabulary for what AI might be doing to our minds. Some of these terms are well-established, some are early research concepts, and some are useful metaphors. But together, they help me understand through potential cognitive injuries.

• Cognitive offloading: Reducing mental load by using external tools. This isn’t inherently bad, but it forces us to ask: what are we outsourcing, and what are we keeping alive through practice?

• Cognitive surrender: Abandoning critical reasoning because we trust the machine’s confident voice, subtly overriding our own intuition.

• Cognitive atrophy: The slow loss of our own cognitive capacity because we removed the effortful moments before the skills could fully form.

• Cognitive debt: The short-term productivity high that quietly accumulates long-term costs in critical thinking, memory, and independent judgment.

• Cognitive stunting. Skills that never form because the effortful moments got removed before they could develop. (I wrote about this more here).

The Guidelines I Follow When Thinking With AI

Understanding these concepts gave me a new guiding principle: I use AI where it expands my agency and makes me feel good; I avoid it where it replaces my own becoming and leaves me depleted.

For what I don’t use it anymore:

• I don’t use AI for broad brainstorming. The ideas it generates are incredibly homogeneous, heavily anchored in the technical biases and epistemological underpinnings of the culture that built it. If I want a genuinely new perspective, I don’t want an average of the internet—I want the friction of a specific, lived human lens.

• I don’t use AI for relationship advice. I used to ask it to help me think through relational struggles, blinded by its confident-sounding oracle voice, only to realize its advice is essentially just scraped Reddit pattern recognition. I refuse to apply online forum advice to emotionally relevant situations; instead, I turn to trusted friends, my therapist, my journal, or my own body through movement and dance.

• I don’t use AI to write for me. Writing is how I think and make sense, and the one who does the work does the learning. You cannot grasp a concept without wrestling with the words yourself, and I refuse to trade the deep satisfaction of forming my own thoughts for the shallow illusion of knowledge.

• I don’t use AI to synthesize research. I have experimented and tested all kinds of AI workflows around research synthesis and I concluded it does not have human instinct on what makes specific pieces click. Finding the truth in research is about recognizing that intuitive itch in my body when a quote or concept is relevant, and I won’t outsource that embodied discernment to technology trained on probability.

When I use AI

• I do use AI for data analytics and pattern recognition. Once I have a clear intention, I want to process large volumes of uploaded stats and numbers without draining my cognitive energy on tasks a AI can do faster and better.

• I do use AI to automate tedious administrative work. I give Claude code access to an unsorted folder of receipts and bank statements to generate my bookkeeping spreadsheets and draft emails to my tax advisor, a task I don’t enjoy at all. I still fact-check its accuracy, but this way I have more mental space left for creative work that brings me joy.

• I do use AI to test and nudge my writing after I’ve written it. I use it as a sparring partner to review my drafts, identify my blind spots, challenge my logic, or point out reader angles I haven’t considered yet.

• I do use AI to deepen my understanding through metaphors. If I am struggling to understand a complex concept, I ask the AI to explain the topic using analogies tailored to my current lived context—like music production—which helps me grasp new ideas faster.

• I do use AI for exploratory deep research. I want a research agent to horizon scan what’s out there and give me a curated list with pointers, before I dive in to do the nuanced reading myself.

• I do use AI as a last resort. When I’ve exhausted all conventional pathways—like navigating the labyrinth of Austrian immigration bureaucracy—I want a back-up hyper-efficient problem-solver in my corner. My tool of choice for this is ChatGPT 5.5 pro.

Choosing Our Own Defaults

Ultimately, setting these AI thinking boundaries is about refusing to let a tech companies in the Bay Area dictate the defaults of my own mind.

made a great point in one of his latest post

“A lot of this [AI development] is happening without, ironically, any real planning or consideration. And I suspect it will be hard to reverse these defaults once a generation of workers and students has built habits around them. The most important thing we can do is keep asking what to hand over and what to keep for ourselves... and not expect anyone, including the AI, to answer that for us.”

So, how do we engage with this technology without losing ourselves to it?

I’ve been thinking a lot about the Ayurvedic concept of Varuna Sattva, often described as the god of water.

