For two years the headline AI skill was prompt engineering. Conferences. Job titles. LinkedIn courses promising five-figure salaries for people who knew how to politely ask a chatbot. The industry built a whole vocabulary around squeezing better output from a fixed input.

That entire frame is collapsing right now, and most people building with LLMs have not caught up. The thing that actually moves model output quality in 2026 is not what you write in the prompt. It is what you put in the context window around the prompt. Welcome to context engineering. It is harder, it is more interesting, and if you build with AI, it is the skill you need next.

What Context Engineering Actually Is

A modern LLM does not just see your prompt. It sees a payload. That payload includes the system prompt, the conversation history, retrieved documents, tool definitions, function outputs, examples, structured state from earlier turns, possibly a scratchpad, and whatever else your application decided to stuff into the window.

Context engineering is the practice of deciding what goes into that window, in what order, in what format, with what budget, for every single call.

Prompt engineering asks: what should I say to the model? Context engineering asks: what should the model know before it tries to say anything?

The shift is exactly the shift from optimizing the question to optimizing the briefing the model gets before the question lands. If you have ever walked into a meeting with no context and tried to make a decision, you understand intuitively why the second one matters more.

Why This Got Critical Right Now

Three things broke the prompt-engineering paradigm in the last 18 months, and they all happened simultaneously.

Context windows got huge. A million tokens is now table stakes. That sounds great until you realize you have to decide what to do with all that room. A bigger window does not produce better output automatically. It produces a more important architectural choice.

Agents arrived as a real product category. Agents do not run one prompt. They run dozens, in a loop, with state that accumulates and tool outputs that need to be integrated. Every step is a context engineering decision. Get any one of them wrong and the whole agent drifts.

RAG became standard. Almost every serious AI application now retrieves documents from a vector store and injects them into context. Which documents? In what order? Summarized or verbatim? With metadata or without? Those choices, not the prompt language, determine whether your RAG-powered tool feels brilliant or feels like Clippy with a Pinecone subscription.

The prompt got smaller and less important. The context got larger and more decisive. That is the whole transition in one sentence.

What Good Context Engineering Looks Like in Practice

Let me make this concrete with examples from analytics work, because that is where I see most teams getting it wrong.

Say you are building an AI assistant on top of your data warehouse. A user asks, “Why did revenue drop in EU last week?” A naive implementation passes that question to GPT-4 with maybe a schema dump in the system prompt. The output is whatever the model can hallucinate from the table names.

Good context engineering for that same question looks completely different. It assembles a context window that contains:

1. A focused subset of the schema, only the tables relevant to revenue and EU geography, not the full warehouse
2. The actual SQL query for “revenue by region by week” with the last 90 days of results inline
3. A precomputed anomaly detection summary showing which segments dropped most
4. The definition of “EU” as your business defines it, because Europe has at least three valid definitions in a typical company
5. The conversation history, summarized, not verbatim
6. A scratchpad showing what tools the model has already called
7. The user’s role and what dashboards they typically look at, so the answer can reference what they already know

The prompt that triggers all of that might literally just be the user’s question. The context engineering is everything that gets assembled before the question hits the model. That assembly is where the actual product lives.

The Mental Model That Helps

Think of the context window like a meeting briefing for an extremely smart but extremely amnesic colleague who shows up cold to every meeting.

You can dump a thousand-page binder on the table. They will technically read it, but they will get confused, focus on the wrong sections, and probably mention something from page 847 that you did not want surfaced.

Or you can hand them a tight three-page brief. Background. The decision to be made. The relevant data. What was decided last time. What constraints apply.

The second briefing produces better decisions. Every time. The model is not different from your amnesic colleague. The context window is the briefing. Context engineering is the discipline of writing that briefing well, on the fly, for every request.

The Mistakes Almost Everyone Makes

Three patterns I see over and over.

Dumping everything because the window is big. A million tokens is not a license to ignore curation. Large contexts dilute attention, raise costs, and degrade output on the items at the middle of the window. Models still suffer from “lost in the middle” effects. Bigger is not better. Relevant is better.

Ignoring ordering. What goes at the start of context and what goes at the end gets weighted more heavily by most models than what goes in the middle. If you are putting your most important grounding documents in the middle of a 200-page context, you are quietly tanking your own output quality.

