ABSTRACT
We arrive at the current state of software and the social world. The critical stance pursued through all the eras so far remains just as important in what has now become an era of artificial intelligence. The collector economy, described in the previous chapter, has now found a new target of interest and a way to present the results of this collection to enthusiastically engage new audiences. What is now being collected is the content that has been created through network-based software. The contents of Wikipedia and Reddit have, among with many other sources, become the input that then contributes to the training of large language models (LLMs). At this point, the spectacle is no longer simply about the circulation of images and messages but is actively folding these “things” back into the tools that can ceaselessly generate new and very plausible fragments of text and image, automating the production of what Debord would describe as pseudo-events and pseudo-history (Debord 1970). However, we cannot simply dismiss these outcomes as inauthentic. If, as Baudrillard argues, the distinction between reality and simulation has looped back in on itself and models now generate a hyperreal that precedes what we take to be “the real”, then we need to unpick a more critical pathway through these tools and their outputs (Baudrillard 2021). Viewed through the lens of Lyotard’s analysis, LLMs appear as exemplary postmodern knowledge machines. These machines do not claim to ground their response in a shared metanarrative but maximise their fluency and usefulness with particular language games, optimising for a form of performative success in search, customer service or creative assistance rather than specifically presenting any stable correspondence to an external reality (Lyotard 1984) – the relationship is a fortunate combination of words drawn from training rather than representing a conscious purpose. If it helps, and put more simply, we can think of generative AI as a very large fruitcake. Each post, message or entry in the training data becomes an individual raisin or pecan. The baking process, that is the training, mixes and transforms all of these ingredients, with occasional “secret” additions by the chef. Guardrails and interface design are the icing that make the cake presentable. The final cake is the LLM. The result is a single, apparently coherent artefact that hides the unevenness of its ingredients and the transformations that they have undergone. Individuals cannot slice the cake directly but instead, they must describe the slice that they want. Clear instructions will yield an acceptable piece, while poor descriptions produce strange, sometimes unsettling results.