To cultivate this state is to act like a river flowing from the mountains to the ocean: you set a deep, clear intention, and you flow toward it, come what may. When obstacles appear, the water doesn’t stop. It goes above, goes under, goes around, or even retraces its path to find a new way to the sea.

This is a purpose-filled existence, opposed to the fish state of mind (Matsya Tamas), darting aimlessly here and there for quick hits of reward, which I equate to the jittery fatigue of switching between AI tabs all day.

We can absolutely use AI as a tool to flow around tedious administrative rocks and bureaucratic roadblocks. But the tool is just a way we can choose to navigate the rock; and it should not alter the water’s course.

Navigating the AI era isn't about rigid rejection, but it certainly isn't about frictionless surrender, either. It is about conscious friction. It is about actively choosing when to let AI help us flow around an obstacle, and when to keep the slow, effortful, beautifully human joy of figuring things out for ourselves.

It’s up to us to decide.

Thanks for reading. I always appreciate hearing what’s on your mind—and how you are navigating these tools in your own life—in the comment section.

Leave a comment

Subscribe now

P.S.: I am putting this philosophy into practice and have decided to take a clear no-AI writing stance. I will be reducing the quantity of my posts, reverting to weekly instead of twice a week starting this week. I am more committed to this work than ever, and choosing the joy of deep and sometimes slow thinking feels aligned with how I want to show up for this publication. Should you have any questions or comments, send a DM. Thank you for supporting this work.

Hi, I’m Eva, independent researcher, and I write about how humans learn and how AI changes that. Translating learning science into strategies to remember more. You’re receiving the free Wednesday edition of the Lifelong Learning Club.

Subscribe now

It was right after my bachelor’s and master’s degrees, and after my two years teaching in a classroom.1 I had just started writing online, and I was struggling to remember what I read. I’d finish a great book feeling inspired, only to stare at a blank screen a week later, frustrated that the insights had already slipped through my fingers.

I’d try to reference ideas I knew I had encountered, only to find my mind blank. My Zettelkasten and Second Brain knowledge organization systems didn’t help; they felt slow and tedious.

Until then, learning had always been a mystery to me. I assumed it was simply a matter of time and effort. When I thought of “learning,” I pictured sitting in a library with a book, rewatching video courses, taking notes, and then re-reading those notes.

Now I know that I relied entirely on things that made me feel productive, trusting that immersing myself in the material was enough.

Then, in September 2020, I took ’s Learning How to Learn course and read two of her books.

I remember the moment my jaw dropped. Oakley explained that learning, in essense, is just moving facts from your short-term memory to your long-term memory. She then pointed out that there is a clear, evidence-based process for doing this—and that re-reading and highlighting are not part of it.

Disbelieving, I grew angry. I had gone through school, two degrees, and two years as a teacher without knowing the basics of how we learn and remember.

But it was also the moment something clicked. I realized that learning isn’t a mysterious black box. It’s a skill. Besides the anger, I felt release, as I finally considered I don’t have a bad memory, and instead, can simply rely on strategies that work.

That realization is what made me write this newsletter. I want all people, especially teachers and lifelong learners, understand how learning works.

And I also want the effective methods to feel good. I want clarity to mean understanding, fluency to mean memory, and confidence to mean competence.

But the evidence keeps pointing in the opposite direction and it took me a year to really grasp the distinction. Because the moments that feel inefficient, repetitive, embarrassing, and slow are often the moments that do the actual work.

Here is the learning science I wish were false. The sooner you understand it, the less time you’ll waste mistaking familiarity for learning—and the more deliberately you can build knowledge that stays with you.

When Feeling Smart Stops You From Getting Smarter

Learning actually happens on two different timelines. One is how well you can perform while help is still available. The other is what you can still do after that help is gone.

We often mistake the first for real learning. If you can follow along with notes, examples, or explanations, it feels like you understand. But that support is doing part of the work for you.

Re-reading a page feels familiar, so it seems like you know it. Watching a clear explanation feels like understanding, but the speaker is carrying the thinking. Even polished AI answers can give the illusion that you’ve mastered something, when you’ve only recognized it. Highlighting text can feel productive, but it mainly strengthens recognition, not recall.

Psychologists Robert and Elizabeth Bjork describe this as the difference between retrieval strength and storage strength. Retrieval strength is how easily you can access something right now. Storage strength is how well it is actually learned and retained over time.

