DUBAI—In his seminal 1956 paper “The Magical Number Seven, Plus or Minus Two,” American psychologist George Miller made a deceptively simple argument: our working memory can hold only seven pieces of information at once. In effect, Miller identified a hard constraint on the human mind’s processing capacity, showing that short-term cognition operates within surprisingly narrow limits.
At roughly the same time, the Nobel laureate economist Herbert A. Simon arrived at a strikingly similar conclusion. His theory of bounded rationality held that decision-makers never optimize in the sense that classical economics imagines, because cognition itself is a scarce resource. Faced with more variables than they can simultaneously process, human beings do not search for the best possible answer. Instead, they settle for an answer that is good enough within the limits of their cognitive resources. As Simon put it, “A wealth of information creates a poverty of attention.”
In the 1980s, educational psychologist John Sweller pushed this logic further with his cognitive load theory. Sweller’s theory holds that when informational demands exceed the limits of working memory, the mind becomes overwhelmed and performance deteriorates.
Cognitive Capitalism
Despite approaching the problem from different directions, Miller, Simon, and Sweller all described the same underlying condition: a widening gap between the complexity of modern societies and people’s cognitive capacities.
It was against this backdrop that heterodox economists in the early 2000s began to argue that capitalism was entering a new phase. In an influential 2005 paper, Carlo Vercellone—an economist at Université de Paris 8 Vincennes-Saint Denis—drew on Karl Marx’s concept of the “general intellect” to contend that collective human intelligence has displaced the factory as the central engine of value creation, giving rise to what he termed “cognitive capitalism.”
But Vercellone’s thesis extended beyond the rise of the knowledge economy. Capital, he argued, could never fully own or control the most productive dimensions of cognition: tacit knowledge, relational judgment, and lived experience. Unlike machinery, knowledge could not be fully separated from the workers who possessed it and thus could not be codified into procedures or transferred at will. Cognition, in his view, remained the one productive input resistant to complete commodification precisely because it was irreducibly human.
Vercellone, of course, could not have foreseen the rise of AI. Before the emergence of large language models, the limits of codifying human cognition manifested as informational overload: too many variables to process, too much data to interpret, and too much complexity for workers and decision-makers to navigate. Today, however, what once seemed inseparable from human intellect can increasingly be extracted, replicated, and deployed at scale.
In a sense, AI has introduced a form of cognitive compression, or “zipping,” that converts tacit human understanding—once confined to individuals and institutions—into something that can be sold by the token. As a result, a new market has emerged around what Vercellone believed could never be commodified: human cognition itself.
The Market for Human Thought
From the clay tablet and the printed book to the encyclopedia and the search engine, humans have sought to externalize cognition. But while a textbook contains knowledge and a database stores information, both remain passive repositories. AI systems, by contrast, can perform increasingly complex cognitive tasks without continuous human guidance.
Cognitive science provides a valuable lens for understanding what AI firms actually package and sell. British psychologist Alan Baddeley’s model of working memory is particularly useful in this regard, showing that human cognition consists of several distinct but interconnected functions.
Baddeley’s model identifies three main cognitive layers. The first is the capacity to hold and process verbal information, which allows us to read a sentence, retain it, and build on it. The second is the ability to synthesize information from multiple sources, integrating disparate inputs into a coherent whole. And above both sits what Baddeley called the “central executive”: the part of the mind responsible for setting goals, allocating attention, responding to new situations, and deciding what matters and why.
These layers are not interchangeable, and they are being commodified in very different ways. AI can now reproduce the first two, processing and synthesizing vast amounts of information without the constraints of human memory and at a fraction of what companies currently pay research analysts, policy advisers, and strategy teams. What it cannot yet do is define priorities, exercise judgment, and navigate uncertainty.
The real disruption lies beyond routine automation, in the growing commodification of cognitively demanding, knowledge-intensive labor. While workers whose economic value lies primarily in processing and synthesizing information are increasingly vulnerable to AI-driven displacement, those whose value lies in directing and overseeing complex processes remain far less vulnerable—at least for now.
But obsolescence may pose a bigger threat to business models built on selling cognitive labor by the hour than displacement, as workers and organizations move from the demand side to the supply side of the cognitive market. An independent researcher who sells their analytical capacity by the hour could serve multiple clients at once without working more hours by packaging their expertise into an AI-augmented product. Likewise, a small consultancy that relies on analytical depth alone could harness AI tools to operate at a scale previously available only to much larger firms.
The New Knowledge Economy
The knowledge economy was conceived for a world in which expertise commanded a premium, not one in which cognition itself has become infrastructure. This shift marks the emergence of a new form of economic participation, defined less by performing cognitive labor than by controlling the systems that package, scale, and distribute it.
The cognitive market, however, is not a single market. It has two distinct layers governed by conflicting logics. At the lower layer, organized around the general cognitive functions Baddeley identified, something resembling what Vercellone and his co-authors called “commonfare” is beginning to emerge, though not by political design.
