Bochum, February 27, 2025

I was initially trained not only in philosophy but also in French and German literary studies. I have thus spent a lot of time over the course of my academic studies learning to read both theory and literature, and to paying close attention to the logical-formal structure of an argument and its rhetoric. I learned how to apply narratological, structural, and linguistic tools to texts of various lengths, or to distinguish between énoncé and énonciation. I learned methods of close reading or explication de texte.

Reading is the foundational skill of the humanities. This skill has undergone tremendous changes since the introduction of digital media and the systematic digitization of texts. Katherine Hayles analyzed this in depth already thirteen years ago, in her book How We Think: Digital Media and Contemporary Technogenesis. Ever a groundbreaking theorist of media, literature and critical posthumanism, Hayles has since continued to explore how humans and machines form what she calls a “cognitive assemblage” that challenges the idea of autonomous, self-possessed subjectivity in processes of meaning-making. I couldn’t be happier to have her give the keynote talk this evening.

Our colleague and director of the KWI, Julika Griem, has equally inspired this workshop with her concept of the “reading scene,” where the practice of reading is explicitly thematized in literary texts and visual media. She proposes that this media reflexivity enables us to analyze the changing forms, valuations, and norms assigned to reading as a cultural practice.

The task at hand as I envisaged it for these two days is to analyze the new reading scenes emerging with large language models (LLMs) and the research practices that surround them. How is reading transformed in terms of modes, methods, but also valuations, when machines read for us, that is, “interpret” content, summarize text, propose essay questions? What does the format of the machine output—the breakdown of content in bullet points, the use of bold typography to attract our attention to “what counts”, but counts following what norms?, the chat with a bot—do to our understanding and valuation of reading? What do we read of a text, a philosophical position, a source in a foreign language, when we read it through machine reading? Is this reading through the machine a reading with the machine? That is, do we form an assemblage of shared cognition and co-constitution of meaning-making when we think and read through an LLM?  

One is reminded of a certain Friedrich Nietzsche, who, in 1882, chiselled away on his Hansen Writing Ball the following words: “Unser Schreibzeug arbeitet mit an unseren Gedanken.” “Our writing tool works with us at shaping our thoughts.” Martin Stingelin has pointed to this passage as one of the foundational “writing scenes” where the writer acknowledges how the machine contributes in a very embodied, material way, to co-shaping what he thinks and what he writes. 

So are we reading with the machine or have we started to read from it, such that we are being trained to align with LLMs valuations and the interpretations it generates? And could we arrive at a point where we consider LLM reading the default mode of what it means to read? Perhaps not for us, but what about for the next generation?

I am convinced that LLMs will have a massive impact on the ability to read and interpret. That is why I am skeptical about the introduction of these models in educational settings at a stage when teenagers are still learning how to read—grappling with how to parse information, how to understand and interpret metaphors, and how to develop an awareness for the rhetorical dimension of any text. The rationale behind this introduction, apart from clear economic interests, is that LLMs constitute a chance as long as students learn how to—I quote—“use them critically.” 

Critical thinking is the operative, yet under-defined concept that does a lot of work in the institutional documentation around LLMs for schools and universities. Critical thinking functions as the safe-gard of students' autonomy, a value still considered essential in pedagogical settings. At the same time, it is expected that critical thinking toward AI doesn’t hinder the readiness to use the tools so as to optimally prepare students for the job market.

I did some research on the proposed activities supposed to teach students “critical thinking” towards LLMs outputs. What I found so far is that these activities often consist in discovering the inaccuracies and errors that the model generated in response to a prompt. So high-school students may be asked to compare AI output with information existing elsewhere on the web or in books. Such activities train students to become not the active producer of meaning, of analysis, of arguments, but the evaluators, modulators, ameliorators of the model’s output. These evaluations, modulations, and improvements expected from the students are made on the basis of and thus determined by the kind of normative rationality that characterizes current AI.

Two precisions before I leave the stage to our first guest; the first regards the question of expertise, the second the idea of LLMs as normativizing machines:

1. Expertise

Serious output evaluation presupposes a degree of expertise that must be at least equal or superior to the “expertise” of the machine. Without expertise, one has to either fact-check the entirety of an output or take the model’s word for it. The acquisition of the necessary expertise rests on the existence of an exteriority to the model. However, this exteriority is currently being itself populated by generated content. 

Speaking of my own experience, I have yet to get an output that does not contain hallucinations, errors, or inaccuracies. To mention but one example, I asked GPT to explain inflation from a Marxist perspective. The model first generated somewhat unsurprisingly a supply-and-demand-based explanation. But even after the third prompting/correction, the model still wasn’t able to shift its attention head to the location in its vector space containing what is apparently a very minoritized position: the marxist theory of economy. At the same time, LLMs are fine-tuned through reinforcement learning to repeat again and again in an apologetic tone that they always aim to generate content that is “neutral and balanced.” 

2. LLMs as Normativizing Machines

This claim to neutrality and balance must be challenged for several reasons. The first challenge is not, as one could expect, to expose the existence of biases. Indeed, their existence is gladly admitted by companies, which frame biases as a temporary problem that is the occasion to perform their idea of liberalism: where every subject deserves equal representation in the eyes of the model; a good occasion to gather even more data.

Instead, the challenge to the “neutral and balanced” claim would be to show how this claim invisibilizes the fact that LLMs and other generative AIs are normativizing machines. My hypothesis would be that this new algorithmic rationality is a normative rationality. That artificial intelligence is an inherently normative endeavor marks a departure from earlier forms of rule-based artificial intelligence that pertained to what historian Erickson and colleagues have called in their book How Reason Almost Lost Its Mind “Cold War Rationality,” a rationality devoid of human judgement and interpretation and thus considered particularly apt to minimize uncertainty. 

In contrast, current machine-learning-based AI explicitly relies on the integration and automation of moral judgment, valuation, and interpretation. So when a LLM asserts that it aims to be neutral and balanced, it forecloses not only the normative work going into the systematic production of this claim, that is reinforcement learning often based on human evaluations; but it also works at neutralizing the question of positionality and partial, situated knowledge, which has been a major epistemological contribution of feminist and postcolonial scholars. 

In the meantime, these models are fine-tuned to fulfill evaluative tasks such as the automated evaluation of resumes, job applications (Gan, Zhang, and Mori 2024) and more recently college applications in the U.S., as well as the automated grading of students' homework in our very own German schools. Rainer Mühlhoff will present his research on one of these tools. As a heads-up, if you want to pass all these standardized tests, you better make sure that your application materials reflect the statistical norm and ethico-technical valuations that constitute a LLM.

To Conclude on A Methodological Note

I am convinced that close reading is not over: but a close reading that attends to machine output, to the research output of the machine learning community and to the technology itself. To this, I would add genealogies, that is, histories that insist on the contingency of the trajectory of technology while accounting for the specific valuations–in every sense of the word–that have contributed to shaping its course.