Epistemic Agency and the Substrate Problem: Realization of Reading Comprehension with LLMs
Third in a series on Floridi's Levels of Abstraction and literacy pedagogy
Luciano Floridi's recent theorizing about LLMs and my speculative application of his frameworks to literacy pedagogy might evoke moderately friendly professional development discussions among teachers. With deft philosophical moves, Floridi soft-pedals the roar from the crowds who fear the takeover of life as we know it by AI, ranging from the renowned theoretical physicist Stephen Hawkings, who issued stark warnings about artificial intelligence for years before his death in 2018, to Geoffrey Hinton, known as the "godfather of artificial intelligence" and a pioneer of deep learning, who warned the world about AI risks in 2023.
In this essay, the third in my series devoted to exploring potentially useful links between Floridi's Levels of Abstraction (LoA) and literacy teaching and learning, I want to focus on a crucial pedagogical problem: students are encountering two fundamentally different types of text that require entirely different comprehension strategies, yet we're teaching them as if all texts emerge from the same generative process.
Starting the Discussion with Realization, Not Fixed Skills
Two competing frameworks seek to define LLMs according to Floridi. One seeks to define what LLMs are. A relevant question might be "Are LLMs human?" If LLMs are intelligent epistemic agents in the same category as humans, then Stephen Hawkings and Geoffrey Hinton are right. We have an existential problem.
Floridi calls this position the ARI thesis (Artificial Realisability of Intelligence). We may have or are on the brink of creating non-biological intelligence. From this position flows the implication that we should expand our conception of intelligence to include artificial forms of it. The objective of further research should be to better construct AI entities with human intelligence. The alternative position, the position Floridi occupies so compellingly, is called the MRA thesis (Multiple Realisability of Agency) We should expand our conception of agency to include multiple forms, including artificial agency that lacks cognition, intelligence, intention, or mental states.
Think of LLMs as categorically more like hammers than human beings. What type of questions are relevant? Floridi might lead us to think about questions like "Are hammers good carpenters?" Are LLMs award-winning journalists? We could quibble over different types of hammers, we could ethically debate whether trees should be sacrificed to make hammer handles, but obviously, the question is absurd. Alternatively, we might ask "Can hammering be done by different kinds of agents?" Can high-quality journalism be done by an LLM? A quick trip to the 19th century tells us we aren't the first to have asked this question about who owns the rights to the hammer gun. Who owns the rightful patent of the hammer gun could fuel a doctoral dissertation.
Realization of agency from this Multiple Realization of Agency perspective is a phenomenological concern—focusing on what agency means rather than what intelligence is. Arguing for this thesis doesn't imply the alternative thesis is uninteresting. Scientists can continue to work on the intelligence question. But teachers are not expected to answer what David Chalmers called "the hard problem" of consciousness. Easy problems like explaining cognitive functions, e.g., how to command attention, how to develop organized memory, how to represent spatial constructs, etc., are mere entertainments alongside the hard problem: Explaining why and how subjective, first-person conscious experience is realized at all.
The Multiple Realizabilities of Reading Comprehension
Different Textual Substrates, Different Comprehension Processes
Floridi explains that agency can be multiply realized through different domains of substrates—rivers through hydraulic processes, thermostats through programmed responses, LLMs through computational pattern recognition—and so reading comprehension itself can be multiply realized depending on the substrate of the text being comprehended. By substrate, I mean the underlying generative source that produces the text: human consciousness with its intentionality, lived experience, and meaning-making processes, versus artificial computation based on human-generated prompts.
When students read human-authored texts, comprehension realizes through what we might call “substrate intent recognition." Students approach these texts knowing they were written from conscious rendering of thought into language, biographical experiences, and purposeful human communication. This knowledge activates comprehension processes that monitor meaning for authorial intention, cultural context, and the temporal integration that Floridi identifies with human agency. Students interpret not just semantic content but the human thought that generated it, engaging in what might be called "consciousness-to-consciousness" meaning-making—mind reading using linguistic methods.
Conversely, when students read LLM-generated texts, comprehension necessarily realizes differently. The substrate is no longer human consciousness but Floridi's "data-driven adaptability" through calculations that select words based on the probability that the next word is the right word. Students cannot seek authorial intention because none exists—only algorithmic optimization. Rather than seeking deeper human meaning behind textual complexity, students must approach these texts as dataprints where ambiguities and hedges represent computational limitations rather than intentional meaning. LLMs are designed to patch over poor-probability output with less precise terms and hedges.
This distinction becomes clear when examining LLMs against human agency across Floridi's three observables of agency: interactivity, autonomy, and adaptability. While humans demonstrate what Floridi calls "epistemic agency" where "participants recognize both a personal and a collective responsibility for success" and engage in "rational deliberation" and "moral reasoning," LLMs generate responses without intentions or moral responsibility for their outputs. Epistemic agency offers credibility on the part of the author as texts are realized in mental models, which can promote readers’ remodeling prior knowledge. Human interactivity encompasses what Floridi describes as "multifaceted social and informational interactions" that extend far beyond textual exchange, the sole interaction accessible by LLMs. Humans engage with their environment through embodied experience—they can discuss a reading with another human, emotional resonance, and cultural participation that shapes meaning-making in ways LLMs cannot access.
Most critically, human adaptability operates through conscious learning, creative problem-solving, and meaning-making that LLMs cannot replicate. Floridi notes that humans possess "learning from experience, causal reasoning, complex decision-making capabilities" alongside "future scenario simulation" and "long-term goal setting." Human adaptability involves what he describes as "temporal integration" that allows humans to "learn from history, develop complex personal identities and maintain biographical coherence." LLMs adapt within their training parameters, but they lack conscious reflection, intentional learning, and durable meaning-construction. Their adaptability remains syntactic rather than semantic, processing patterns regardless of the significance of the patterns.
Pedagogical Implications of Substrate-Dependent Comprehension
This distinction has practical implications for literacy instruction. Traditional reading comprehension strategies—seeking authorial purpose, making biographical connections, interpreting symbolic meaning—remain appropriate for human-authored texts but may mislead students when applied to artificial texts. Students need new literacy frameworks that help them recognize textual substrates and adapt their comprehension strategies accordingly.
The realizability framework suggests that reading comprehension itself exhibits the in interactivity, adaptability, and autonomy that Floridi identifies as a key observable of agency. Agentic human readers could be taught to enact substrate-sensitive comprehension, modifying their interpretive processes based on recognition of the text's generative mechanism. Literacy pedagogy will need to begin to research questions in addition to human approaches toward what we might call "substrate-aware reading agency."
Looking Forward: In Medias Res Analysis
My fourth essay will extend this analysis through a Levels of Abstraction examination of "reading comprehension in the context of immediate LLM interaction"—what candidates for observables can we construct when students engage with AI-generated text in real time, as the text unfolds through their queries and the system's responses? This in medias res analysis will explore how comprehension realizes differently when students are not just reading AI text but actively participating in its generation, creating new forms of human-AI collaborative meaning-making that require entirely novel pedagogical frameworks. I'm motivated to work hard on this task by the possibility I may come up dry and appreciate any thoughts or feedback you can give me. I'm not sure this sort of analysis can be carried out in the absence of empirical data. But to unearth thoughts and questions, it could be worth a try.