Using a large AI model can feel like encountering a sudden break with the past. We ask a question in ordinary language, and a machine can explain a scientific concept, write code, translate a text, analyse a contract, propose a software architecture or discuss philosophy.
The breakthrough is real. But it did not come from nowhere.
Contemporary artificial intelligence is the culmination of an extraordinary accumulation of technologies, scientific discoveries, public and private infrastructure, and knowledge produced by billions of human beings.
To understand the political question now facing us, it helps to reconstruct that accumulation, layer by layer. If AI becomes essential infrastructure, one question becomes difficult to avoid:
To what extent should a technology with such deeply collective origins be controlled by a very small number of private actors?
First layer: the machine
Before artificial intelligence, we needed computers. That required an understanding of electricity, the development of electronics, and the invention of transistors, integrated circuits, microprocessors, electronic memory and storage devices.
We then had to learn how to manufacture billions of transistors at microscopic scales, with a precision that would have been unimaginable a few decades earlier. And we had to multiply their computing power.
Graphics processing units, or GPUs, were originally designed to carry out large numbers of calculations in parallel for images and video games. They proved particularly well suited to the mathematical operations used by neural networks.
Training the largest AI models depends on vast data centres containing thousands of these processors, connected by high-speed networks and supplied with substantial amounts of electricity.
The first layer of artificial intelligence is therefore industrial infrastructure.
Metallurgy, chemistry, solid-state physics, semiconductors, electricity generation, telecommunications, optics, cooling systems and precision manufacturing: a remarkable share of twentieth-century industrial history is, in a sense, condensed into a modern computing facility.
Second layer: learning to program the machine
Hardware alone is not enough. Generations of computer scientists and engineers created the abstractions that turn electrical signals into usable systems: programming languages, compilers, operating systems, files, databases, network protocols, graphical interfaces and software libraries.
A substantial part of this software infrastructure is developed and shared under free or open-source licences. Unix influenced generations of operating systems. Linux and countless open-source libraries underpin much of the world’s computing infrastructure.
Languages, protocols and formats became standardized. And computers learned to communicate.
Third layer: connecting humanity
The Internet connected computers around the world. The Web then made that infrastructure a widely accessible system for publishing and retrieving information.
Tim Berners-Lee proposed the World Wide Web at CERN in 1989 to make it easier for researchers to share information. In 1993, CERN released the Web software into the public domain, a decision that helped enable its global expansion. CERN recounts that history here.
That choice deserves to be remembered. One of the most consequential economic and cultural infrastructures in modern history rests on protocols that were not locked behind their inventor’s exclusive ownership.
The Web then helped bring about something that would become crucial to AI decades later: an enormous body of human intellectual work became digital and machine-accessible.
Books, scientific papers, newspapers, photographs, software, conversations, encyclopedias, technical documentation, music, videos and archives gradually entered this global information space. Without knowing it, we were building part of the training material for future AI systems.
Fourth layer: teaching machines to learn
Another history was unfolding alongside these developments.
The idea of reproducing aspects of intelligence in machines is old. But 1956 is an important landmark: at the Dartmouth Summer Research Project, John McCarthy and other researchers helped establish artificial intelligence as a distinct field. Dartmouth looks back at that meeting.
Over the following decades, different approaches developed. Some sought to reproduce reasoning through explicit symbolic rules. Others tried to make machines learn patterns from examples. Neural networks belong to the latter tradition.
An important advance was the spread of methods that allowed networks to adjust their parameters in response to errors. In 1986, David Rumelhart, Geoffrey Hinton and Ronald Williams published a now-classic paper on learning through backpropagation. Their paper appeared in Nature.
The central idea was striking: instead of explicitly programming every rule needed for a task, we could make a machine learn patterns implicitly by showing it enough examples.
For a long time, however, three ingredients were still missing at scale: computing power, data and sufficiently effective architectures.
Fifth layer: the data
The growth of the Web and the mass digitization of information gradually changed the situation. Researchers gained access to quantities of data that would once have been unimaginable.
ImageNet illustrates this transition. In 2009, its researchers described a database of 3.2 million annotated images across thousands of categories. Building it involved researchers and computing infrastructure, but also human workers using Amazon Mechanical Turk to classify images. The original ImageNet paper describes that process.
This detail matters. Machine learning depends on considerable amounts of often invisible human labour: classification, annotation, evaluation, correction and, more recently, assessments of model responses.
Machines learn because people have produced the information from which they learn.
Sixth layer: the architecture
In 2017 came an innovation that would contribute directly to the current expansion of AI: the Transformer.
In Attention Is All You Need, a team of researchers working at Google proposed an architecture built around attention mechanisms, allowing training to be parallelized much more effectively. Read the original paper.
This innovation did not create modern AI on its own. It provided a particularly effective way to bring together decades of neural-network research, massively parallel processors, large data centres, enormous digital corpora, distributed computing software and enough capital to operate the whole system.
Models could now be trained on unprecedented quantities of data.
