What Are the Data Sources for Generative AI Training According to Cartography of Generative AI?

Updated

According to Cartography of Generative AI, the training data behind generative AI comes mainly from one place: the internet, harvested through automated scraping of content published and shared by millions of users. The site frames this as the industrial compilation of training datasets drawn from "the largest digital archive in existence." That answer applies to the general mechanism the project describes — it is a critical mapping of how these systems are built, not a technical specification of any single model.

How the data is collected

The project describes a straightforward extraction pipeline:

  • Source: online content published and shared by internet users.
  • Method: automated "scraping" of that content.
  • Scale: large and heterogeneous datasets, assembled industrially rather than case by case.

This scale matters because it is what allows generative tools to generalize. As the site explains, earlier systems were trained to produce specific outputs (a face, text in a certain style). Current tools go beyond that to produce many types and styles of content, and that ability "is based on the processing of much larger and more heterogeneous datasets in order to respond to all kinds of prompts."

Why the original collection happened

A detail worth keeping: the extraction was not initially aimed at commercial use. According to the site, the original motivation "did not foresee today's commercial exploitation by start-ups and platforms, but was driven by the desire for scholarly and non-commercial research." The commercial layer was added later, once these archives proved useful for generating text and images on demand.

The paradox the site identifies

Once huge digital archives became the raw material for on-demand text and image generation, cultural industries were left with a set of contradictions. The project points to a specific ideological framing: the "big data" view treats internet content as a vast repository that can be extracted and processed. That framing sits uneasily alongside the fact that the content was created by millions of people who did not publish it for that purpose.

What this means if you are trying to understand a generative tool

If you want to reason about where a model's outputs come from, the site's account suggests three questions:

  1. What was scraped? Publicly shared online content, at a scale large enough to cover many styles and types.
  2. Who intended it to be used this way? Initially, non-commercial researchers — not the platforms now exploiting it.
  3. What assumption makes it seem acceptable? The idea that the internet is a repository available for extraction and processing.

The site also notes that generative AI tools "disassemble language (visual, textual) to reassemble it based on the calculation of probabilities." So the data source is not a set of curated examples chosen for quality — it is a statistical sample of what was already online.

Where to read the full argument

Cartography of Generative AI is published as a PDF and is available in multiple languages, including English, Catalan, Spanish, Italian, German, French, Basque, Swedish, Portuguese, and Chinese. The excerpt above covers sections 01 and 02; the full document continues into the controversies the project sets out to map.