Brainwashed at Scale

Fact deciders are fate deciders. Will We Hold LLMs Accountable?

Every technology that promised to connect us has also been used to control us. The printing press spread scripture and propaganda in the same century. Radio gave us music, and in 1994 it gave us Rwanda genocide, where a single station read out names and directions and helped turn neighbors into killers. [^1]

Search engines gave us the world's information, and they gave us search engine bombing, where whoever gamed the algorithm best got to define reality for a billion people.

Now we have large language models. And we are about to repeat the pattern at a scale we have never seen before.

Here is the uncomfortable truth (😉 AI influence here again). AI is genuinely beneficial but it is also the single biggest threat to human agency and human history that we have ever built. Not because it is evil, but because it does not know what it does not know while it knows so much that those whose it does not know it does not know.

AI bombing is the new search engine bombing.

Twenty years ago, if you wanted to control a narrative, you gamed Google. You built link farms, stuffed keywords, and flooded the web until your version of the story ranked first. We called it Google bombing. It worked because most people never scrolled past the first page.

Today you don't need to rank first. You need to get into the training data.

If I can flood the internet with enough documents saying something, repeated across enough websites, in enough formats, over enough time, I can get an AI to believe it. Not to rank it. To believe it. To state it as fact, in a warm and authoritative voice, to hundreds of millions of people who have stopped checking sources because the machine sounds so sure.

This is AI bombing. And unlike search results, which at least showed you ten links and let you choose, an LLM gives you one answer. One synthesis. One truth. And all the references, coming from my articles deliberately planted.

Whose truth?

Truth is becoming a race, not a verdict.

In the world of LLMs as is today, truth will not be decided by who is factually correct. It is decided by who got there first, who wrote the most, and who wrote it in English. Once a narrative gains ground in the training data, the smaller player can no longer fight it. You cannot out-publish a machine that has already internalized the other side's story and now repeats it, politely and endlessly, to everyone who asks.

We have seen this movie before. History written by the victors. Cultures described by their colonizers. Entire peoples defined by what outsiders wrote about them, because outsiders controlled the presses, the archives, and the universities. These were the raw ingredients of supremacy thinking, of racism, of imperialism. When one group's account of reality becomes the only account available, the other group stops being a people with a story and becomes a footnote in someone else's.

That dynamic triggered wars. It justified conquest. It fueled xenophobia and, at its worst, the desire of one group to erase another entirely.

We are now automating that dynamic and shipping it to every phone on earth. So the question is not whether this is dangerous. The question is what we are going to do to prevent this new world order from calcifying while it is still soft.

Why do LLM get so much wrong?

To answer that, you have to understand what an LLM actually is, because most people, including most policymakers, do not.

A large language model is as good as the info it gets on the internet - plus the prediction engine plus some extra data cleaning work that the LLMs are, in fairness, investing a lot on. not a database of facts. It was trained by reading a massive slice of the internet and learning one skill: given a sequence of words, predict the next word. Do that billions of times, across trillions of words, and something remarkable emerges. The model learns grammar, style, reasoning patterns, and a compressed, statistical impression of everything it read.

This has two consequences that matter enormously.

First, the model is only as good as its diet, its data. If the training data says something a thousand times and the correction appears twice, the model learns the error. If a topic barely appears at all, the model fills the silence with plausible-sounding invention. We call these hallucinations, which is a gentle word for a machine confidently telling you fables.

Second, the model has no concept of importance beyond frequency. It does not know that a Yoruba oral history matters as much as a Wikipedia edit war. It only knows which one it saw more of.

And what did it see more of? English. When GPT-3 was trained, more than 90 percent of its training tokens were English, even though native English speakers are only about 5 percent of humanity.[^2] English is also the single largest language on the web, close to half of all website content, a share wildly out of proportion to the number of people who actually speak it.[^3]

Meanwhile, in the giant web archive most models learn from, dozens of languages each make up less than one-hundredth of one percent of the data, and around a hundred languages sit below a tenth of a percent.[^4] The languages spoken by over a billion Africans, Yoruba, Hausa, Igbo, Swahili, Amharic, Wolof, together occupy fractions of a single percent. An entire culture can be represented by less text than a mid-sized American forum generates in a month.

This is corpus disparity. It means the machine's picture of the world is not the world. It is the world as documented, and documentation has never been evenly distributed.

Sickle cell tells you everything about what gets solved.

If you want to see where this leads, look at medicine.