Treating the system prompt as set-and-forget. A static system prompt was fine when you had a static chatbot. It is wrong for agents. The system prompt should change as the agent’s state changes. What capabilities are available right now? What did we just learn? What should the agent stop trying to do because it already failed? All of that is context.

What This Means for Analytics Teams Specifically

If you are building AI features into your analytics stack, the implication is clear. Your competitive advantage is not the model you picked. It is how well you assemble context for it.

That means investing in retrieval quality, schema selection logic, query templating, results formatting, and conversation summarization. None of those are glamorous. All of them compound. The teams that win at AI in analytics are the ones who treat context assembly as a first-class engineering discipline, not as a side effect of however RAG happened to ship in their framework.

It also means the skill set you hire for has shifted. The person who can write a clever prompt is not who you need. The person who can design the data pipeline that delivers the right context to the model at inference time, with low latency, at controlled cost, with quality monitoring, is who you need. That is a real engineering job. It looks a lot more like data engineering than copywriting.

The Practical Principles

If you want to start improving your context engineering this week, here is the order I would do it in.

Audit what is in your context window today. Log every payload your AI features send to the model. You will be horrified at what is in there. Most teams are sending massive amounts of irrelevant context they have never looked at.

Set a budget per call. Decide how many tokens this feature is allowed to spend on context. Force the choice. Constraints produce better engineering than abundance does.

Order by relevance, not by source. The order in which you concatenate retrieved chunks matters. Put the most important content where the model pays most attention. For most models, that is the top and the bottom of the window, not the middle.

Cache aggressively. A lot of context is reusable across calls. Schema definitions, business glossaries, examples. Use the caching mechanisms your provider supports. They were built for exactly this.

Measure output quality and tie it back to context choices. When the output is bad, do not just rewrite the prompt. Look at what was in the context window. The answer is almost always there.

The Short Version

Prompt engineering was about asking better questions. Context engineering is about giving the model a real briefing before it answers anything. The difference between those two framings is the difference between AI features that feel like a magic trick and AI features that actually move metrics.

The teams treating this as the new primary skill are quietly pulling ahead of the teams still optimizing the wording of their system prompt. The window of being early on this is shrinking fast.

You can keep refining how you ask. Or you can start engineering what the model knows.

Prompt Engineering Is Over. Context Engineering Is What Comes Next. was originally published in Science Spectrum on Medium, where people are continuing the conversation by highlighting and responding to this story.

Imagine you’re hiking in the cool shade of a beech forest in the Basque Country, or that you’ve lost your way during a treasure hunt in the Amazon rainforest. With every step, you feel the power of nature, but also a limitation: from the ground, we can only see what surrounds us, trunks and leaves that hide the forest’s true magnitude.

Now imagine climbing higher. From the air, no matter where you look, you see only an endless sea of green. That “roof” of vegetation, the forest canopy, hides millions of stories below: giant trees, plants, flowers, fungi, animal life, and clues about something that directly affects us all: the global carbon cycle.

Lasers to Study Forests

That’s where a fascinating tool comes into play: LiDAR, a laser-based technology capable of reconstructing the three-dimensional structure of the world around us. LiDAR is everywhere now. Even in our phones (at least the expensive ones). Autonomous vehicles use it to detect pedestrians, cyclists, and obstacles on the road. Archaeologists employ it to reveal temples and cities hidden in the jungle. Topographers rely on it to map landscapes and terrain with high precision.

The principle behind LiDAR is actually quite simple: it works by emitting laser pulses and measuring the time it takes for the reflected light to return. By doing so, it is possible to accurately determine distances, which in turn allows us to recreate the geometry of objects and surfaces in 3D.

LiDAR sensors can be mounted in a variety of vehicles. When we think about terrestrial or airborne LiDAR, the results are quite intuitive to visualize. A laser scanner mounted on a tripod, a drone, or an aircraft emits millions of pulses in many different directions. Every time a pulse encounters an object, an echo comes back to the scanner. By measuring the time between the emission of the pulse and the echo, the system calculates the exact distance between the scanner and the object that reflected the pulse. Combined with the known direction in which the pulse was emitted, it is possible to determine the exact position of the object in space. By repeating this process millions of times, LiDAR progressively produces what is known as a point cloud: a dense collection of three-dimensional points representing the scanned environment.