Seventh layer: scale
A further idea gained ground: instead of building a different model for every problem, why not train a very large model on a wide range of data, then use it for many different tasks?
GPT-3 offered a striking demonstration in 2020. With 175 billion parameters, it could perform a wide variety of language tasks from task descriptions and a few examples, without being retrained for each one. Read the GPT-3 paper.
In 2021, a Stanford team proposed the term foundation model for models trained on broad data that can serve as the basis for many applications. The researchers also highlighted the consequences of many systems depending on a small number of underlying models. Read the Stanford report.
This is where the problem reaches beyond computing.
Eighth layer: behind the data, human work
Saying that a large model was simply “trained on the Internet” misses something essential. The Internet did not create what it contains.
Behind a physics textbook lie centuries of scientific inquiry. Behind a computer program lie decades of work on languages, algorithms and methods. Behind a novel lie a language, a culture and a literary tradition that may stretch back millennia.
Behind an encyclopedia are thousands of contributors. Behind an answer on a forum may be decades of professional experience. Behind a photograph stand a photographer, a subject, a visual culture and two centuries of photographic technology.
A contemporary model draws on fragments of these productions and learns statistical patterns from them.
This does not mean a model contains all of humanity’s knowledge. Training corpora represent only part of it, unevenly distributed across languages, cultures and fields. Learning patterns from those corpora does not guarantee understanding or accurate answers.
Speaking of “millions of years of human work” evokes the scale of this inheritance, but it is not an established estimate. What we can say without assigning a number is that these systems depend on intellectual labour extending far beyond the teams that train them.
Over centuries, we accumulated an intellectual inheritance. Over decades, we digitized it. Through the Web, we connected it. Eventually, we built machines capable of extracting patterns from it at a previously impossible scale.
The final layer: capital
There is still another layer. Turning these resources into frontier models requires substantial investment. This constraint varies across systems: smaller models can be developed and used with far fewer resources.
At the frontier, however, expensive processors, data centres, energy, specialist engineers and the ability to sustain lengthy training runs all matter. This creates a paradox:
The intellectual contributions are widely distributed; the ability to turn them into extremely powerful models is highly concentrated.
Billions of people have contributed directly or indirectly to the information from which these systems learn. Public institutions have funded the fundamental research that makes them possible. Open-source communities have produced much of their software infrastructure.
Yet some of the most powerful models can belong to a handful of companies whose decisions are made by relatively small groups of executives, investors and engineers.
From ownership to governance
This does not necessarily mean that all AI should be nationalized, or that private ownership of models is inherently illegitimate. The companies building these systems take real risks, invest substantial sums and produce important innovations of their own.
Acknowledging their contribution does not require us to ignore everything that came before it. We can frame the question differently:
How much of the value created by AI should legitimately belong to those who assembled the final layer, and how much should be understood as deriving from a collective inheritance?
That distinction could become essential. AI may not remain a product comparable to office software or a search engine. It could become cognitive infrastructure: an intermediary through which people learn, write, program, design, administer, research, diagnose, produce and make decisions.
At that point, asking who owns the model leads to a larger question:
Who controls the infrastructure through which a society produces and uses its intelligence?
Sharing this power
One path treats AI as a new private industry. Companies build the most capable systems, users pay for access, and markets gradually determine how those systems are distributed.
Another recognizes that beyond a certain degree of capability and generality, these systems become a form of collective infrastructure.
That recognition does not settle the solution. Acknowledging a collective inheritance raises a political question; it does not, by itself, establish an ownership regime or determine who would represent humanity.
It could lead to open models, public computing infrastructure, publicly or university-funded models, interoperability requirements, universal access to certain capabilities, redistribution mechanisms or new forms of international governance.
The first step may simply be to recognize what we have built. Contemporary AI did not appear from nowhere. It is the latest layer of a collective undertaking that began long before the companies now commercializing it existed.
If it becomes a fundamental infrastructure of the twenty-first century, we will eventually have to decide whether this extraordinary collective inheritance should produce widely shared power, or an unprecedented concentration of it.
A historic bifurcation
We may be approaching one of the most important turning points in human history.
Not simply because we have created a powerful new technology, but because we have created a tool capable of intervening directly in the production, organization and circulation of knowledge itself.
Agriculture transformed our relationship with food. Printing transformed the circulation of ideas. Steam power and electricity transformed our ability to act on the physical world. The Internet transformed our ability to communicate.
Artificial intelligence could transform something more fundamental still: our collective ability to think, design, decide and create.
This turning point is therefore institutional and political as well as technological.
Will we treat this new capability as a collective inheritance whose benefits should be widely accessible, or as a strategic resource controlled by those who own the infrastructure and capital required to exploit it?
Decisions made over the coming years could determine which products we use, but also how power, knowledge and the capacity to act are distributed for much of the twenty-first century.
Perhaps that is the real question artificial intelligence raises.
Not simply how far it can go.
But who will decide the direction we take with it.