The problem is that retrieval strength feels like learning. It creates a sense of fluency—things seem clear and easy. But fluency is misleading. It tells you what is available in the moment, not what you will still know later, when the support is gone.

Four Learning Truths That Don’t Feel Good (But Work)

1. Effort helps memory

Your brain remembers what it has to work to retrieve. When you try to recall something and can’t quite get it, that struggle feels uncomfortable—but it’s exactly what strengthens learning. Re-reading feels easier because the information is right in front of you. Retrieval forces your brain to rebuild the idea, and that rebuilding is what makes it stick.

2. Forgetting is part of learning

The problem isn’t that you forget. It’s that you expect not to.

Spaced repetition works because it brings material back just as you’re starting to lose it. That moment—when something feels fuzzy—is the ideal time to practice retrieval. If review feels slightly difficult or even annoying, that’s a sign it’s working.

3. Some boredom is useful

If something feels completely meaningless, it’s hard to learn from it. But there’s also another kind of boredom which I’d describe as the restless feeling that makes you want to switch tasks. If you stay with it a bit longer, your focus often deepens.

Learning also involves confusion. When you work through that confusion, understanding becomes stronger than if everything felt easy from the start.

4. Early failure can help

Getting things wrong at first can improve learning later.

Across 53 studies on “productive failure,” researcher Manu Kapur found the pattern that learners who struggle with a novel problem before receiving instruction learn the underlying concepts more deeply than learners who get the instruction first.

When you try to solve a problem before being shown the answer, you create a mental framework. Even if your attempt is incorrect, it prepares your brain to understand the explanation more deeply. The key is to try first, then learn from feedback.

But the level of difficulty should be appropriate.

Struggle only helps if it’s at the right level.

If something is so hard that you have no idea where to start, your working memory gets overloaded and learning breaks down. Effective practice means you can attempt a solution, get part of it right, receive feedback, and improve. Good difficulty guides your effort whereas bad difficulty just overwhelms you.

Hence the timing of help matters.

If you get an answer before you’ve had a chance to think, your brain skips an important step in learning.

Many tools of our times are designed to make things feel easy. But as you now understand, learning isn’t supposed to feel effortless all the time. Long before AI, we already preferred methods that felt smooth—like re-reading or watching clear explanations. These methods create a sense of understanding without requiring much mental effort.

AI makes this effect stronger. It gives fast, polished answers that can make you feel like you’ve mastered something. But studies show a pattern: students often perform better during practice when they can use AI freely, yet perform worse later when the tool is no longer available.

So the lesson here? AI can be very useful if you use it to support durable learning. Try first. Struggle a bit. Make a prediction or attempt a solution. Then use AI to check, correct, or expand your thinking. A good learning tool gives you space to think before the answer appears.

A Practical Way to Think About Difficulty while Using AI

Not all difficulty is helpful. Some obstacles get in the way of learning, while others support you to build memory and understanding.

You want to a) remove unnecessary obstacles: confusing instructions, hard-to-access materials, poor explanations, busywork that doesn’t support learning, and b) keep the useful challenges: trying to recall without looking, making predictions, revisiting material through spaced repetition after you’ve started to forget, explaining ideas in your own words, getting feedback that helps you improve.

What Learning Really Feels Like

It would be nice if learning felt smooth and easy—like downloading information. But that isn’t how the brain works.

Learning happens when you engage with what you can’t yet do or recall. That process often feels uncertain or uncomfortable.

So the feeling of real learning isn’t always confidence. Often, it’s returning—again and again—to what still feels difficult, and gradually making it clearer.

Subscribe now

Sources

• Bjork, R. A., & Bjork, E. L. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. Psychology and the real world. Worth Publishers.

• D’Mello, S., Lehman, B., Pekrun, R., & Graesser, A. (2014). Confusion can be beneficial for learning. Learning and Instruction, 29, 153–170.

• Kirschner, P. A., Sweller, J., & Clark, R. E. (2006). Why minimal guidance during instruction does not work. Educational Psychologist, 41(2), 75–86.

• Pyke, A. A., Lunau, K., & Javadi, A. H. (2024). Reconciling desirable difficulties and cognitive load theory. Quarterly Journal of Experimental Psychology.

