cartography-of-generative-ai.net

What is Cartography of Generative AI?

Cartography of Generative AI is a visual and textual explainer that maps how generative AI works and what it depends on. Rather than presenting AI as a self-contained technical marvel, it traces the chain of data, labour, resources and infrastructure behind tools that write text or generate images. It is published as a freely accessible PDF in many languages, including English, Catalan, Spanish, Italian, German, French, Basque, Swedish, Portuguese and Chinese.

The project's core argument, as described on the page, is that popular AI imaginaries invite "alienation and mystification." Its response is to ask concrete questions: what set of extractions, agencies and resources allow us to converse with a text-generating tool or obtain images in seconds? The available excerpt outlines two sections:

• 01 — how generative AI works: these tools automate tasks not by programming explicit steps but by learning from examples. Statistical networks analyse recurring patterns in large training datasets of words, pixels or sound frequencies. The result is a model that disassembles language and reassembles it through probability calculation. The page notes that generalisation across many content types and styles relies on much larger, more heterogeneous datasets.
• 02 — datasets: industrial training datasets are built by scraping online content published and shared by millions of internet users. The original motivation for this extraction was scholarly and non-commercial research, not today's commercial exploitation by start-ups and platforms. The page frames the resulting conflicts within cultural industries as paradoxes and controversies.

Who it is for. Readers who want a critical, structural overview of generative AI rather than a product guide or technical tutorial: students in media, design or STS; journalists and cultural workers; artists and writers concerned with how their work enters training data; and anyone trying to explain AI's material dependencies to a non-specialist audience. The multilingual PDF format makes it usable in classrooms, reading groups and workshops across several language communities.

How to use it. Start with section 01 if you need a plain-language account of how these systems learn from examples, then section 02 for the data-extraction story. The mapping approach is most useful as a discussion prompt: pick one tool you use regularly and trace the categories the project raises — training data sources, extraction, labour and infrastructure — to see where your own knowledge is thin.

Trade-offs to expect. It is an argument and a map, not a neutral reference: it foregrounds extraction, power and mystification, so readers looking for a balanced survey of benefits and risks may find the framing one-sided. As a PDF, it is suited to linear reading or printing rather than quick lookup, and it explains concepts rather than evaluating specific products. The excerpt does not state a publication date or licence, so check the file itself before redistributing it.

Next step. Download the PDF in the language you will actually read and share, and pair it with a current, product-focused account of the same tools so the structural critique meets up-to-date specifics. For a complementary critical perspective on AI and its infrastructures, AI Now Institute publishes related research.

What are the main environmental and social impacts of generative AI?

The site frames generative AI as an extraction chain rather than a purely digital product: models are trained by scraping content published by millions of people, then reassembled through statistical patterns. That framing points to two linked impact areas — environmental and social — that often get discussed separately.

Environmental impacts

• Energy and compute: Training and running large models requires significant computation, and the site stresses that scale has grown as models become more general-purpose.
• Material extraction: The site asks what "extractions, agencies and resources" make a text or image tool possible, implying hardware, minerals and infrastructure beyond the screen.
• Water and land: Data centres and chip fabrication consume water and land, though the page evidence does not quantify this.
• E-waste: Rapid hardware turnover is a downstream consequence the site's "more than human territories" framing invites you to consider.

Social impacts

• Data provenance: Scraped content comes from internet users who did not necessarily consent to commercial training use.
• Cultural-industry controversy: The site explicitly notes "paradoxes and controversies within the cultural industries" once archives built for research become commercial inputs.
• Labour: Automation of writing and image-making shifts work rather than simply removing it.
• Concentration of power: The reliance on "new economies and accelerated reliance on ecosystems" suggests dependence on a few large platforms.

Where the site is strongest — and its limits

Cartography of generative AI is a mapping project available as PDFs in multiple languages (English, Catalan, Spanish, Italian, German, French, Basque, Swedish, Portuguese and Chinese). It is useful for conceptual orientation, not for statistics. If you need figures on energy per query or water per training run, you will need to look elsewhere.

A practical next step

Pick one tool you use and trace three inputs: the data it was trained on, the hardware it runs on, and the labour it displaces or creates. The site's own question — what extractions allow us to converse with a text generator or get images in seconds — is a good checklist for that exercise.

How does the extraction of training data for generative AI affect data privacy and copyright?

The site frames training-data extraction as a two-sided problem: it depends on scraping the open internet, which creates both privacy exposure and copyright conflict. As Cartography of generative AI puts it, datasets are built by extracting content from the largest digital archive in existence, originally for scholarly and non-commercial research rather than today's commercial exploitation.

Privacy

• Scraped material can include personal writing, photos, conversations and metadata that people never intended as training input.
• Because the extraction is automated and large-scale, individual awareness and consent are largely absent.
• The harm is not only inclusion but inference: models can reproduce or leak patterns tied to identifiable people or groups.

Copyright

• Text, images and audio are copied into datasets and used to train systems that may later compete with the original creators.
• The site notes that this has produced paradoxes and controversies within cultural industries, especially where the original motivation was research rather than commercial use.
• Questions of fair use, licensing and remuneration remain contested and vary by jurisdiction.

A concrete scenario: a photographer finds her portfolio images in a model that generates similar stock imagery. She faces a privacy-adjacent issue (her name and style linked to outputs) and a copyright issue (reproduction and derivative use). Her practical options are to document the outputs, check the model's data statement, and seek advice on licensing or takedown routes in her jurisdiction.