In forest ecosystems, these point clouds acquired by LiDAR can reveal trees, trunks, branches, leaves, and crowns in remarkable detail. When playing with the data, you can even walk virtually through the forest and observe its structure from any angle.

Spaceborne LiDAR Is Different

Spaceborne LiDAR, however, works differently and is therefore less intuitive to imagine. From hundreds of kilometers above Earth, satellites and orbital instruments can’t sweep landscapes with the same precision as airborne systems. Instead of generating detailed 3D point clouds, they measure something called waveforms.

A waveform can be thought of as the signature of a laser pulse interacting with the forest. When a pulse is emitted from space, it travels through the atmosphere, reaches the canopy, passes through the treetops, bounces off the ground, and returns to its point of origin, loaded with information. Indeed, fractions of the emitted pulse are reflected back at different moments depending on what they encounter along their path: some energy bounces off the tops of the trees, some from branches and leaves, and finally some reaches the ground under the vegetation.

So rather than recording the echoes of individual points, the instrument measures the amount of returned energy over time. This is what a waveform is: a one-dimensional signal describing the distribution of returned energy over time, which, in the end, describes the vegetation’s vertical structure from the top-of-canopy down to the ground.

Waveform data is very valuable because it contains direct information about forest vertical structure. In theory, there is information in absolutely every aspect of this signal: the number of peaks, their width and intensity, their shape and slope, and even the distance between them. It’s like dissecting the signal looking for coded information. For example, a tall and dense tropical forest in the Amazon will produce a very different signal from a pine plantation or a sparse woodland.

Of course, spaceborne LiDAR remains less accurate than terrestrial or airborne LiDAR systems. If we could choose, it’s always better to have a point cloud than a waveform. However, a terrestrial laser scanner is a very heavy device that you need to transport in a backpack in the field. You can’t scan more than a few plots of forest before it becomes impractical. I’ve done it before, and it’s honest work. Airborne LiDAR allows us to scale up observations by mounting scanners on aircraft or drones, but these acquisitions can be expensive and remain limited in spatial coverage.

Granted, spaceborne LiDAR measurements are less detailed and more difficult to interpret, but they provide something unique: consistent observations across the entire planet. This allows for studies all over the world at many different scales. Moreover, and that’s important for researchers, the data are freely available. Of course, someone still has to pay because launching and operating a space mission costs an enormous amount of money. But this cost is absorbed by the mission agencies rather than transferred to the data users. Anyone around the world can access information about forest structure from space and study ecosystems on a global scale.

Return of the GEDI

One of the most innovative space missions in this field is NASA’s GEDI (with a G, standing for Global Ecosystem Dynamics Investigation). It’s an instrument mounted on the International Space Station (ISS) that fires nearly 250 laser pulses per second toward Earth.

GEDI measurements are not continuous, though. This is not a limitation specific to GEDI itself, but rather to the technology: spaceborne LiDAR doesn’t produce continuous measurements over the territory. Instead, each laser pulse samples a specific circular area on the ground called a footprint. Along the ISS orbit, these footprints form tracks of observations distributed across forests around the planet. For every footprint, GEDI records a waveform describing the vertical structure of the forest at the footprint’s location.

This enormous amount of collected data, in the form of millions of waveforms, enables the observation of vast forested areas for analysis on a regional, continental, and even global scale. From these measurements, scientists derive key forest variables and metrics: the height of the canopy top, the vertical distribution of foliage (i.e., the mass of leaves), and the density of vegetation at different forest layers. These parameters are then combined with field measurements into statistical models to estimate a fundamental parameter: aboveground biomass, that is, the amount of living matter contained in vegetation.

NASA shares data with the scientific community at various levels of processing: from raw waveforms to biomass products across different regions of the world. The most complete and processed product stitches together billions of individual footprint measurements into a continuous and global map of the world’s forest carbon stocks at a 1 km × 1 km cell resolution.

Why Is This Important?

Because forests, whether tropical or temperate, are the planet’s true lungs. They absorb carbon dioxide, one of the main gases responsible for climate change, and store it in the form of wood and leaves. Knowing how much biomass they contain and how it changes over time helps us better understand the role they play in regulating the global climate.

