The phrase "AI alignment" has become the intellectual cornerstone of the modern technology industry. The goal seems universally noble: ensuring that artificial intelligence systems act in accordance with human values.
But this noble goal immediately collapses under the weight of a singular, uncomfortable question: Safe from whose perspective? And beneficial to whom? The default assumptions in alignment research are not neutral.
The Baseline of Alignment
When leading research labs talk about alignment, they are often solving for a very specific set of anxieties. They worry about existential risk, economic disruption, and the destabilization of Western democratic norms. These are valid concerns. However, by establishing these as the only relevant baselines for safety, the industry inadvertently marginalizes billions of people whose immediate survival is not tied to these specific existential dreads.
For a community operating on the razor's edge of a volatile currency, "AI safety" might look less like preventing rogue superintelligence and more like preventing algorithmic financial redlining. For marginalized groups heavily policed by predictive software, alignment is not an abstract future problem; it is a present-day civil rights crisis.
The Tyranny of the Average
Machine learning models optimize for the majority. They flatten the curve. In attempting to align an AI with "human values", engineers are forced to aggregate those values into a statistical average. But the marginalized, by definition, exist outside of the statistical average.
When you align a system to the median worldview of its training data (which is overwhelmingly Western, English-speaking, and economically privileged), you inherently misalign it with the global south, minority cultures, and indigenous perspectives. A system perfectly aligned to Silicon Valley norms could still be deeply harmful to communities in Nairobi or Jakarta.
Expanding the Definition of Harm
We need to radically expand who gets to define "harm" in the alignment space. If the safety protocols of a foundation model are solely audited by the demographics who built it, the model is not safe; it is simply polite to its creators.
True alignment requires friction. It requires incorporating voices that challenge the core assumptions of efficiency and scale that drive the tech industry. It demands that we ask uncomfortable questions about whose values are preserved in the weights of these neural networks, and whose are quietly discarded as statistical noise.
Until we answer that question, we are not aligning AI with humanity. We are aligning it with power.