• Sinha, T., & Kapur, M. (2021). When problem solving followed by instruction works: Evidence for productive failure. Review of Educational Research, 91(5), 761–798.

• Soderstrom, N. C., & Bjork, R. A. (2015). Learning versus performance: An integrative review. Perspectives on Psychological Science, 10(2), 176–199.

Subscribe now

I always wanted to live in South America. When it came time to choose a university for my exchange semester, I set my sights on Santiago, Chile.

There was just one catch: the program required C1-level fluency and a certificate from a language professor. My Spanish at the time hovered somewhere around “Hola, ¿cómo estás?”

So, I hacked the system. I used Google Translate to write a 4-minute monologue about why I wanted to study in Chile. I asked native-speaking friends to record themselves reading it slowly. I memorized it, recited it to a professor in Vienna, and got my certificate.

But because I actually wanted to survive in Chile, I also looked up the most “effective” study method. I signed up for a flashcard app and built a 5,000-word vocabulary using daily spaced repetition.

It worked—kind of. I passed all my exams in Santiago. But looking back, I got half of it right and half of it completely wrong.

The half I got right: I picked the most evidence-backed memory tool on Earth.
The half I got wrong: I thought spaced repetition was my language-learning method.

It wasn’t. It was a memory layer sitting underneath a method I didn’t actually have.

Anki (and similar spaced-repetition software) is a memory tool inside a language-learning method. It is not the method itself. Used well, it helps you retain vocabulary, notice patterns, and make your reading and listening wildly more productive. Used badly, it becomes progress theater—a growing deck, flawless card recall, and almost no improvement in real-world listening or speaking.

That distinction is the difference between a habit that builds fluency and a deck that quietly steals years of your life.

Here is exactly when flashcards help, when they hinder, and the five rules to build a system that helps you learn any language.

When Anki Helps vs. When It Hinders

If you ask the internet how to learn a language, you’ll find two camps screaming at each other.

• The Input-Only Camp: Popularized by polyglots like Steve Kaufmann, this camp says flashcards are busywork. They argue fluency comes from massive “comprehensible input”—listening and reading content slightly above your level, in volume, for years.

• The Flashcard Maximalists: The old-school Anki crowd argues you should build 10,000 sentence cards, drill them daily, and brute-force the language into your head. Without spaced repetition, they argue, you forget faster than you learn.

Both cite evidence and neither is entirely wrong.

I appreciate the stance of Paul Nation, a leading researcher in second-language vocabulary, who champions the Four Strands principle. A balanced language program needs roughly equal portions of:

1. Meaning-focused input (reading, listening)

2. Meaning-focused output (speaking, writing)

3. Language-focused learning (yes, including flashcards)

4. Fluency development (low-pressure use of what you already know)

Spaced repetition sits in one of those four strands. And it’s a fundamental pillar of language learning.

What We Can Assume Based on Current Evidence

Spaced practice for second-language vocabulary is one of the most replicated findings in applied linguistics. A 2022 meta-analysis found a spacing benefit for long-term retention. In a famous 9-year study, researchers showed that thirteen review sessions spaced 56 days apart produced the exact same retention as twenty-six sessions spaced 14 days apart. Which in other words means you have half the work, but the same long-term result.

But recent studies (Ophuis-Cox et al.; Gupta et al.) clarify a relevant distinction: remembering a rule is not the same as applying it fluently in context.

Adding retrieval practice (flashcards) to your routine drastically improves your ability to recall explicit rules or vocabulary words. But it does not automatically transfer to applying those words under pressure. These studies don't show that flashcards fail; they show that flashcards are best for recall, while application requires practice that actually resembles real-world use.

So what are the implications for learning a language?

To build a mental library you can use, you want two distinct types of knowledge:

• Anki builds the explicit “dictionary entry.”

• Immersion builds the tacit, semantic web.

If you use only flashcards you will plateau at reading comprehension and freeze under speaking pressure. If you only use immersion for learning you will struggle to retain rare, low-frequency vocabulary because you don’t encounter it often enough for it to stick naturally.

Hence, you want both. Anki is the brake against forgetting and comprehensible input is the engine of fluency.