Developed across a series of collaborative works during the 2010s, the commonfare concept rests on the idea that cognitive value is produced collectively and draws on a shared intellectual inheritance. Therefore, the systems that generate and distribute it should be governed as a common good rather than as private property.
Market forces are now producing a version of this on their own. Once an AI model is trained, adding users costs relatively little, incentivizing firms to scale as aggressively as possible. Unsurprisingly, the most reproducible layers of cognition are also the ones being rapidly democratized: a researcher in Nairobi can perform many of the same analytical tasks as a partner at McKinsey.
Yet the same technology that democratizes access to expertise is also eroding the scarcity on which many professionals built their careers. The tacit knowledge of millions of analysts, researchers, lawyers, and advisers—the very thing Vercellone argued could never truly be separated from human judgment and lived experience—has been harvested to train AI models owned by a handful of corporations.
Capital, in this sense, has not merely commodified human cognition. It has appropriated it, much like the enclosure of common land once dispossessed those whose labor had made it valuable.
This dynamic is already visible at the cognitive market’s specialist layer, where AI systems are being trained to replicate highly specific forms of expertise: the oncologist’s ability to recognize rare patterns, the geopolitical analyst’s strategic intuition, and the M&A lawyer’s feel for hidden risk.
Such hard-won capabilities are now being systematically extracted and encoded into commercial AI products like Bloomberg’s financial-analysis models and the legal-services platform Harvey. These are proprietary systems built by capturing and concentrating forms of cognitive capital that were unevenly distributed in the first place.
The cognitive market is therefore democratizing at the bottom while concentrating at the top. Those most vulnerable are mid-level knowledge workers—analysts, researchers, junior professionals—whose market value derives from the cognitive functions AI can perform more cheaply. The primary beneficiaries are those already at the top of the cognitive hierarchy, who can now monetize their expertise at a scale they would never have achieved on their own.
The Scarcest Commodity
Whether or not one accepts their broader political premise, Vercellone and his co-authors raise a fundamental question: If the knowledge used to train AI systems is collectively produced, who is entitled to the returns?
That question becomes even more pressing as cognitive compression improves over time. In a recent working paper, Nobel laureate economist Daron Acemoglu and his co-authors argue that when AI systems consistently outperform average humans, the incentives to develop and maintain expertise begin to erode. After all, why spend years developing internal analytical capacity when you can rent it?
The result is what the authors call “knowledge collapse.” The more effective cognitive zipping becomes, the less incentive institutions and individuals have to preserve the raw material on which it depends. In their telling, analytical capacity risks atrophying not because AI has failed, but because it has succeeded to such an extent that society no longer feels the need to cultivate it.
None of this is inevitable, provided institutions follow three key principles. The first is to adopt AI gradually while simultaneously developing the internal capacity to challenge, supervise, and evaluate increasingly agentic systems before entrusting them with consequential tasks.
The second principle is to bolster institutional decision-making. While AI can handle tasks like reading, synthesis, and pattern recognition far more efficiently than humans and with relatively little risk, strategic judgment and interpretive reasoning are different. Those capacities must be deliberately maintained and strengthened, not quietly handed over to machines.
Lastly, institutions must invest in oversight. The ability to question, verify, and interpret AI outputs is itself a cognitive skill, and an organization that cannot tell good analysis from bad effectively loses the capacity to govern itself.
What Marx Got Right
In his unfinished notebooks from the 1850s, known as the Grundrisse, Marx predicted that collective human knowledge would become the central productive force in capitalist economies. What he did not foresee was that this knowledge would be compressed, priced by subscription, and sold back to institutions drowning in more information than they could process. The cognitive surplus that organizations can no longer absorb has become, in effect, the commodity being bought and sold.
This emerging economic reality not only reshapes how institutions function but also determines which forms of cognitive labor retain value, which kinds of expertise remain scarce, and who gets to participate in the market as sellers. Once the cognitive functions required for routine knowledge work become universally available at negligible cost, what remains genuinely scarce is that which cannot be reproduced: the ability to ask the question no one thought to ask, see what others have overlooked, and know what to do when there is no precedent.
The real risk, as Acemoglu and his co-authors note, is that institutions may not recognize what they are losing until it is too late. As AI assumes more of the analytical burden, the distinctly human capacities at the top of Baddeley’s model of the mind could atrophy. Over time, we may lose the ability to make sense of complexity by ourselves.
Yet those capacities are also becoming more valuable than ever. In a global economy where AI can package, replicate, and distribute knowledge at near-zero marginal cost, the ability to think independently may soon become the scarcest commodity of all.
Sami Mahroum, Founder of Spark X, previously held posts at INSEAD, the OECD, and Nesta.
Copyright: Project Syndicate, 2026.
www.project-syndicate.org