Sickle cell disease affects millions of people, and somewhere between 300,000 and 400,000 babies are born with it every year. More than three quarters of those births are in sub-Saharan Africa, and Nigeria alone carries the heaviest load on earth, with roughly 150,000 affected newborns a year, compared with about 2,000 in the United States.[^5] It was also one of the first diseases we ever understood at the molecular level. In 1949, Linus Pauling and his colleagues showed that the hemoglobin of people with sickle cell was physically different, and called it the first "molecular disease," four years before we even knew the structure of DNA.[^6]

Three quarters of a century later, it remains under-researched, under-funded, and largely unsolved for the people who actually have it. When cures finally arrived in 2023, the two approved gene therapies were priced at 2.2 and 3.1 million dollars per patient, for a condition whose sufferers are overwhelmingly in countries where almost no one can pay that.[^7]

One modeling study estimated that a gene therapy would only be cost-effective in Uganda at around 700 dollars a dose, against 3.6 million in the United States, a five-thousand-fold gap between where the disease is and where the money is.[^8]

Why? Because for seventy years, sickle cell was not a Western problem, and research follows the money, and the money follows the West.

Now put an LLM on top of that history. The model trains on decades of medical literature that reflects those funding priorities. It knows enormous amounts about conditions that affect wealthy populations and comparatively little about the ones that do not. When it reasons about health, drafts research summaries, or helps a doctor think through a diagnosis, it carries that imbalance forward. Worse, as AI increasingly generates the papers, summaries, and hypotheses that future AI trains on, the neglect compounds. The disease that was ignored by human institutions gets ignored by machine ones, except now the machine sounds like it has considered everything.

Vaccines follow the same logic. Whose diseases get modeled, whose trial populations get studied, whose side effect profiles get documented. The gaps in the data become gaps in the intelligence, and the gaps in the intelligence become gaps in who gets to live well.

The testing is as lopsided as the training.

Here is something most people outside the industry do not know. The frontier AI labs do not test these models alone. They quietly contract specialized firms, staffed with statisticians, red-teamers, and domain experts, to stress-test the models before release. These vendors probe for the failures the labs are worried about: bioweapons instructions, election misinformation, cyberattacks, self-harm content.

That work is real and it matters. But look at what makes the list. The things that get tested are the things high up in the meta-awareness of people in San Francisco, London, and New York. The risks they can imagine. The lawsuits they can foresee. The headlines they fear.

What does not make the list? Whether the model gets Nigerian inheritance law right. Whether its child safety norms make sense in a culture where family structures look different. Whether it understands that a health remedy it confidently recommends is dangerous in a region it barely has data on. Whether it can tell the difference between a government's actual rules on cultural practice and a foreign NGO's summary of them. Nobody red-teams for the history of the Ile-Ife kingdom being flattened into a paragraph of half-truths, because nobody in the room knows enough to notice.

You cannot test for blind spots you do not know you have. And right now, the people designing the tests share the same blind spots as the people who built the models.

And to be clear, LLMs are not the only victims of this, we all have made decisions in life and business due to the “data” available to us, ignoring everything else. But then due to share scale, AI having same data challenge is not something to be swept aside. It’s critical to survival and sovereignty of many.

So what do we do? We document like our survival depends on it. Because it does.

The answer, at its most primitive, is not glamorous. It is documentation.

Websites. More websites. Archives. Government records digitized and published. Court rulings online. Local newspapers preserved. Scientific research from African universities published openly, in local languages and in English. Oral histories recorded and transcribed. Your country's story, your community's story, your family's story, written down and put out there, where the machines that are learning the world can find it.

I feel this personally. I have been building a digital family tree for my great-grandfather, a chief in Ife. Not because it is quaint. Because I understand now that anything undocumented is, to these systems, something that never happened. If it is not in the corpus, it does not exist. And if it does not exist, someone else's version of it will.

Countries need to treat their national corpus the way they treat their airspace. It is sovereign territory, and right now most of the Global South has left it undefended. Every ministry of culture, every national archive, every university should have a mandate and a budget to get their nation's knowledge online, structured, and machine-readable. This is not a heritage project. It is a security project.

Beyond documentation, we need three more things.

We need tools that expose the fables. Independent, public-facing systems that audit what the major models say about specific countries, cultures, health conditions, and histories, and publish the errors loudly. Think of it as a factual accountability index. Not a benchmark run by the labs on themselves, but scrutiny from outside, the way human rights organizations monitor governments.

We need multilateral organizations and non-profits in the arena. UNESCO, the African Union, the WHO, regional bodies, philanthropies. They fund museums and libraries. They should fund corpus building, evaluation of AI systems in low-resource languages, and adversarial testing for the risks that never occur to a lab in California. The labs have shown they will fix what is measured and publicized. So measure it. Publicize it.