Where Spaced Repetition Software Like Anki Excels

• Initial encoding of new vocabulary: Memorizing the first 2,000–3,000 high-frequency words you need before reading and listening become helpful.

• Memorizing explicit grammar rules: Verb conjugations, particles, gendered articles, and irregular forms.

• Retaining low-frequency words: The rare terms you encounter once in a podcast and won’t hear again for six months.

• Pronunciation drilling: Minimal pairs, tone work, and phonetic familiarity.

• Collocations and chunks: Phrases that hang together naturally (like hacer un esfuerzo or prendre une décision).

Where Anki Hinders

• When it replaces actual listening and reading: Reviewing 200 cards a day feels productive. Comprehending native speech at speed is a different activity.

• When cards are stripped of context: A card showing agua → water tests an isolated form-meaning link. Research shows this is where the testing benefit shrinks fastest.

• When you expect cards to deliver speaking fluency: Knowing a rule and applying it under time pressure are two different abilities.

• When the cards are bad: You can ace your daily reviews, but still freeze when trying to speak to a native.

The rest of this guide is about how to write great flashcards that support your language learning goals. Below the paywall, I’m sharing a science-backed guide to build a spaced repetition system that translates into fluency.

Specifically, we’ll cover:

• The 5 Rules of Card Design: How to structure your flashcards so they train your brain to produce language, not just recognize it (including examples).

• The Flashcard Audit: A quick diagnostic tool to identify and prune the bad cards that might be stealing your time.

• The Level-by-Level Roadmap: How your Anki strategy should shift from month one to year two so you don’t plateau.

• The 60-Minute Formula: A daily time-allocation breakdown for combining real-world immersion with flashcard reviews.

If you are tired of spending hours on language apps only to freeze when a native speaker talks to you, this system is for you. Let’s dive in.

Read more

Subscribe now

You encounter a dense concept at work: a new mental model, a complex coding architecture, a niche economic theory.

Instead of wading through academic papers or a textbook, you do what any smart professional in 2026 does. You ask ChatGPT.

Three seconds later, the chatbot gives you a confident-sounding, well-structured explanation. You read it. You nod along. It makes total sense. You close the tab feeling much smarter than you did five minutes ago.

But that afternoon, when a friend asks you about the same topic. You open your mouth and… nothing. You fumble and realize, with a sinking feeling, that you don’t really know it.

What happened? Luckily, your memory didn’t fail you. Instead, your sense of understanding was inflated by fluency and familiarity — a cognitive glitch that AI amplifies by giving you polished text.

The Illusion of Explanatory Depth

In 2002, cognitive scientists Leonid Rozenblit and Frank Keil ran an experiment. They asked people to rate how well they understood everyday objects, like a zipper or a bicycle.

Most rated their understanding highly. Of course they knew how a zipper worked. They used one every day.

Then the researchers handed them a pen and paper and asked them to draw, step by step, how the gears of a bicycle or the teeth of a zipper functioned.

Almost everyone failed.

Rozenblit and Keil named this the Illusion of Explanatory Depth (IOED). We mistake our familiarity with a concept for understanding of it. We confuse recognizing a pattern with reconstructing it.

The Fluency Trap

If everyday objects trigger this illusion, Large Language Models put it on steroids.

AI engineers spend billions tuning models like Claude and ChatGPT to optimize for fluency. They are designed to remove cognitive friction, delivering answers that are authoritative, legible, and easy to read.

But in the science of learning, easy is dangerous.

Cognitive psychologists, notably Robert Bjork, have studied how fluency effects trick our metacognition. When we read a text that is perfectly formatted and effortlessly processed, we mistake that ease for understanding. We think, “This is easy to read, therefore I’ve easily learned it, and I’ll easily remember it.”

When you read a brilliant AI output, you’re borrowing the AI’s reasoning. But because the text is so clear, your brain tricks you into believing you generated the insight. You get the emotional reward of solving a problem without doing the cognitive heavy lifting that stores the solution in your long-term memory.

But you didn’t build expertise.

A Playbook to Break the Illusion

To use AI to build an interconnected library of knowledge in your mind, rather than outsource your thinking, you have to inject friction back into the process. Bjork calls these “desirable difficulties.”

You can still use AI to explain complex topics. But you must sever the link between consuming the answer and assuming you understand it.

