Atlas of AI

An AI application seems almost weightless on screen: enter a question, receive an answer. Hidden from view are data centres, energy and water consumption, mineral resources, globally distributed labour, data collections and institutions that decide on development and deployment.

In Atlas of AI Kate Crawford shifts the focus from the individual algorithm to the entire system. She describes AI as material and political infrastructure, not autonomous, neutral intelligence. Anyone assessing AI responsibly must therefore look beyond model quality: what resources, labour conditions, classifications and power relations make the system possible?

1. Make the material infrastructure visible

Digital systems need physical foundations: mining for hardware, factories and transport, data centres, electricity and cooling. Even a seemingly small query is part of this infrastructure. Individual resource-use figures depend heavily on the model, location and measurement method – the fundamental material footprint remains.

A responsible perspective therefore asks about life cycles, not just convenience: what hardware and energy are needed? How often is a model trained or queried? Which environmental impacts are measured, published or outsourced?

2. Recognise data and categories as decisions

Training data do not simply represent the world neutrally. Someone decides what is collected, excluded, labelled and used as a target variable. Categories can contain historical prejudices, institutional interests or unsuitable simplifications.

This becomes particularly consequential when systems assess or classify people. High technical accuracy does not yet answer whether the task is legitimate, what errors arise for whom, or whether those affected can understand or contest a decision. The question of fairness therefore begins before the model.

3. Include labour and power in the assessment

AI relies on human labour: data are collected and annotated, problematic content reviewed, systems maintained and results implemented in organisations. Some of this work remains poorly visible or takes place under unequal conditions.

At the same time, large systems concentrate data, computing power and decision-making authority. Responsibility must therefore not be delegated solely to individual users. Procurement, corporate governance, regulation and public oversight help determine which systems are built and which alternatives remain possible.

Your exercise: map an AI application

Choose an AI system you use or are considering. Create an initial map in three minutes:

1. Minute 1 – resources and labour. What data, hardware, energy and human activities might stand behind the system? Mark what is unknown.
2. Minute 2 – decisions and affected people. What does the system influence, who benefits and who bears possible errors or costs?
3. Minute 3 – responsibility. Formulate one question the provider or organisation would need to answer and one limit for your own use.

The map need not be complete. Its value lies in translating the invisible into concrete questions for assessment and design.

What we take from this

Atlas of AI adds systems thinking and responsibility to digital competence. Within WEITER7 the map can form part of the reality check: consider not just immediate benefits, but include affected people, side effects and scope for design in the next experiment.

Which prerequisite or affected group remains invisible when I assess an AI application only by speed and output quality?

Context: This insight distils selected arguments from Crawford's critical analysis in our own words. The exercise and WEITER7 allocation are our own applications. The book offers a societal and political perspective, not a comprehensive technical evaluation of every AI system. Specific environmental and social impacts differ by technology, provider, location and use, and should be assessed with current, verifiable data.

Sources and further reading: Kate Crawford's official book page ↗ · Yale University Press ↗