Revisiting Joseph Weizenbaum’s Performance and Theory Modes in AI

The third paper in this AoIR 2023 session is Matthew Salzano, presenting a paper co-authored with the late Misti Yang. Their work focusses on Joseph Weizenbaum’s critique of AI, the creator of the Eliza chatbot and prominent AI theorist whose work offers a valuable vocabulary for the current AI discourse.

Weizenbaum theorised a performance mode of AI, where output is valued over input or means; this risks overvaluing engagement with undertheorised models rather than understanding how these systems work. AI evangelists exemplify this performance mode; their evangelism evades questions about the ethics or implications of such systems. An alternative to this is the theory mode, which questions AI products against an understanding of how such systems should work. This is exemplified by current calls for a participatory turn in AI design which goes beyond mere stakeholder or end user consultations.

Theory mode might be epistemological and based on deep expertise in specific fields of knowledge; or it might be axiological, and focus on how AI outputs represent and reveal cultural biases and prejudices. This also re-emphasises the responsibility of AI designers for the outputs produced by their products – a responsibility that has often been eschewed. This can create further pressure in favour of ethical AI development, too. We should work together to denaturalise performance mode and encourage a stronger embrace of theory mode.