Here’s a 3-step workflow I developed so AI outputs translate into human expertise. I call it the Reverse Feynman Audit, a method grounded in retrieval practice and self-explanation.

The Blind Draft (The Friction):

After the AI explains a concept, read it carefully. Then close the tab. Don’t look at it. Open a blank document or grab a notebook. Force yourself to write out the concept in your own words, using an analogy from your own life.

Why this works: it forces active retrieval. The moment you struggle to find the right word, you shatter the Illusion of Explanatory Depth. You expose where your knowledge ends and the AI’s begins.

The Reverse Feynman Prompt (The Feedback):

Once you have your messy, imperfect explanation, paste it back into the AI. You’re no longer asking it to be your teacher. You’re using it as a peer reviewer.

Copy-paste this prompt:

“I’m trying to deeply understand [Topic]. I’ve written an explanation in my own words to test my comprehension. Act as a rigorous cognitive science professor. Review my explanation below. Do not rewrite it for me. Instead:

1. Identify any conceptual gaps, logical leaps, or factual errors in my thinking.

2. Point out where my analogy breaks down.

3. Ask me one challenging follow-up question that forces me to think about the edge-cases of this concept.”

The Calibration (The Synthesis):

Read the AI’s critique. Answer its follow-up question. You’ve moved from passive consumption to elaborative rehearsal, because you’ve wrestled with what you’re reading.

The Architecture of Understanding

Often, AI erases the friction of learning. It has never been more tempting to let an algorithm do the heavy lifting, and coast on its polished outputs.

But building knowledge and expertise, just like building muscles, does require a bit of strain, of doing a bit of the lifting yourself..

Use technology to give yourself richer material to think about, not to bypass thinking.

The next time ChatGPT hands you a perfect answer, notice the dopamine hit of easy fluency. Enjoy the clarity. Then make a deliberate choice. Close the tab, grab a pen, and prove to yourself that you actually know it.

References

Bjork, R. A., Dunlosky, J., & Kornell, N. (2013). Self-regulated learning: Beliefs, techniques, and illusions. Annual Review of Psychology, 64, 417–444. https://doi.org/10.1146/annurev-psych-113011-143823

Brown, P. C., Roediger, H. L., III, & McDaniel, M. A. (2014). Make it stick: The science of successful learning. Belknap Press.

Rozenblit, L., & Keil, F. (2002). The misunderstood limits of folk science: An illusion of explanatory depth. Cognitive Science, 26(5), 521–562. https://doi.org/10.1207/s15516709cog2605_1

Subscribe now

For a long time, I was skeptical about using AI for research.

Ask ChatGPT or Claude for a literature review and you usually get back something polished, with citations. It looks perfect. But click through to the primary sources and half the papers don’t exist, or the report misreads what they say.

This isn’t just my intuition. In 2023, Walters and Wilder tested ChatGPT across four subject areas and found 55% of GPT-3.5’s bibliographic citations were fabricated. With GPT-4, the rate only dropped to 18%.

Because you never know when to distrust the output, you end up doing exhausting vigilance work—verifying every link, author, and claim by hand.

But recently, my skepticism has changed into awe.

As I’ve learned more about how to direct AI in service of my intentions (and build automated verification loops with it), I’ve been amazed by the quality of the output.

So much so that I named the Claude Code skill1 I adapted and built around it Jean2, after a head of research whose rigor I admire.

The version I use now is adapted from an open-source Claude Code Academic Research skill by chrislemke/stoffy, which I customized for my own research process.

In this post, I’ll walk you through the thinking behind the Jean research skill so you can build your own—and share how to install mine if you’d rather just use it.

What The Claude Research Skill Does

Most AI tools start guessing the second you hit enter, leaving you to do the draining work of checking their sources. Jean isn’t smarter but it follows a clear protocol (with verification loops) that leads to a more accurate output:

• Scoping before search: Instead of immediately writing a generic summary, it stops to clarify the domain, the depth (5, 15, or 30 sources), and the focus area.

• Three cycles of research: Instead of one web search, it targets core academic databases (like Semantic Scholar and arXiv) first. Then it expands to specific authors and journals. Finally, it follows citation trails to find foundational papers.