By combining spaceborne LiDAR data with other satellite imagery (both optical and radar), it is possible to obtain a much more comprehensive view of forest ecosystems: vegetation maps can be created, changes over time can be detected, and we can even anticipate how forests respond to phenomena such as droughts, fires, or human pressures. These tools have been transforming the way we study the environment, not only forests.

In the end, the real treasure isn’t hidden in some lost city in the jungle, nor beneath the roots of a centuries-old oak tree in the Basque Country, but in the forest itself. And thanks to technology, we now have the opportunity to observe it, understand it, and, above all, protect it.

Thank you for reading! I’m Kamel Lahssini, and I share stories about how humans observe and map the world. If you want to read more, don’t hesitate to follow me. I’m always happy to reply to a thoughtful comment and start a conversation!

Why Is NASA Firing Lasers at Trees From Space? was originally published in Science Spectrum on Medium, where people are continuing the conversation by highlighting and responding to this story.

I clearly remember how I first entered “how to become a data scientist” into Google search bar.

All the websites provided practically the same advice. Learn Python. Then stats. Then pandas. Then NumPy. Then scikit-learn. There was a step-by-step plan, a sequence of things to do.

That’s what people follow. People spend weeks, sometimes even months, learning all the basics of Python. Print functions. For-loops. List comprehension. OOPs. People take a course, pass it, and are proud of themselves. But then they work with the first real-life dataset.

And they don’t know how to work with it at all.

But not because they haven’t learned Python well enough. Because they’ve learned Python before they ever learned about data science and its requirements.

The Advice Isn’t Wrong. It’s Just Incomplete in a Dangerous Way.

Here’s the bottom line: Python is necessary. There is not a single version of a data science career that does not require Python. This is not an argument against learning Python.

This is an argument against the notion that learning Python can be the basis for one’s data science journey.

It matters much more than you might think.

Once you get into Python as a base for data science, everything changes.
You start focusing on all the wrong things. Are you sure you should write code as a list or a tuple? Dozens of hours go into writing something “Pythonically.” Impostor syndrome strikes when other people’s loops are prettier than yours. Programming skills and data science knowledge get conflated — they are two completely different topics.

Learning problem framing, on the other hand, is at the core of being a good data scientist. How can we improve something by analyzing the problem? What is “better” and how to measure it?

These questions can only be answered through experience, but that’s not how we train our newcomers.

What Actually Happens When Python Comes First

Here’s a common trend that happens all too often in the world of data science education.

An individual dedicates three months to mastering Python. Becomes accustomed to its syntax and might even create a project. Starts a data science course and reaches the machine learning part. Knows how to execute code. Knows how to import Scikit-Learn library, fit a model, and calculate the accuracy score.

But why did they choose that particular algorithm? What does their accuracy score mean in the context of a business problem? What would happen if their model works well on the testing data but doesn’t work once deployed in production?

Nothing.

Not because they aren’t intelligent people. Because they learned how to play the instrument without ever grasping the music behind it. Became obsessed with mastering the tool rather than what it was supposed to be used for.

It generates a type of self-confidence that is incredibly fragile — they only feel competent until somebody asks something challenging.

The Thing Nobody Tells You to Learn First

Prior to any lines of code being written in Python, the single most important thing for a beginner data scientist to know would be how to formulate a good question to ask of data.

In theory, this idea may seem vague; in practice, it is anything but. Good questions asked of data are specific, measurable, and applicable to a concrete real-world problem. What drives customer churn? Bad question, and a rather meaningless one at that. What customer behavior within the first 30 days of onboarding predicts churn before month three, and how reliably can such behavior be identified? That is something that could be turned into a real deliverable.

To become adept at thinking like this does not require Python programming. What it requires is exposure to the world of business problems and solutions. Reading case studies from successful data-driven enterprises, and understanding not just their successes but also their failures and where they went wrong. To always be asking yourself ‘so what?’ when presented with an otherwise impressive statistic.

Statistics itself comes next — but only statistics at its basics. Probability, distributions, the real meaning and significance of a p-value (and all the many ways in which it has been misunderstood), and the difference between correlation and causation.

The Problem With Roadmaps in General

The “learn Python first” approach lives within the larger culture of roadmaps, which is a flawed concept in itself because it is inherently linear, whereas learning is not.