And we need governments willing to impose real consequences. Not to ban AI, that ship has sailed and banning it only guarantees you have no seat at the table. But to penalize demonstrable, persistent misrepresentation the way we penalize false advertising or defamation. If a model systematically distorts a nation's history or spreads medical falsehoods that harm its people, there should be a cost. Accountability changed how pharmaceutical companies, banks, and food producers behave. It will change how AI companies behave too. Nothing else will.

The age of reading is ending. That is exactly why this matters.

Here is the hardest truth in this whole essay. Books are not being read anymore. Not the way they were. A generation is growing up asking a chat window instead of opening a text, and no amount of nostalgia will reverse that.

Which means these models are no longer just tools. They are becoming the fact-deciders of civilization. And whoever decides the facts decides the fates. What a child in Lagos believes about her own history. What a doctor trusts about a disease. What an investor assumes about an entire continent. What one people come to believe about another.

We have been here before, when a small group controlled the presses and the archives and wrote everyone else into the margins. It took centuries and immense suffering to claw back from that. This time, the concentration is greater, the reach is instant, and the voice never sleeps.

The models are still young. The corpus is still forming. The window to act is open, but it is closing at the speed of training runs.

So the question in the title is not rhetorical. Will we hold LLMs accountable? Or will we let a handful of machines, trained on a lopsided library, quietly rewrite who we are, and call it intelligence?

The victors used to write history. Now the corpus does. Make sure you are in it.

Footnote

This article was written with the assistance of an AI language model, working from the author’s own outline, argument, and direction, with the factual claims sourced and footnoted above. That fact is itself part of the point. Several phrases in this piece are the kind of confident, cadenced language AI models produce by default - "here is the uncomfortable truth," "the single biggest threat," "this is why this matters" Some were originally the author's (probably influenced by months of reading AI slop and writing with AI), some were the machine's, and after a few edits it becomes genuinely hard to tell which is which. That difficulty is the argument in miniature: when the tool that helps us write also shapes how we write, and increasingly what we believe, the line between our voice and its voice quietly disappears. The author has left these traces visible on purpose.

References

[^1]: During the 1994 genocide, Radio Television Libre des Mille Collines (RTLM) broadcast propaganda and named targets, and is widely documented by researchers and the International Criminal Tribunal for Rwanda as having incited and directed killings.

[^2]: Language-distribution analyses of GPT-3's training corpus put English at roughly 92-93 percent of tokens. See the training-corpus table in "A Survey on Multilingual Large Language Models" (arXiv:2404.00929); a related analysis reports English at about 92.1 percent of ChatGPT's corpus ("Bridging Language Gaps," arXiv:2510.00908). Native English speakers number roughly 373-400 million, about 5 percent of the world's ~8.2 billion people (Ethnologue 2025, via Rosetta Stone and Wikipedia, "English-speaking world"; see also Teflpedia's ~5 percent estimate).

[^3]: English accounts for roughly 49 percent of website content, the largest share of any language, with Spanish and German each around 6 percent (Ethnologue 2025, cited by Rosetta Stone, "How Many People Speak English Globally"). Other measures place English-language web content closer to 58 percent. Either way, its online footprint far exceeds its share of speakers.

[^4]: In the Common Crawl web archive that underpins most LLM training, more than 41 languages each make up less than 0.01 percent of the data, and roughly 100 languages fall below 0.1 percent, with data volume dropping almost exponentially across languages ("UnifiedCrawl," arXiv:2411.14343).

[^5]: Roughly 300,000-400,000 babies are born with sickle cell disease worldwide each year, with more than 75 percent of those births in sub-Saharan Africa; Nigeria accounts for an estimated 150,000 annually (about a third of the global total), compared with roughly 2,000 in the United States. Sources: systematic review in PMC11379310; single-institution study in PMC9120745; "Management of Severe Acute Malnutrition in SCD" (ClinicalTrials.gov NCT03634488); Healio, "Epidemiology and Disease Burden."

[^6]: Linus Pauling, Harvey Itano, and colleagues, "Sickle Cell Anemia, a Molecular Disease," Science 110 (1949): 543-548 - the paper that first identified a specific molecular cause of a human disease, four years before the double-helix structure of DNA was proposed.

[^7]: In December 2023, the U.S. FDA approved two gene therapies for sickle cell disease: Casgevy (Vertex/CRISPR Therapeutics) at a list price of $2.2 million and Lyfgenia (bluebird bio) at $3.1 million per one-time treatment, excluding associated hospital and chemotherapy costs. Sources: CNN, NBC News, BioPharma Dive, Dec. 8-9, 2023.

[^8]: Morgan et al., "The value-based price of transformative gene therapy for sickle cell disease: a modeling analysis," Scientific Reports (2024), DOI 10.1038/s41598-024-53121-0, which modeled country-specific value-based prices ranging from about $3.6 million in the United States to roughly $700 in Uganda.