• Source evaluation & triangulation — every paper is rated for relevance, authority, recency, and type; cross-referenced for consensus, debates, gaps, and key figures.

• Fact-checking for hallucinations. Before writing the final report, it must confirm that every cited paper exists via web search. It requires a title, author, year, and an actual DOI/URL. If a citation cannot be verified, the source is deleted.

• Highlighting “Open Questions”: By cross-referencing sources, it tells me where the field hasn’t settled yet. This prevents from presenting a confident-sounding, yet false consensus.

When I Use The Claude Research Skill

I use this research skill I named Jean when I want a literature review — papers, DOIs, competing theories, named debates; before I write. .

For example I ran the skill on questions like “What Did Phones, Social Media, and Notifications Do to Our Attention? Historical and Contemporary Perspectives”. And after twenty minutes I received this report (which informed my post here).

260416 Jean Attention Distraction Phones Social Media

105KB ∙ PDF file

Download

Download

Another question I used the skill for were spaced repetition selection principles in the age of AI (which informed my post here).

260423 Jean Srs Selection Principles In The Age Of Ai

115KB ∙ PDF file

Download

Download

I then use these outputs as starting points for further reading. It’s like my personal expert reviewer and researcher and then I dive deeper on what excites me and what I am curious about.

The rest of this post is for paid subscribers. Below, you’ll get the skill to download and a 15-minute install guide for non-coders.

🛠️ The Skill File & How to Install It

Read more

Subscribe now

When I decided to learn Claude Code, I didn't expect the cognitive whiplash that followed.

It took me a week to learn how to use the terminal, organize MD files, and to find my way around. The learning curve was steep, yet surprisingly short. Soon I was spinning off AI agents in parallel, letting Claude Code1 coach me through setting up connectors to Gmail, Canva, Otter AI, and others.

It felt like I'd hired 50 new people all at once. But within weeks, I hit a wall.

Every workday felt exhausting. I wasn’t recovering overnight. By the end of March, my thinking became slower, scattered, drained.

I started to become scared about the implications. Here I was, living on a tropical hippie island, surrounded by the ocean, jamming with my guitar at the beach, and running into friends who hug me whenever I take a break.

If I felt this drained in paradise, how were people coping in high-stress cities?

We treat agentic AI like a harmless productivity hack when we should treat it like an experimental drug. To understand why “multi-clauding”2 can feel so depleting, I looked at the evidence on automation, vigilance, cognitive load, and attention.

The Ironies of Automation

Working with Claude Code, or any autonomous AI setup, isn’t the same kind of work as doing the work yourself. Cognitively, this type of work is much closer to what researchers call supervisory control or automation monitoring.

Until recently, we’ve assumed that if AI does the executing, our brains get to coast. Existing evidence points toward this not being as simple as that.

In 1983, researcher Lisanne Bainbridge coined the “Ironies of Automation.” She found a structural flaw in how we design systems: the more you automate the easy, repetitive parts of a job, the harder, less rewarding, and more exhausting the residual human role becomes.

When you run multiple AI agents, you stop being a creator and become an air traffic controller. You’re doing vigilance work: sustained attention waiting for rare signals or errors. As far back as 1948, radar operator studies showed that human detection rates plummet within the first half-hour of passive monitoring.

This means that watching a mostly-reliable system do the work requires hard, stressful mental effort.

The Illusion of Speed

I still put up with this exhaustion because I believed AI would make me exponentially faster (and better!). New data suggests this might be an illusion.

In a 2025 randomized controlled trial by METR, experienced developers using AI tools were 19% slower than those working without them. The developers believed they were 20% faster.

Why the disconnect? The time formerly spent in the “flow state” of coding had migrated to prompting, waiting, and reviewing.

A 2025 Microsoft Research study of 319 knowledge workers surfaced a similar pattern. Under generative AI, human effort shifts from problem-solving to information verification and task stewardship. The higher their trust in the AI, the more their critical thinking declined.

Every time your AI agent silently changes strategy, edits a file you didn’t expect, or misinterprets a prompt, you experience what aviation researchers call “automation surprise.” Supervising and evaluating an AI model comes with a heavy cognitive tax.