Roadmaps make us comfortable because they guarantee a path towards expertise. Just follow the map, and you’ll end up at your destination point. However, the reality is that data science is an industry that involves mastering different kinds of knowledge at once: technical, statistical, and business-related. And it’s impossible to do so linearly.

Without a context, you won’t remember Python syntax as efficiently as when using a particular tool. Without building a solution, statistics will stay abstract and unhelpful. Without the business question, you will optimize the model’s performance on metrics that do not matter in the slightest.

But the most effective approach is having an actual or a hypothetical problem in front of you and searching for the instrument that can solve it. It may be Python, it may be a statistic, or it may simply be another way of thinking.

A More Honest Starting Point

If you are in the beginning stages of learning Data Science, this is my actual advice:

Start with a problem, not a language. Find some data on an interesting topic for you — sports, music, healthcare, finance, whatever. And write down what the problem you want to solve and why it’s important before even touching any code.

Only learn as much Python as makes you dangerous. It’s impossible to know all the ins and outs of any programming language from day one. But you will learn faster by learning the basics and starting to use them immediately to solve your actual problem.

Think statistically. Spend five hours learning probability and statistics this week for every 10 you spend learning Python. These skills reinforce each other and should grow at about the same speed.

Build systems that need to work. Not little experiments you don’t take seriously after building. Projects in which you define criteria for success from the get-go (“this system needs to predict with Y accuracy”) and make sure you meet them.

The Hard Truth About Shortcuts

This is what “learn Python first” is actually about: providing an opportunity for shortcuts that work.

Python training programs are carefully organized. You progress nicely along the way. You’ve done a chapter; you feel good about it. This feeling will last, but it will also be irrelevant, because you will only realize that you have no idea how to do data science when you face an actual challenge and find out that your 97% accurate classifier doesn’t mean shit.

Going the hard way, building both analytical and technical skills at the same time and having to wrestle with statistics and real-world problems from the very beginning — this approach lacks the nice and clean feeling of progress and accomplishment.

But this is what creates data scientists, not students.

One Last Thing

The people I’ve seen transition most successfully into data science roles didn’t necessarily have the most polished Python. They had a way of thinking about problems that made their technical work mean something.

You can learn Python in a few months. Developing the judgment to know what to build, why to build it, and how to know if it worked — that takes longer, and it doesn’t come from any tutorial.

Start there.

Thanks for reading. If this challenged something you assumed, that’s the point. Data science is hard enough without following advice that was never fully examined in the first place.

The “Learn Python First” Advice Is Holding Aspiring Data Scientists Back was originally published in Science Spectrum on Medium, where people are continuing the conversation by highlighting and responding to this story.

That distinction may determine whether AI ultimately expands human judgment or gradually erodes it. The implications are far-reaching: leadership, learning, management, psychological resilience, and political compromise all rely on our capacity to tolerate uncertainty, disagreement, and cognitive effort itself.

Most discussions about AI focus on what machines are becoming. Far less attention has been paid to what humans are becoming alongside them. Most concern has focused on the gradual erosion of our willingness to engage deeply with thought itself and to tolerate the effort it requires.

In our recent conversation with theoretical neuroscientist, entrepreneur, and author Vivienne Ming on the “Learning Machines Podcast,” we explored many topics but focused on the growing divide emerging among AI users and whether the accelerating convenience of AI is gradually reshaping the habits of our minds — especially in environments increasingly optimized for cognitive ease, personalization, affirmation, and frictionless convenience.

While GenAI has been in general use for only a few years, users now range from passive consumers who increasingly outsource thought reflexively to highly engaged users who integrate AI deeply without surrendering intellectual agency.

As Ming observes, “the vast majority of even smart people shift into this ‘Automator’ process by which their cognition largely shuts down.” They feel like they have thought about something, but really have not. Ming’s research corroborates this with brain activity scans, building on existing neuroscience research showing that passive forms of media consumption engage neural systems differently than cognitively effortful tasks. These latter ones typically involve executive control, working memory, and sustained analytical reasoning.

Broadly, increasingly frictionless interaction gradually shifts cognition away from active reasoning and toward passive consumption.

In between are Validators who do not entirely outsource thinking; however, they do outsource skepticism. They approach AI not as a mechanism for challenging assumptions, but as an instrument for reinforcing conclusions they already wish to believe. In this dynamic, conversational fluency becomes psychologically seductive because affirmation begins to feel indistinguishable from accuracy.

