Paper short abstract:

Departing from positivist data science in favor of a situated stance, we 1) examine colonial histories of domination and broader political economies, in which AI is shaped. Based on this, we 2) open a spectrum of alternative visions for AI, conceptualizing it as a material-semiotic web of practices.

Paper long abstract:

This paper presents a humanities perspective on the recent developments of generative AI, challenging positivist data science in favor of a situated stance (Haraway 1988).

The first half makes the contingency of power visible, examining the enduring colonial histories of domination shaping current advances in generative AI (Benjamin 2019). And it analyses the broader political economies in which AI is developed by large tech companies (Poell et al. 2019) to better understand, critique, and situate the current instantiations of generative AI—from data generation to model development, infrastructuring to standardization, evaluation to deployment, business models to ethico-political consequences.

The second half focuses on alternative visions. Starting with the possible knowledge cultures that AI may amplify, distribute, and generate, a situated approach to AI responds to varying needs and capabilities (Sen 1985). It incorporates local expertise to develop models that reflect context. Situated AI also depends on training data. Not through extraction or augmentation, but by empowering local actors to curate data on their own accounts, ideally leading to the creation of tailored applications, the multiplication of perspectives on what can be done and made computable and what should not be done.

Situating AI is crucial to expand our perspectives: From AI as a technological destiny, placed upon us by powerful actors, towards AI as a material-semiotic web of practices (Law 2019), thereby engendering a multiplication of imaginations for how to live together as a pluriversal collective, characterized by both: interdependence and mutual responsibility, and respect for profound otherness (Escobar 2019).