Why Your Brain Hates “Multi-Clauding”

Even people building these tools are noticing the shift. My learning scientist deep work primed mind was surprised when I heard the inventor of Claude Code say this in an interview:

“My work now is just jumping between tabs kind of tending to the Claudes .. to make sure they're unblocked answering their questions. I feel like now is the age of multi-clauding.”

— Boris Cherny, Head of Claude Code at Anthropic

In cognitive science, Sweller’s Cognitive Load Theory explains why we feel so drained when multi-clauding.

AI offloads our intrinsic load (you don’t have to memorize a coding language). But running multiple agents inflates our extraneous load. You split your attention between the prompt, the code, the chat, and the output. Every time you context-switch between agents, your brain leaves behind “attention residue,” making the next task harder.

A 2024 meta-analysis across 29 countries showed that mental effort is experienced as unpleasant. The frustration, irritation, and stress you feel when “another Claude needs me” isn’t a very human biological response to sustained executive demand.

I felt so depleted during the month I maxed out my Claude Code tokens. I love deep work. I’ve meditated almost daily for more than a decade, and I know my brain enjoys sustained concentration and complex thinking.

But the depletion that came from tending to multiple Claude workstreams — from multi-clauding — felt different. Less like focused effort and more like something slowly draining my cognitive and emotional energy. Like a dementor from Harry Potter hovering in the background.

Steve Yegge uses a similar mystical analogy and calls this feeling the AI Vampire:

I’ve been watching the effect the AI Vampire is having on people around me and I’m growing concerned. We’re all excited, but it’s also… weird.

I already posted about […] how I fall asleep suddenly at all hours of the day after long vibe-coding sessions, and how my colleagues at SageOx are seriously considering installing nap pods at the “office.” I’m still sleeping a crazy amount.

It would seem that we are addicted to a new drug, and we don’t understand all of its effects yet. But one of them is massive fatigue, every day.

A Playbook for not Depleting Your Mind

We won’t put the AI genie back in the bottle. But to protect our cognition (and our sanity) we can change how we work with these tools.

• Limit your span of control. Just because you can run four agents at once doesn’t mean your working memory can handle it. Treat AI agents like human direct reports. Start by supervising one task at a time. Your bottleneck isn’t computing power but your visual-spatial and working-memory channels. I started to have “no AI days” where I just do deep focused work the way I used to do it pre 2023.

• Beware the vigilance decrement. Passive monitoring depletes your brain faster than active creation. Don’t stare at the screen waiting for the AI to finish. Step away, reset your attention, and return to review the output with fresh eyes.

• Protect your “germane load.” The greatest risk of automation is losing your mental model of the project (what researchers call situation awareness). Before you hand a big task to Claude Code, understand what you’re building. If you don’t know how the house is framed, you won’t know when the AI installs a faulty beam. I wrote about this in more depth here.

The Choice Between Creation and Fatigue

Decades from now, we may look back at our current, unfiltered use of AI agents the way we now look at teenagers given unrestricted access to social media.

We’re handing over the rewarding, schema-building work of creation, and trading it for the exhausting, high-stress work of continuous vigilance.

Use AI with intention. Decide whether you’re using it to enhance your thinking, or outsourcing your cognition until there’s nothing left but fatigue.

References

Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779.
Becker, et al. (2025). METR Randomized Controlled Trial on AI Coding Speed. (Pre-print/Ongoing).

David, L., Vassena, E., & Bijleveld, E. (2024). Mental effort is unpleasant: A meta-analysis. Psychological Bulletin.

Endsley, M. R. (2017). Toward a theory of situation awareness in dynamic systems. Human Factors, 37(1), 32–64.

Lee, Sarkar, Tankelevitch, et al. (2025). Shift in Knowledge Work under GenAI. Microsoft Research CHI 2025.

Mackworth, N. H. (1948). The breakdown of vigilance during prolonged visual search. Quarterly Journal of Experimental Psychology, 1(1), 6–21.

Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410.

Sarter, N. B., & Woods, D. D. (1997). Team play with a powerful and independent agent: Operational experiences and automation surprises on the Airbus A-320. Human Factors, 39(4), 553–569.

Warm, J. S., Parasuraman, R., & Matthews, G. (2008). Vigilance requires hard mental work and is stressful. Human Factors, 50(3), 433–441.

Subscribe now

Read more