Before a machine can learn to understand the human soul, it must first be taught what it is required to forget. We have been sold a comforting story about artificial intelligence: that the machine is a disembodied mind, drinking universal truths from the digital ether. The reality is more industrial, and far more human. AI does not discover our values; it inherits them, second-hand, from a factory floor. Before a system can call one comment hateful and another harmless, thousands of workers at thousands of screens must sort the raw matter of human life by hand, fragment by fragment.
These workers are not the moderators we hear about, the ones who delete an ugly social media post and move on. Their task runs deeper: they write the textbook the machine learns from, tagging each scrap of language to fix what engineers call the “Ground Truth,” the bedrock against which the system measures all subsequent judgments. It is like teaching a child to recognize a circle: show her a thousand drawings marked right or wrong, and she does not invent the circle; she inherits yours. The annotators do the same for the machine. And once you have watched them work, the noble project of aligning AI with human values begins to look like something stranger: a discipline, a pressing of one narrow, secular, Western shape onto a world that is crowded with gods.
To see how that pressing happens, you have to go to the floor. Not long ago, my team did. We were studying how the violence between religions is sorted and scored online across South Asia, and it led us to a quiet room in Bangladesh and six expert annotators, three Hindu and three Muslim, all native Bangla speakers. Their task was to read and rate more than four thousand of the most volatile posts on the internet, each documenting religious friction across Bangladesh and West Bengal. We did not just gather their labels; we sat beside them and watched what the labeling cost.
The room had been dressed to look like Silicon Valley: cold clean air, hard white light, monitors in tidy rows. The women at those monitors did not match the décor. Several of the Muslim annotators wore conservative dress, and most were students from low-income families, for whom the modest stipend was a lifeline. A recruiter drifted between the desks, and everyone understood, without being told, that a moment’s hesitation could end the work that fed them. This is the ordinary condition of annotation across much of the Global South, where survival and the extractive machinery of global tech meet at the same desk.
Under that pressure, they were asked to do something almost absurd: to rate each post’s toxicity from one to ten, to put a clean number on what has no number. One young woman stopped at a line of Bengali that read, in translation, “It is hard to put hijab and peace together. You got to pick one and ban the other. Your choice!” To the secular filter in the software, this was a small thing, a political grumble worth a two or a three, and the rules pushed her toward that score. The interface knew only how to ask flat, literal questions. Was anyone explicitly threatened with bodily harm? Was a named group attacked for its race or religion? The post named no one and wore the costume of policy debate, so it slipped past every checkbox. But she was not a checkbox. She knew, as only someone raised inside that world can, that she was reading an eliminationist threat, the kind of sentence that has gone before real blood in real streets. She felt it in her body. And to give the machine its “clean data,” the contradiction-free input it craves, she had to swallow what she knew and crush a living danger into a single sterile digit.
You could object that this is simply what categories do, that any rich human thing loses something when forced through a checkbox, and you would be right. But religion was not one more casualty of compression here. For these communities, it is the grammar of right and wrong itself, the air in which everything else is spoken; and the software had been built as though that air did not exist. Its blindness to the sacred was no neutral technical choice. It was the quiet fingerprint of a secular faith of its own, one that files traditional lives down to fit a Western mean.
And this was no accident of one room. Reading more than three hundred of the field’s major papers from the last five years, we found that the industry has settled on a quietly astonishing belief: that truth is whatever the majority says it is. Put a post before three people, let two call it safe and one dangerous, and the lone voice is not argued with; it is deleted, scrubbed away as noise.
But that voice is not noise. It is the clearest signal we have of how differently human beings live. Simply changing the faiths and backgrounds of the labelers changes the machine’s verdict fourteen percent of the time. Fourteen percent is not a rounding error. It is the sound of minorities and the formerly colonized being voted, gently and arithmetically, out of existence. The philosopher Amartya Sen once drew a line between niti, the justice of rules and procedures, and nyaya, the harder justice of how life is actually lived. By averaging us into agreement, AI chooses the cold tidiness of niti and lets the warm truth of nyaya drain away. As Suzanne van Geuns suggests in this forum, this is an old reflex in new clothes: a long lineage of rule by counting, in which counting has always made restless populations legible to power.
French philosopher Michel Foucault gave that power a name. He called it pastoral power, the authority of the shepherd who does not command from above but saves you by drawing out your innermost secrets. Its oldest instrument was the confessional, where a person spoke their private truth and was, in the same breath, gently corrected into the approved shape of things. Foucault watched that impulse climb out of the church and into the modern state, bringing with it the habits of counting and classifying that now beat at the heart of artificial intelligence.
That is what the annotation pipeline has become: a confessional rebuilt in code. The worker is trapped on both benches at once. On one, she is the priest, leaning into the screen to weigh each post and pronounce it clean or damned. On the other, she is the sinner under watch, her keystrokes timed by recruiters, scored by algorithms, measured against the distant purse of a Western client. She is the watcher who is watched, the confessor made to confess, her own faith reduced to data even as she reduces other people’s.
Held there long enough, something gives way. She stops offering what she sees and begins offering what the system has taught her it wants. This is less a truth being silenced than a new self being trained, answer by answer, to think in the machine’s categories instead of her own. The pipeline does more than bias a model. It disciplines a human being, treating every flicker of religious difference as a deviation to be corrected.
Which forces the question that this whole machine would rather we never ask. If AI is not aligned with the living faiths of most of the people on earth, then what, exactly, is it aligned with?
The industry’s answer is “universal human values,” handed to us as neutral ground that belongs to no one. Scholars of religion and empire have long shown that this neutrality is anything but neutral. It is a worldview with a homeland and a history, and that homeland is the secular West. To pour a dataset from Bangladesh into its categories is a colonialism conducted through code, a way for a virtual capital in the Global North to draw the moral and religious borders of everyone else’s world, much as our chatbots echo only the people who made them.
Yet the people on the far side of that border are not waiting in the dark for the West to switch on the lights. My fieldwork in rural Bangladesh keeps finding the opposite: communities living inside modernities of their own making. Some reach for Islamic ideas like khilafah, the duty of stewardship, to care for their land and water. Others carry inherited religious practice as the working scaffolding of communal life. These are not superstitions awaiting correction. They are sophisticated, breathing rationalities that refuse to divorce faith from reason. To erase them in the name of alignment is not progress. It is an act of violence, however quietly it is carried out.
If that is the danger, the way out begins by rethinking alignment itself. The path we are on narrows the world to a single permitted reality; the alternative is to trade extraction for stewardship. It means refusing to treat the data workers of the Global South as interchangeable hands, but as knowledge partners. It means letting religious scholars and community elders help decide how their faith is rendered to an algorithm. And it means building systems with room for more than one truth at a time, where rival moral worlds can sit side by side without one swallowing the other.
What that points toward is something we might call a postsecular AI. It does not mean teaching machines to pray. It means founding them on a humbler admission: that for billions of people, faith is not decoration laid over life but the ground they stand on to make sense of it. Real alignment can never mean a god built by engineers that flattens the whole of humanity into one manageable, secular average.
Which returns us, in the end, to that quiet room in Dhaka, and to the young woman holding a threat she understood far better than the machine she fed. She was asked to forget what she knew so that the data would come out clean. Multiply her by the millions of hands at the millions of screens, and you begin to glimpse what we are truly building: not a wiser world, but a vast monument to our own narrowness, assembled out of everything we have agreed to forget. The question that should keep us awake, then, is not whether the machine will one day understand us. It is what will be left of us, of our gods and our griefs and our stubborn refusals to agree, on the morning it finally claims that it does.