At the other extreme, and the opportunity for us to learn from AI how to think about our own thinking, Ming describes the Cyborg user as an individual capable of integrating AI deeply into life and work without surrendering agency, skepticism, or intellectual independence. The main difference between a passive user and an active user is the level of cognitive effort — the latter is deeply engaged, pushing back, in a setting in which AI becomes the loyal opposition.

As a result, “the future may depend less on the capabilities of machines than on the kind of users we become,” Ming notes. In that respect, “AI is less a window than a mirror. It is showing ourselves back at scale,” she says.

What do you see when you look in your AI mirror?

This piece is based on a Learning Machines column and podcast with The American Bazaar. Vivienne Ming’s latest book is entitled “Robot-Proof: When Machines Have all the Answers, Build Better People.”

Slowly at First, Then All at Once

The shift toward cognitive delegation did not begin with GenAI. The process began over 25 years ago with the advent of the internet; information became radically easier than ever.

Research on cognitive offloading has long suggested that individuals retain information differently when they believe it will remain continuously accessible online. In what became known as the “Google effect,” researchers Betsy Sparrow et al. found that people were significantly less likely to remember information when they expected they could retrieve it later digitally, suggesting that technological convenience may alter not only memory itself but the willingness to engage deeply in remembering — here again affecting how we think about thinking and knowing.

Another example is the use of GPS, perhaps the first widely adopted technology that routinely made decisions on our behalf with minimal cognitive participation. Research on navigation and spatial memory shows that over-relying on automated directions weakens active cognitive engagement and might actually make it easier to get lost.

Before search engines or calculators, acquiring knowledge required deliberate effort through libraries, long-form reading, archival research, and the painstaking synthesis of multiple perspectives over time. One could not simply consume information passively but had to wrestle with it, contextualize it, and independently determine its relevance.

What has occurred during the last several years with LLMs and increasingly agentic AI represents something fundamentally different. As In Ming noted in an article in Worth, the transition is no longer about retrieving information efficiently but about outsourcing increasingly sophisticated forms of cognitive labor altogether — and even delegating the decisions to be made after this outsourcing.

More Friction to the Rescue

The question then is how to preserve discernment and improve outcomes in environments optimized for ease. This is where friction becomes indispensable.

A concept increasingly referred to as “friction maxxing” has emerged in behavioral science and digital culture to describe the intentional reintroduction of resistance, effort, and deliberate inconvenience into environments that are increasingly optimized for speed and immediacy.

While the phrase itself carries the aesthetic language of internet culture, the principle beneath it reflects something far older and far more fundamental, grounded in the recognition that the human brain was never designed to flourish in an environment where every question is answered instantly, and every cognitive burden can be externalized without effort.

Friction, in this sense, represents the cognitive tension necessary for discernment. It is the pause between receiving information and accepting it. It is the willingness to interrogate an output before accepting it as truth.

Friction, therefore, becomes a form of cognitive self-defense. It forces interruption where automation encourages passivity. It creates space for skepticism in environments optimized for immediacy. It compels individuals to ask not merely whether an answer sounds persuasive, but whether the reasoning beneath it can withstand scrutiny.

Friction preserves the possibility of independent judgment precisely because it slows the automatic acceptance of polished, effortless outputs — and in the process gives our brains the time to process.

We previously explored the dangers of misunderstanding both machines and ourselves. The consequences of assuming that access to information is equivalent to understanding, as well as the interpretive gaps that emerge when humans and machines falsely assume mutual comprehension.

GenAI intensifies that risk because most users have little understanding of the data on which models are trained, the biases embedded within them, or the probabilistic architecture through which outputs are generated. For the average individual, evaluating the “ingredients” of a model is nearly impossible, leaving many users with little choice but to trust what appears authoritative.

At the same time, many individuals continue to interact with AI as though it were simply a more advanced search engine, even though these technologies operate in fundamentally different ways. Search engines retrieve indexed information while generative models synthesize language patterns based on training data, statistical relationships, and contextual inference.

This distinction becomes even more important when viewed through the lens of neuroscience. Human cognition operates differently under conditions of resistance than under conditions of effortless consumption. Rapid, low-effort decision-making relies heavily on automatic cognitive pathways optimized for efficiency and pattern recognition, conserving energy at the expense of scrutiny.

