How Does Generative AI Work According to Cartography of Generative AI?

According to Cartography of Generative AI, generative AI works by learning from examples rather than being programmed with explicit steps. Instead of a developer writing out the concrete instructions for producing a piece of text or an image, the system is fed many examples of a case, and statistical networks configure themselves by analysing the recurring patterns in that data. The tool then "disassembles" language — visual or textual — and reassembles it based on the calculation of probabilities.

Learning from examples, not instructions

The site's first point is that automation in generative AI is achieved "not by programming the concrete steps to be taken, but by using examples." This is the core mechanism:

• If you have many examples of a case, you can process them using statistical networks.
• Those networks configure themselves by analysing recurring patterns in the examples.
• Whether the input is words, pixels or sound frequencies, the result is a statistical model derived from a training data set.

So the model is not following a hand-written recipe. It is generalising from a body of prior material.

Disassembling and reassembling language

The site describes the output side of the process in a memorable way: generative AI tools "disassemble language (visual, textual) to reassemble it based on the calculation of probabilities."

In practice, that means the tool breaks down the patterns it has learned and then reconstructs an output by estimating what is most probable given the prompt. The result is not retrieved from a fixed catalogue of specific expressions — it is generated.

Why generalisation changed the scale

A few years ago, these tools were trained to produce concrete expressions — for example, the image of a face, or text in a certain style. Today they go beyond that specific concreteness to produce many types and styles of content.

This ability to generalise rests on processing much larger and more heterogeneous datasets, so the system can respond to all kinds of prompts. The site notes that the scale of this change has been large enough to require boosting new economies and accelerating reliance on ecosystems.

Where the training data comes from

The site's second point covers datasets. The industrial compilation of training datasets is achieved by extracting content from the largest digital archive in existence: the Internet. Data is obtained through automated "scraping" of online content published and shared by millions of internet users.

Two things are worth noting from the source:

• The original motivation for this extraction did not foresee today's commercial exploitation by start-ups and platforms; it was driven by the desire for scholarly and non-commercial research.
• Now that these huge digital archives are used to generate texts and images on demand, a series of paradoxes and controversies has arisen within the cultural industries.

The site also points to the ideology of big data, which understands Internet content as a vast repository that can be extracted and processed.

The takeaway

If you want a one-line answer: generative AI, as mapped by this site, is a statistical pattern-learning process. It learns from examples, builds a model of recurring patterns, and generates output by calculating probabilities — and its modern generality depends on very large, very heterogeneous datasets scraped from the web.