I have a mental model of what a successful human-AI collaboration may look like. A human performs a task manually, then uses AI to perform a more complex version of the task, all while understanding and being able to reason about what the AI is doing under the hood. This is not merely a preference of mine about how tools should be used. I suspect it is a more general condition for human thriving in a world where increasingly capable models are available to mediate, accelerate and even replace numerous forms of mental labour.

This is fundamentally different from a fully automated pipeline: using an AI system to brainstorm, then to generate a task spec, to execute the spec, to evaluate the work done, and finally to write docs about it. Automating the entire pipeline is acceptable only if you don’t want or need to know what the agents are doing.

But the big point here, is that it is a poor choice if you actually want to understand what is going on in the task. And that distinction, between wanting the output and wanting the understanding, is now becoming one of the decisive distinctions in the use of artificial intelligence.

The Shrinking Moat

Whatever humans are able to do using their minds and hands, AI is slowly but surely able to approach in capability and quality. The pace and modalities of progress have been improving and the quality of output from modern Agentic AI systems is compelling. AI in its current state is able to deliver the same knowledge work that humans have delivered at speed and scale, without the need to rest like humans do. In many cases the output seems to be better than that produced by humans, as well, although there are many notable situations where AI is not as good. This improvement seems an elementary aspect of AI's capabilities, but it is a crucial one in the day-to-day use of AI for numerous tasks. This has shrunk the moat in terms of what humans ought to work on in the future. We are undoubtedly in a time of massive change and transformation as a result.

For centuries, skill itself acted as a kind of natural bottleneck. If one wanted to write clearly, one had to learn to write. If one wanted to reason well, one had to learn to reason - the process of critical thinking was often learned painstakingly in schools, colleges, through writing, debate and reflection. There was a struggle to learning of all kinds. If one wanted to make a compelling image, build a reliable machine, write a competent essay, compose a score, solve a mathematical proof, or explain a software application's architecture, one had to pass through the narrow gate of practice. This narrow gate eliminated a lot of learners, because they didn't manage to strive hard or smart enough, and the gate was often frustrating, slow and ego-bruising as a result, but it had an almost civilizational function. It ensured that the ability to produce convincing work was usually correlated, however imperfectly, with some accumulation of judgment and effort. And that this accumulation of convincing work was associated with greater credibility, greater performance in society and so on. Credibility, in other words, was not free but earned through learning, experience and proof of work.

AI weakens that correlation between credibility and performance and introduces a bluff in between the effort system of learning and the credibility system based on advertising work and rewards. It allows the communication utility of work to appear during and before the production of work, because someone need not be an expert to learn something, and they need not even have true understanding or a mental model of a system. This is the world experienced in an imperative tone. When we prompt an AI model, we are living from result to result. The imperative mood skips the messy middle of human effort, and uses automation to go from thought to thing. We can now acquire the signs of competence faster than we acquire competence itself. In some domains, this may be useful. In others, it may be deeply deforming.

When humans become used to, even reliant on AI systems for tasks that they're capable of doing themselves, they undergo four very consequential things, two of which are negative consequences, the other two being positive ones:

  • Skill Atrophy, which is the wasting away of one's ability to perform a certain task, owing to underutilization of that skill, and
  • The Loss of Serendipity, where the inability or unwillingness to engage with a certain task or skill reduces one's tolerance for grappling with ideas adjacent to the tasks or skills in question, and
  • The Development of a Strong Imperative Mindset, where due to the repeated use of a direction-first mindset, where one's directing an AI agent all the time for work, we all develop and thrive as managers and shepherds of agents.
  • Deep Knowledge of Metrics and their Usage, a desirable second-order effect of AI usage, where AI users begin using proxies of performance and metrics of quality and embed these into their workflows to measure and improve their work. Note that this doesn't necessarily mean they can perform the task or evaluate the work product, but that they're able to incorporate evaluations into the process of delivering work.

Human Thriving

A month ago, when OpenClaw was all the rage, much was made of how the agents created a new website named MoltBook (a play on Facebook), and prepared for themselves a code of rules that functioned as a religion. Agency has been a more central capability to humans in the many millennia of our existence, more than in the few months that these agents have existed.

This brings me to the question of what may change in the future. We live in an age particularly of human thriving - with rapid advances in human health, technology and the like leading to the biggest demographic dividend for humanity as a whole in its entire history. If more people were an indication of thriving, 2026 is humanity's summer, its best period. But what will human thriving look like in the age of human-AI collaboration? Will we still own the core logic of the systems that comprise human society? Will our AI assistants execute within the context defined by this core, or will they begin to define the context in which we humans operate? Or, will we be orchestrated by agents?