Next step: if you are a creator or a small team, keep a dated record of your published work and the model outputs that resemble it, then check whether the provider offers an opt-out or data-removal process before deciding on legal or licensing action.

What are the working conditions and labor practices behind generative AI development?

The site maps generative AI to its material and human inputs, but it focuses on data extraction, not on a detailed account of developer pay or workplace conditions. From its page evidence, the clearest labor point is that training datasets are built by automated scraping of online content published by millions of users—work that is uncompensated and largely invisible. The site also notes that the original motivation for this extraction was scholarly, non-commercial research, not today's commercial exploitation by start-ups and platforms.

That distinction matters for anyone trying to understand labor practices: it separates the visible "AI developer" role from the hidden supply chain. If you are researching this topic, treat the site as a conceptual map of extraction and dependency, then look for dedicated labor reporting to fill in wages, contracts and worker testimony.

What the site supports

• Generative AI automates writing and image production by learning statistical patterns from examples, not by following explicitly programmed steps.
• The scale of current systems depends on much larger, more heterogeneous datasets and on "new economies and accelerated reliance on ecosystems."
• Datasets come from scraping the internet—content created by millions of people without their direct negotiation.

A practical reader scenario

Suppose you are a designer or writer deciding whether to use a generative tool at work. The site's framing suggests asking not only "does this tool save time?" but "whose past output made this possible, and under what terms?" That question is the bridge between the site's cartography and the labor conditions it does not detail.

For further context, compare how different organizations describe AI labor and data provenance, for example Partnership on AI or OECD.AI. For the site itself, start with its downloadable PDFs in your preferred language to follow the extraction argument in full.

How can artists and cultural workers respond to the challenges posed by generative AI?

Artists and cultural workers can respond by treating generative AI as something to map and interrogate rather than simply adopt or reject. The Cartography of generative AI frames the technology as a set of "extractions, agencies and resources" — training data scraped from the internet, statistical models that reassemble language and images, and the new economies and ecosystems built around them. That framing suggests the response is less about mastering a tool than about making its hidden supply chains visible.

Practical directions

• Make the pipeline the subject. Work that reveals where training data comes from, who produced it, and what labour and resources are involved turns an opaque process into something an audience can examine.
• Organise collectively. Individual refusal is weak against platform-scale extraction; shared licensing standards, union positions, and mutual aid among cultural workers carry more weight.
• Separate uses. Using generative tools for ideation, accessibility, or administrative drudgery raises different questions than using them to replace commissioned illustration, translation, or voice work.
• Interrogate the mystification. The project notes that AI's popularisation produces "imaginaries that invite alienation and mystification." Writing, teaching, and curating that counter those imaginaries is itself cultural work.
• Claim provenance and consent as cultural demands. If datasets were compiled by scraping content published by millions of users, then attribution, consent, and compensation are legitimate negotiating positions, not nostalgia.

A concrete scenario

A small gallery plans an exhibition on machine-generated imagery. Rather than showing only polished outputs, the curators could commission a companion piece tracing one image's likely training sources, invite a labour lawyer and a data researcher to speak, and publish a plain-language note on what the gallery will and won't use generative tools for in its own design work. That turns the exhibition from a showcase into a map.

How to choose your approach

For context on how other cultural institutions have approached these questions, see Creative Commons on licensing and Artists Rights Society on artists' rights. As a next step, pick one concrete decision in your own practice — a contract clause, a teaching module, a commissioning policy — and make the extraction behind it explicit to the people it affects.

What are the key differences between generative AI and other forms of artificial intelligence?

Generative AI is best understood as a subset of AI, not a synonym for it. The key difference is what the system produces and how it learns to produce it: generative models create new content (text, images, audio) by learning statistical patterns from examples, while most other AI systems classify, predict, rank, or decide.

How generative AI differs from other AI

The Cartography of generative AI describes the generative approach clearly: automation is achieved "not by programming the concrete steps to be taken, but by using examples," and the tools "disassemble language (visual, textual) to reassemble it based on the calculation of probabilities." It also notes that earlier systems were trained for concrete outputs (a face, text in a certain style), whereas current tools generalise across many types and styles because they process much larger, more heterogeneous datasets.

What follows from that difference

• Open-ended vs. bounded tasks. A classifier answers a fixed question; a generative model responds to almost any prompt, which makes evaluation harder and outputs less predictable.
• Data dependency. The Cartography notes that training datasets are compiled by scraping online content, originally for scholarly and non-commercial research, and that this extraction now fuels commercial exploitation — the source of current controversies in the cultural industries. Classification models also need data, but the scale and the cultural stakes differ.
• Human role. Generative tools shift work toward prompting, reviewing and editing; predictive systems usually feed a decision that a person or downstream system acts on.

A practical way to choose

Ask what you need: a decision (is this transaction fraudulent?) points to a classifier or predictor; a draft or artefact (a first-pass article, a concept image) points to a generative tool — with human review built in.

For a deeper, visual explanation of these distinctions and the resource flows behind them, the project offers downloadable PDFs in multiple languages, including English, Català, Español, Italiano, Deutsche, Française, Euskara, Svenska, Português and 中文, at Cartography of generative AI.