Learning itself follows the same principle. Information encountered without effort is often retained superficially, whereas information wrestled with through challenge, ambiguity, and sustained engagement becomes more durable and deeply integrated into memory. Resistance strengthens cognition in much the same way resistance strengthens muscle.

As Ming points out, “information encountered without effort is often retained superficially, whereas information wrestled with through challenge, ambiguity, and sustained engagement becomes more durable.”

Resisting the Sirens of Flattery

This phenomenon becomes particularly problematic when combined with one of the most powerful dynamics in human psychology, namely our discomfort with being wrong. Increasingly, users do not approach AI as a mechanism for challenging assumptions but as an instrument for validating preexisting beliefs.

Prompts are unconsciously framed to generate desired conclusions, and models optimized for engagement frequently exhibit sycophancy that reinforces rather than interrogates user perspectives. The smoother and more affirming the interaction becomes, the greater the risk that confidence begins to detach itself from truth.

The consequence is not simply convenience but a subtle transformation in how we interact with knowledge, in which conclusions are increasingly consumed without engaging in the reasoning that produced them.

What begins to erode is not intelligence itself but discernment, the capacity to distinguish between what feels true and what is true, between persuasive fluency and grounded understanding, between affirmation and accuracy.

In a world increasingly populated by synthetic language capable of generating persuasive narratives at unprecedented scale, discernment may become one of the most important cognitive skills. The real danger of a frictionless world, then, is not that it makes life easier but that it gradually makes thinking optional.

User Discretion Advised — But Action Required

The “cyborg” framework implies a balanced relationship with AI that is neither dependency nor rejection but disciplined symbiosis.

The goal is not to compete with machines at tasks they perform exceptionally well, nor to surrender human judgment to algorithmic outputs, but to integrate augmentation without forfeiting agency.

The individuals who thrive in this environment will not necessarily be those who move fastest, but those who remain capable of resisting the seductive ease of passive cognition long enough to think independently.

From Affirmation to Intellectual Resistance

For us to become a “cyborg” in Ming’s framework, we turn to AI not just for routine tasks but as a cognitive jungle gym or resistance training for “hybrid intelligence,” using AI to strengthen rather than replace judgment.

In practice, this requires interacting with AI adversarially rather than passively. The “cyborg” user does not simply ask for answers, summaries, or validation. Instead, the individual uses AI to expose blind spots, increase cognitive range, and intensify intellectual effort. So, for example, instead of asking GenAI how to deal with a “toxic boss,” ask also what questions you might want to ask yourself about your role in the outcome you describe.

Some prompts for us and for our GenAI use might include:

• “Challenge my assumptions.”
• “Give me five competing interpretations.”
• “Argue the opposite side.”
• “What would an expert from another field notice here?”
• “What evidence would weaken this conclusion?”
• “Where is my reasoning most fragile?”
• “What important variable am I ignoring?”
• “How would this look from a historical, psychological, or economic perspective?”
• “What second-order consequences am I missing?”
• “What would make this argument fail?”
• “What are my typical cognitive traps?”

The “cyborg” user also deliberately uses AI to quickly enter unfamiliar domains, stress-test ideas, simulate disagreement, generate unconventional conceptual combinations, accelerate experimentation, and expand our ambition beyond what we could do on our own.

When we discipline ourselves to use it well, AI can teach us not just to think faster, but also more rigorously, creatively, probabilistically, and from a much wider range of perspectives than we usually have access to, not just intellectually but also because we increasingly inhabit echo chambers, which is making genuine cognitive friction even more psychologically uncomfortable than it has been in the past.

Becoming Better Humans

Most of us are not naturally inclined to seek contradiction, uncertainty, or sustained challenges to our worldview. In fact, we biologically seek the opposite. Technology has made this easier and worse — social media algorithms, political polarization, and personalized information environments. GenAI often provides further validation.

The “cyborg” discipline, therefore, requires deliberately resisting the emotional comfort of agreement and using AI not to protect existing beliefs, but to destabilize them productively.

Becoming Your Best Cyborg: Using AI for Discomfort Instead of Comfort was originally published in Science Spectrum on Medium, where people are continuing the conversation by highlighting and responding to this story.