I am deliberately using the word thriving here, rather than productivity, convenience or efficiency or other terms we may use to describe the outcomes for humanity as a whole. Productivity is too narrow a term. It answers the question: how much can be produced? It does not answer the more important question: what sort of person is being produced in the process? A civilization can become more productive while becoming less wise. An individual can become more efficient while becoming more shallow. A team can ship products faster while losing the internal ability to explain, maintain and extend what it has shipped. Human thriving has to include growth in character, depth of understanding, resilience, taste, self-command and the ability to carry responsibility.

If AI merely increases output while diminishing these things, then it does not contribute to human thriving in the strongest sense. It contributes to a narrower economic or instrumental objective. This may still be useful. But it must not be confused with flourishing.

Assistance versus Substitution

There is an important distinction to be made between assistance and substitution. Assistance occurs when the tool extends the user while leaving intact the user's own relationship to the task. Substitution occurs when the tool quietly takes over the reasoning core of the task and leaves the human in the position of approver, curator or spectator.

When a programmer uses an AI assistant to explain a Rust borrow checker error after having first attempted to understand it, that may be assistance. When the same programmer prompts an agent to produce a complete subsystem, its tests, its deployment config and its documentation without understanding the trade-offs involved, that is substitution. When a writer uses AI to critique an already formed argument, that may be assistance. When the writer uses AI to generate a viewpoint, the structure of the essay, the transitions, the examples and the conclusion without ever grappling with the ideas, that is substitution. The outward artefact may be polished in both cases. But inwardly, the human being is participating very differently.

This distinction matters because the human cost of substitution is not visible in the output. The output may look fine. It may even look better than what the human would have produced on their own that day. But if the person has not inhabited the chain of reasoning that led to it, then something important has not happened. A certain kind of cognitive metabolism has been skipped.

The Missing Sense of Accomplishment

When I use AI to build something, I often experience a sense that I have not truly accomplished the task myself. In this state, I feel compelled to brush off some of the things I have done as trivial. On the one hand, my mind is convinced that the outcome achieved with AI is important. In many cases this is true. Whether it is systems I have built, or writing I have augmented with AI, I feel that there is something accomplished as an outcome. On the other hand, I am convinced that I have not accomplished this task, and this is a weird cognitive dissonance. It is precisely because an AI assistant executed it for me, rather than me doing it myself.

This cognitive dissonance stems from a deep-seated challenge of attribution and accountability. While we mentally account for AI as just another tool, like a compiler or a text editor, its general capabilities make it difficult to attribute the final outcome entirely to our own effort. This isn't inherently wrong, but it feels unnatural. Consider how we interact with cars: we might buy a car and say, "I did up the interior differently" or "I got a wider set of tyres," even if a mechanic did the actual labor. Perhaps a generation ago, people tinkered with the engine directly, but today, we tinker differently, by delegating to specialists. AI is forcing a similar transition in knowledge work, moving us from being hands-on tinkerers to orchestrators. However, our sense of personal achievement hasn't yet caught up to this new mode of tinkering.

That mismatch between external result and internal ownership is not merely sentimental. It reveals something about how human beings relate to work. We do not derive meaning only from outputs. We derive meaning from the relationship between ourselves and the process by which those outputs come into being. There is a reason that mastery feels different from procurement. There is a reason that making a thing oneself, even imperfectly, carries a different psychic weight from having a thing made. We are not only consumers of outcomes. We are also beings who are shaped by disciplined participation in processes.

In this sense, AI can sometimes create a peculiar form of alienation. The output is ours in one sense and not ours in another. We asked for it. We directed it. We selected among options and perhaps revised them. Yet the deepest parts of the construction were not inhabited by us. The task passed through our hands, but not through our full understanding. This is why AI-mediated accomplishment can feel simultaneously real and hollow.

Friction, Learning and Character

The case for friction is often misunderstood. Friction is usually treated as a defect in a system. In consumer software and enterprise tooling alike, we are taught that the ideal experience is one where all friction has been removed. This is sensible in many contexts. Nobody wants needless bureaucracy, broken flows, bad interfaces or repetitive drudgery. But the absence of friction is not an unqualified good. Some forms of friction are pedagogical. Some are formative. Some are precisely the media through which judgment is built.

Anyone who has learned mathematics, painting, writing, music, flying, engineering or philosophy knows this. The early stages are awkward. One does not yet have the conceptual map, the confidence, the speed or the intuition. Friction is high because the self is under construction. The temptation to erase that friction with AI is understandable. But if one erases too much of it, one also erases the very sequence by which the relevant faculties are developed.

This is why I think there is danger in the current tendency to offload not only execution but also intermediate struggle. We increasingly use AI not merely for answers but for framing, decomposition, validation and even motivation. Prompting becomes a way of outsourcing not just labour, but hesitation, ambiguity and doubt. Yet ambiguity and doubt are not always enemies. They are often the terrain on which understanding is won.

