WHO CONTROLS THE CORPUS: INFORMATION POWER FROM CHARLEMAGNE TO CHATGPT
In the late eighth century, Charlemagne launched a sweeping program of educational and religious reform. A letter known as De litteris colendis urged bishops and abbots to improve the quality of clerical learning. The Admonitio generalis of 789 directed monasteries and bishoprics to establish schools. Across the empire's scriptoria, a clearer and increasingly uniform writing system — Caroline minuscule — spread under royal and ecclesiastical patronage. At Aachen, scholars including Alcuin of York worked to correct biblical and liturgical texts and create a more coherent intellectual culture for the expanding Frankish realm.
History has called this the Carolingian Renaissance. It preserved classical texts that might otherwise have been lost. It laid the groundwork for medieval scholarship. It is, by most accounts, a story of intellectual progress.
It was also the most sophisticated information control operation the Western world had seen.
More than 7,000 Carolingian manuscripts and fragments from the eighth and ninth centuries survive — a fraction of what was originally produced. What was copied wasn't neutral. What was copied depended on institutional priorities, copying demand, material resources, and the judgments of the people controlling the scriptoria. The scholars who decided which texts received attention, which were standardized, and which quietly fell out of circulation were operating inside a political project as much as a scholarly one. Charlemagne needed literate administrators. He needed a unified theological framework to hold a multiethnic empire together. He needed, in the most literal sense, a coherent version of history and doctrine that everyone from the Rhine to the Pyrenees could be taught.
The monks in the scriptoria weren't censors in the modern sense. They were something more powerful: they were the infrastructure through which all knowledge had to pass.
The Pattern
What Charlemagne built wasn't unprecedented and it wasn't the last time.
Knowledge lost through conquest, suppression, neglect, institutional collapse, and the destruction of libraries didn't disappear randomly. What survived in any given era reflects the priorities of whoever controlled the infrastructure of preservation, standardization, and distribution. The Catholic Church's dominance of manuscript production for centuries following the Carolingian period meant that survival depended heavily on ecclesiastical relevance and institutional favor. Texts that challenged doctrine faced different odds than texts that reinforced it.
The printing press, introduced to Europe in the mid-fifteenth century, was celebrated as the great democratization. And it was — briefly. Within decades, the Counter-Reformation had organized systematic responses to control what could be printed, distributed, and read. The Index Librorum Prohibitorum wasn't a desperate rear-guard action. It was an adaptation. The infrastructure changed; the impulse to control it did not.
Colonial historiography represents perhaps the most thorough execution of this pattern at scale. Entire civilizations — their histories, cosmologies, medical knowledge, legal systems — were either destroyed or rewritten into the archive of their conquerors. The history of the Americas, Africa, and Asia that entered European libraries was curated by the people doing the conquering. What survived was what served the narrative of the people who controlled the corpus.
The twentieth century brought mass media concentration. A handful of broadcast networks, newspaper chains, and publishing conglomerates determined what millions of people understood about the world. The dream of the open internet was, in part, a reaction to this — the promise that no single entity could again control the infrastructure of information.
That promise has not aged well.
The New Scriptorium
Today, the pipelines that transform the world's information into widely used artificial intelligence systems are controlled by a small number of companies, while the datasets, filtering decisions, and training procedures behind the most capable models remain largely invisible to the people who use them.
This invisibility is not a bug. It is the nature of the infrastructure.
What makes the current moment historically distinct is the feedback loop. Previous information control systems were relatively static — a manuscript, once copied, didn't rewrite itself. AI systems are different. Models train on internet content. Their outputs increasingly become internet content. The next generation of models trains on that combined corpus. Stanford's 2026 AI Index reports that industry produced more than 90% of notable frontier models in 2025, and that the most capable systems are often the least transparent about their datasets, parameters, and training procedures. Researchers have documented that training models recursively on their own outputs can produce "model collapse" — a degradation of the underlying knowledge base over successive generations. Errors, biases, and distortions don't just persist. They can compound.
The research community has a name for the deliberate version of this: data poisoning. And it is no longer theoretical.
In a 2025 Nature Medicine study, researchers found that replacing just 0.001% of a model's training tokens with medical misinformation measurably increased harmful outputs while leaving standard benchmark performance largely unchanged. In one experiment, 2,000 malicious articles — generated for about five dollars — were enough to affect a four-billion-parameter model. The researchers estimated that a comparable attack against a much larger training run could require roughly 40,000 articles costing under one hundred dollars. Separate research in other machine learning settings has shown effective poisoning attacks involving approximately 0.1% of a dataset. The common thread: even small, hidden manipulations can survive curation and testing, only to resurface later as persistent behavioral changes that are difficult to detect and harder to reverse.
The Carolingian monks needed armies and papal authority to standardize the corpus. The modern equivalent may require a few thousand well-placed documents and patience.
The Concentration Problem
The technical vulnerability matters less in isolation than in context. Data poisoning is concerning as a cybersecurity threat. It becomes something qualitatively different when considered alongside the documented concentration of AI infrastructure in the hands of a small number of actors with explicitly held views about governance, democracy, and the relationship between information and power.
The World Economic Forum's Global Risks Report 2026 placed mis- and disinformation among the top short-term global risks globally — describing it as "seemingly the risk that catalyses or worsens all other risks." Evidence from the 2024-2025 electoral cycle shows how AI systems optimized content for maximum emotional impact across multiple countries. This is documented. It is not projection.
The recently leaked records of the private Dialog network — a gathering of figures from government, technology, finance, intelligence, and AI-adjacent industries — do not establish a coordinated operation to manipulate AI training data. What they do illustrate is how densely political, technological, financial, and governmental power can overlap outside normal public visibility. That overlap matters when the same sectors increasingly shape the infrastructure through which information is processed and distributed.
We are not claiming that coordinated corpus manipulation is currently underway. The honest position is more unsettling than that: the infrastructure that would allow it to happen largely without detection already exists, is already concentrated, and the process would be, by design, difficult to distinguish from organic bias emerging from who builds these systems and what they value.
The monk in the scriptorium didn't need to be malicious. He just needed to copy what he was told to copy, omit what he was told to omit, and trust that the people giving the orders had good reasons.
The Methodology Question
There is a counter-strategy, and it is not new.
Every era of information control has produced its own forms of resistance — the samizdat manuscripts of the Soviet Union, the underground presses of occupied Europe, the oral traditions that survived colonial erasure by staying outside the written record entirely. What these have in common is not technology. It is methodology: explicit sourcing, redundant preservation, transparent confidence levels, and the willingness to label what is known versus what is suspected.
The historians who have done the most durable work on understanding Charlemagne's information project didn't do it by trusting the Carolingian manuscripts uncritically. They did it by triangulating across sources, noting what was absent, and treating the shape of the silence as evidence.
That same methodology applies here. The questions worth asking about AI training data are not primarily technical. They are the same questions any historian asks about any archive: Who assembled this? What were their interests? What is conspicuously absent? Who benefited from what was included and what was left out?
The internet promised to make everyone a publisher. It may have instead made everyone a monk — producing content that enters a corpus controlled by others, shaped by others, and deployed in ways that serve ends we cannot fully see.
Charlemagne's court scholars saved a great deal of classical knowledge. They also standardized an empire. Both things were true simultaneously.
The question worth sitting with is which of those functions the current infrastructure is primarily serving — and whether, a thousand years from now, the historians will be able to tell the difference.
