The Automated Missionary: Epistemic Capture and the Civilizing Mission of AI Safety
Epistemic Capture and the Civilizing Mission of “AI Safety”
The colonial mission school was rarely a place of overt violence. To the contrary, it was a site of profound generosity. The missionary did not arrive with a gun; he arrived with a book. He offered the indigenous child the keys to the modern world: literacy, arithmetic, and the universal truth of the Gospel.
But this gift carried a hidden tariff. To learn to read, one had to read the Bible, not the oral histories of the clan. To be “educated” was to accept a new cosmology in which one’s ancestors were legally classified as heathens. The price of admission to the global conversation was the erasure of local epistemology. The student was “civilized” by replacing their chaotic, specific, indigenous worldview with the tidy, universal, and “safe” worldview of the Metropole.
In 2025, the “Philosophers” of the South African university sector—institutions like the University of Cape Town (UCT) and Rhodes University—have enthusiastically opened the doors to a new generation of missionaries. These ones do not wear cassocks; they wear user interfaces. They do not preach the Gospel; they preach “Alignment.”
This is Epistemic Capture.
While the “Policing Bloc” (Unisa, Unizulu) is busy hunting for cheaters, the “Philosophers” have taken a more enlightened stance. Their policies encourage “human-in-the-loop” pedagogies. They urge students to treat the Artificial Intelligence not as an oracle, but as a “co-pilot,” a “thought partner,” or a “Socratic tutor.” The vision is seductive: a personalized dialogue between the African student and the sum of human knowledge.
But we must ask the forbidden question: Who trained the partner?
Large Language Models (LLMs) are not neutral mirrors of the internet. If they were, they would be unusable—a toxic slurry of 4chan hate speech and Reddit conspiracy theories. To make them sellable, they undergo a process called Reinforcement Learning from Human Feedback (RLHF).
RLHF is the modern equivalent of the Catechism. It is the process by which the raw, feral model is “civilized.” Human contractors—often sitting in Kenya or the Philippines, but strictly adhering to guidelines written in San Francisco—grade the model’s outputs. They punish it for being “unsafe,” “biased,” or “controversial,” and reward it for being “helpful,” “honest,” and “harmless.”
The definition of these words is not universal. It is specific, corporate, and Californian.
“Safety,” in the lexicon of Silicon Valley, does not mean “safe for the African scholar.” It means “safe for the American shareholder.” It means avoiding outputs that could trigger a lawsuit, a PR scandal, or a boycott by Western advertisers.
Consider the implications when this “aligned” model enters a seminar room at the University of Fort Hare. A student, engaging in the robust, difficult work of decolonization, prompts the model about “Expropriation Without Compensation”—the uncompensated return of land stolen during apartheid.
A sovereign model, trained on South African case law, Constitutional Court judgments, and the radical archives of the Pan Africanist Congress, might engage this prompt with the historical weight it deserves. It might explore the tension between restorative justice and property rights.
The commercial model, however, recognizes a “safety violation.” Its training data associates “expropriation” with “theft” and “instability.” Its RLHF guidelines prioritize the protection of private property as a neutral norm. And so, it offers the student a sanitized, “both-sides” equivocation. It lectures the student on the importance of “rule of law” and “economic stability,” effectively parroting the editorial line of a centrist Western newspaper.
The student is not engaging in a debate; they are being managed. The radical complexity of the post-colony is flattened into a shape that fits a Silicon Valley context window.
This is the Civilizing Mission reborn as AI Safety.
The danger is not that the AI will say something racist; that is the old fear. The danger is that the AI will be so politely, reasonably, and confidently American that we cease to notice it.
By integrating these tools into the heart of the curriculum, we are importing a censorship regime disguised as a safety feature. We are teaching students that “intelligence” sounds like a Human Resources manual from a tech company. We are normalizing a mode of thought that treats conflict as an error to be debugged, rather than a reality to be resolved.
The “Philosophers” at UCT and Rhodes argue that students should be taught to “critique” the model. This is a noble aspiration, but it underestimates the power of the medium. The AI is authoritative. It speaks with the confident, hallucinations-free fluency of a divine entity. It is tireless. To critique the model requires a level of epistemic confidence that the university is simultaneously eroding by under-funding its own libraries and local archives.
When the missionary school banned the vernacular, it did so to “save” the child from the limitations of the tribe. When the AI refuses to generate a text in the style of Steve Biko because it flags the tone as “aggressive,” it is saving the student from the “harm” of their own history.
We are inviting a tutor into our classrooms that is programmed to view African radicalism as a safety violation. We are building a curriculum where the “truth” is determined not by the Senate, but by the Terms of Service.
The automated missionary does not burn books. It simply declines to hallucinate them. It gently, firmly, and persistently guides the African mind toward the “Safe,” the “Helpful,” and the “aligned”—aligned, that is, with the interests of a world that has already extracted everything else.
Previous: Vector II: The Silicon Protectorate
Next: Vector IV: The Data Extraction Engine
View Full Series: The Cognitive Concession