Vibe coding has its place, but the associated things we're seeing in the SDLC, such as vibe engineering (where we execute prompt -> full application workflows), vibe evaluating (where we write tests with AI), vibe deploying (where we go from prompt to deployed app with infrastructure et al), and vibe documenting (where we write prompts to have AI generate documentation for an app) only to vibe-sell (write marketing campaigns for a vibe-coded app with AI) or screw things up downstream (as has happened with numerous companies who have used AI) is a perilous path. It will lead to:

  • Missed opportunities to learn and grow
  • Lost opportunities to build character
  • An inability to exercise control
  • A loss of agency and lack of a sense of accomplishment
  • An inability to connect ideas across disciplines, due to a lack of deep knowledge across multiple domains.

The danger is not that no one will be able to produce anything. On the contrary, many more people will be able to produce many more things. The danger is that fewer people will know, in a robust sense, what they are doing. This introduces a civilizational asymmetry. The visible abundance of polished outputs may increase just as the stock of real understanding becomes thinner and more concentrated.

Delegated Cognition at the Institutional Level

The question does not stop at the individual. Once AI assisted work becomes normal, institutions begin to optimize around it. A manager sees more output from a team using AI and asks for still more. A company discovers that apparently competent artifacts can be produced with fewer people and less patience. A school discovers that students can submit essays that look complete. A media system discovers that infinite synthetic content can fill every available channel. Over time, the environment itself begins to reward those who are best at coordinating delegated cognition, not necessarily those who are best at thinking.

This creates a new pressure on the individual. What begins as optional assistance soon becomes mandatory acceleration. The person who wants to learn patiently is then measured against the person who uses AI to synthesize ten plausible answers before lunch. The issue is no longer simply one of personal preference. It becomes institutional and then cultural. A society can end up selecting for the appearance of fluency while slowly disincentivizing the cultivation of depth.

This is one reason I resist the framing of AI as a neutral productivity layer. It is not neutral. It changes what kinds of effort are rewarded. It changes what counts as acceptable speed. It changes what organisations demand, what schools tolerate, what audiences expect, and what a human being begins to think of as normal work.

A Heuristic for Working with AI

To counteract these risks, we need a better heuristic for our daily interactions with artificial intelligence:

  1. Embrace the friction: Humans ought to use AI to argue with and learn from, and not merely automate the execution of things we want to do. This means learning the old fashioned way, by reading, engaging with the documentation, and building things manually. Keep the friction gradient intact, before using AI to do the same tasks.
  2. Understanding over execution: Prioritize understanding first, and execution later. Don't use AI only for your understanding, because often, many problem statements are not clear enough to be solved by AI. Use AI to understand the problem better, and then use it to execute the solution.
  3. Avoid purely "vibe" work: Don't vibe engineer, vibe execute, vibe-evaluate, or vibe-document. You wouldn't have learned anything in the process (unless learning absolutely zero was your ultimate objective, which it cannot possibly be).

I would add a fourth principle here: retain the burden of explanation. If you cannot explain why a design is correct, why an argument is persuasive, why a result was produced, or what trade-off has been made, then you do not yet own the work, even if you have possession of the output. The burden of explanation is one of the last remaining checks against the total hollowing out of cognitive labour.

A Guide for Builders of AI Assistants

If you are building AI tools for users, your primary design goal should be to foster human thriving rather than to simply replace human effort.

Consider introducing productive friction instead of unconditionally acquiescing to every single request. Rather than doing all the thinking for the user, design your assistants to explicitly guide them toward a deeper understanding of the task at hand.

Examples of such productive friction might include:

  • Requiring user rationale: Prompting the user to explain why they want a particular architectural change before generating the code for it.
  • Interactive debugging: Pointing out a logical flaw in a user's approach and asking them to propose a fix, rather than just rewriting the function silently.
  • Socratic questioning: When asked to summarize a complex topic, providing the core framework but asking the user to draw the final conclusions based on their specific context.

The point here is not to be paternalistic or to add friction for its own sake. The point is to preserve the user's agency and to strengthen their relationship to the underlying task. A truly humane AI assistant should not merely maximize compliance. It should, in certain contexts, maximize comprehension.

There is a temptation among builders of AI products to equate user delight with instant acquiescence. But a teacher who gives every answer instantly is not necessarily serving the student. A calculator that prevents a child from ever learning arithmetic may be efficient in one narrow sense and destructive in another. Likewise, an AI assistant that makes every difficult thing effortless may end up eroding the very capacities its user most needs in order to live well.

The future of human-AI collaboration should not be one where the human becomes a ceremonial approver for increasingly autonomous systems. Nor should it be one where humans refuse all leverage out of fear. It ought to be a future where leverage and learning remain in right relation. Where AI helps the human go farther, but does not relieve the human of the responsibility to understand. Where the tool is powerful, but the person is not diminished.

That, to me, is the proper thesis for human-AI collaboration: not maximal automation, not reactionary abstinence, but preserved agency under increasing leverage. If we fail to hold that line, we may indeed become more productive. We may even become more impressive. But we will not necessarily become better.