There is no shortage of AI content. The technical guides cover the how. The news creates urgency and, at its best, forces the kind of thinking that urgency alone rarely produces. That layer has its value.
What is largely missing is something deeper: not what AI can do, but what it actually is. What it changes about how decisions get made. What it demands from the people who lead through it. No workflow guide or LinkedIn thread gets there.
Books do - because of what happens when one is actually read. Not summarised. Not skimmed as key points. Read. There is something qualitatively different about following a full argument across 300 pages: sitting with the counterarguments, letting the reasoning accumulate, arriving at a conclusion that feels earned rather than handed over. A summary gives familiarity with an idea. A book gives the capacity to think with it.
None of the five books below are about AI. Several predate the internet. What they share is a set of ideas about emergence, feedback, systems, fragility and judgement that cuts closer to the present moment than most things written about AI this year.
Understanding Media - Marshall McLuhan
McLuhan's aphorism - "the medium is the message" - is widely quoted and rarely understood. His actual argument is worth sitting with.
When a new medium appears, people instinctively evaluate it through the lens of whatever came before. The first cars were called horseless carriages. The first television programmes were filmed radio. The early internet was used to replicate print. In each case, the inherited frame captured something real but missed what was actually changing - not the content the medium carried, but the environment it created. Television did not just deliver different content than radio. It reorganised how authority presented itself, how families spent evenings, how an entire generation understood the relationship between image and trust. Those effects were underway before anyone noticed them.
AI is a medium in exactly this sense. Asking whether it can match human performance on existing tasks is the horseless-carriage question - judging the new thing entirely in terms of the old thing it most resembles.
McLuhan called this "rear-view mirror" thinking: driving into the future while looking backwards, measuring the new medium against the standards of the previous one.
The more structural question is what the new environment makes possible that was not possible before - and what it quietly changes about how thinking and decision-making actually happen.
AI does not just change what we do, it changes what we notice, what we reward and what we gradually stop practising. That is the medium's real message.
The more interesting question the medium invites is not how to do existing things faster. It is what becomes possible that was not seriously attemptable before - new ways of understanding how someone moves through an experience in real time, new relationships between content and the context it lands in, new ways of staying relevant to an audience whose expectations are being quietly reshaped by tools they use every day. None of that is visible from inside the rear-view mirror. It comes into focus when the question shifts from "how do we use this?" to "what does this make possible that wasn't before?" - and that is the question worth building strategy around.
Out of Control - Kevin Kelly
Kelly's answer, written before the web existed, is that AI is an emergent system. And emergence changes the entire frame.
The core idea: large numbers of simple components, following simple rules, can produce behaviour that no individual component could produce or predict. A single ant follows a handful of chemical signals and contributes nothing that looks like intelligence. A colony of a million ants builds climate-controlled chambers, maintains fungal farms and coordinates logistics at a scale no individual ant planned or comprehends. The intelligence does not live in any part - it emerges from the interactions between parts. Large language models work on exactly this principle. Kelly himself, rereading the book recently, observed that today's LLMs are "the prime example of the bottom-up, distributed, decentralised, emergent systems" he described three decades ago.
Treating AI as a deterministic tool means trying to control it input by input - and being perpetually surprised when outputs diverge from expectations.
Treating it as an emergent system shifts the design question: not "how do I specify the output?" but "what conditions produce useful emergence?"
Kelly's phrase - give up control to gain control - is not a paradox. It is a description of how emergence gets managed in practice.
That shift in frame changes what feels worth paying attention to. Rather than asking why a particular output missed the mark, the more useful question is what conditions produced it - the context provided, the constraints in place, the feedback that did or did not exist. The organisations getting the most consistent value from AI are not necessarily the ones with the most optimised workflows. They are the ones that have developed a feel for the environment they are asking the model to operate in - and learned to shape that environment with the same care they bring to the output itself.
The Human Use of Human Beings - Norbert Wiener
Wiener founded cybernetics - the study of feedback and control in systems - and this book is his attempt to explain what that means for the relationship between people and machines. Written in 1950, when the most powerful computers filled entire rooms, it has aged with unnerving accuracy.
The central concept is the feedback loop: a system that receives information about the results of its own actions and adjusts accordingly. What Wiener recognised early is the asymmetry at the heart of human-machine collaboration. Machines are extraordinarily capable at tasks that are repetitive, specifiable and measurable. What they cannot navigate without human guidance is situations where the rules are ambiguous, the stakes are relational and the right answer depends on context that cannot be fully written down.
