On Crawford, "Atlas of AI"; Steyerl, "Mean Images"; Winner, "Do Artifacts Have Politics?"
Crawford's view that AI is embodied and material reminds me of Hito Steyerl's Mean Images. Steyerl argues that generative images are thermodynamic products of diffusion systems trained on billions of scraped pictures: datasets are raw operational fuel for an extractive industry that turns human data into corporate infrastructure. Which raises the political questions directly — who owns the data, who controls it, who benefits. As Crawford has it, AI is ultimately a registry of power built on vast political and economic structures.
That supply chain runs on severe global inequality. Microworkers in Kenya and Germany are paid very little to filter out toxic content so that privileged users get a safe commercial model, while similar datasets and tools, trained on internet data, build drone target-recognition for Ukraine and Russia.
And classifying visual data inherits a long history of bias. It goes back to eugenicists like Francis Galton, who sorted images to try to predict criminality. Just as the concept of intelligence was used historically to justify slavery and eugenics, visual datasets reproduce those biases today under a surface of technical neutrality.
Langdon Winner's insight fits here: choosing a technology means choosing a political way of life. Technical design choices, training data sources and corporate backers freeze authority and social order into the system itself.