The Dataset Is a Colonial Map

A developer sits in a dark room, the monitor’s glow illuminating their face. With a few keystrokes, a command is sent: load_dataset('c4'). Terabytes of text, scraped from every corner of the web, begin to flow. It feels like magic, like accessing a public library the size of a planet. They see themselves as a pioneer on a new frontier. But the map they are using to navigate it was drawn by pirates.

The massive datasets that power foundation models are not neutral collections of human knowledge. They are artifacts of appropriation. They are colonial maps, created by claiming a territory that did not belong to you, drawing borders that serve your own interests, and then presenting the result as an objective view of the world. The entire AI industry is built on this single, convenient fiction: that the internet is terra nullius, empty land, there for the taking.

For decades, we built a digital commons. People shared their thoughts on personal blogs, posted their art to online galleries, debated endlessly on forums, and uploaded their code to public repositories. They did this under a loose social contract of human-scale interaction. No one who posted a family photo on Flickr in 2007 imagined it would be used to teach a facial recognition system. No author who put their life’s work online thought it would become undigested grist for a machine designed to mimic them.

Yet that is precisely what happened. Projects like Common Crawl, LAION, and The Pile hoovered up this world of human expression, treating consent as an externality. This wasn't curation; it was strip-mining. Just as colonial powers extracted resources from new continents without regard for the inhabitants, AI companies have extracted our collective digital lives and called it a training set.

Like any colonial map, the resulting geography is deeply distorted. It wildly over-represents the English-speaking world and the cultures that dominate the web. It enshrines the biases, bigotries, and bad ideas present in its source material not as bugs, but as features of the terrain. When a model generates stereotyped images or confident falsehoods, it is not malfunctioning. It is faithfully reporting the distorted topography of the map it was given. Attempts to "fix" this with alignment and filtering are like building a few sanitized tourist resorts in a conquered territory. The underlying reality of the appropriation remains.

The lawsuits from artists, authors, and newspapers are not just commercial disputes. They are the first resistance from the native population. They are a belated attempt to declare sovereignty, to say: this land is not empty. We live here. Our work has value beyond its utility as raw material for your models.

We are rushing to integrate this technology into every facet of our lives, from search engines to legal research to medical diagnostics. Every time we ask an AI a question, we are asking it to consult this plundered map. The danger is that we will eventually forget it is a map at all, and begin to mistake its convenient, distorted, and stolen representation for the territory itself. The lines being drawn by scraping bots today are setting the borders for the reality of tomorrow. The real fight is not about who can build the most powerful model, but about who gets to draw the map.