The people extracting the most value from AI are not necessarily those who interact with it most efficiently - they are the ones who give the best feedback, who can sense when a plausible output is subtly wrong.
That capacity for calibrated, high-quality feedback is a distinctly human competence, and it is the primary leverage point in any human-AI collaboration.
Wiener's deeper argument quietly dismantles the anxiety underneath most AI commentary: human judgement is not slower, more expensive computation - it is something different in kind, navigating situations that cannot be reduced to calculation at all.
The skill that compounds most in an AI-assisted environment is not tool fluency. It is the capacity to read something that sounds right and know whether it actually is. That takes domain depth - the kind that only comes from years of doing the work. The people who tend to get the most from AI over time are the ones who brought the most genuine expertise to it. The feedback loop Wiener described runs in both directions. What comes out is shaped, more than most people expect, by the quality of judgment brought to the evaluation.
Thinking in Systems - Donella Meadows
Meadows was a systems dynamics researcher at MIT, and this book is one of the clearest accounts of why complex systems consistently produce outcomes nobody designed and nobody wanted. Two of her ideas are particularly sharp for anyone navigating AI adoption.
The first is the trap she calls "shifting the burden." A difficult underlying problem exists. A faster intervention treats the symptoms without touching the root cause. The symptoms ease, the pressure to address the real problem dissipates, dependency on the workaround grows - and the underlying capability quietly atrophies. When analysis, synthesis and complex writing are consistently handled by a model, the daily output looks fine. What erodes is the capacity to do that work independently: the ability to form a considered view, to sense a flawed assumption inside a well-constructed argument, to know when something sounds right but is actually wrong.
The second is her distinction between stocks and flows. Stocks accumulate over time - domain expertise, professional judgement, the pattern recognition built from years inside a problem space. Flows move through a system - daily output, tasks completed, decisions processed.
AI dramatically accelerates flows - daily output can increase by an order of magnitude.
What it can simultaneously do, if used without intention, is draw down the stocks - the accumulated capabilities that took years to build and are not rebuilt quickly.
The question Meadows forces is not "how do we use AI to produce more?" but "what are we preserving while producing more?" Most organisations are only asking the first.
The organisations in the strongest position in five years will likely be the ones that treated capability preservation as a design choice rather than something to think about later.
Antifragile - Nassim Nicholas Taleb
Taleb draws a distinction between three types of system: those that break under stress (fragile), those that resist stress without changing (robust) and those that actually improve when exposed to stress, volatility and uncertainty (antifragile). The third category describes the only genuinely durable position in an environment that keeps changing.
The AI landscape changes fast enough to make specific technical knowledge obsolete within months. A workflow optimised for one model becomes redundant when the next is released. Attaching professional identity or organisational strategy to mastery of specific tools is not a conservative position. It is a fragile one.
Taleb's prescription is the barbell: invest heavily at the two extremes and avoid the middle. At one end, the things that do not expire - domain expertise, sound judgement, the ability to evaluate an argument, the taste to know what good looks like. These capabilities become more valuable as AI raises the volume of mediocre output; genuine judgement becomes scarcer as generated content becomes abundant. At the other end, aggressive experimentation with new tools and approaches - small bets with bounded downside and open upside.
What to avoid is the middle: moderate investment in any specific tool or workflow that feels safe but is one model update away from irrelevance.
The most durable position in an AI-saturated environment is not deep expertise in the tools themselves - it is the combination of capabilities AI cannot easily replicate, paired with enough fluency to know when and how to apply it.
The barbell keeps both ends strong. The soft middle is where the fragility accumulates without announcing itself.
What Taleb's frame makes visible is the middle - the comfortable attachment to a specific workflow that feels productive but is slowly accumulating fragility. It is worth knowing where that sits.
What connects these five books across seven decades is a consistent argument: the most useful response to a powerful new capability is not mastery of the capability itself. It is understanding of the system it operates within.
McLuhan shows why the obvious frame for a new medium is almost always the wrong one. Kelly shows that AI is an emergent system, which changes how it should be directed. Wiener shows that feedback quality is the primary human leverage point in any human-machine collaboration. Meadows shows where the invisible costs accumulate when that is ignored. Taleb shows how to stay durable when the environment keeps changing faster than any specific playbook can keep up with.
None of them are about optimising for output. All of them are about thinking. In a moment that rewards the latter far more than the former, that distinction is what separates the strategies that compound from the ones that don't.
Learn from us
Join thousands of other Product Design experts who depend on Adrenalin for insights
