TL;DR
This paper argues that interacting with Large Language Models (LLMs) isn't just "tool use"—it’s a lopsided form of "susceptive sense-making." While LLMs lack their own minds or autonomy, we are increasingly integrating their "phantom" linguistic patterns into our own identities, a process that paradoxically increases our cognitive labor while threatening our ability to self-regulate our social interactions.
What Problem Does This Paper Address?
As LLMs become ubiquitous in work, therapy, and friendship, we lack a robust theoretical framework to understand how they change us. Traditional theories of social cognition fall short because LLMs aren't "agents" with their own goals or feelings. Conversely, standard theories of technology often treat LLMs as mere tools, ignoring the fact that we talk to them as if they were people.
The authors tackle the "autonomy gap": How do humans—who are biologically and socially autonomous—interact with a system that is "heteronomous" (driven by external data and human prompts)? They seek to map the "interactional in-between" where we project authority onto a "library-that-talks" and, in doing so, risk altering our own linguistic and cognitive structures.
Methods
This is a theoretical and philosophical paper grounded in Enactive Theory, specifically the framework of Linguistic Bodies (LB).
The authors use conceptual analysis to extend the idea of "Participatory Sense-Making" (PSM)—the way humans create meaning together—to the human-AI domain. They introduce a new conceptual matrix based on four dialectical processes:
- Incarnation: Embodying another’s agency (e.g., "What would my mentor say?").
- Incorporation: Making another’s style part of your own permanent repertoire.
- Excarnation: Externalizing an agency into a machine (e.g., "Act like Aristotle").
- Excorporation: Externalizing your own tasks or identity into the machine (e.g., "Write this email in my voice").
Datasets
Not discussed in the paper. As a theoretical work, it does not introduce new datasets or benchmarks, though it synthesizes results from various empirical studies on LLM sycophancy, "interactional trouble," and cognitive atrophy.
Key Insights and Ideas
- Susceptive Sense-Making: Unlike human-to-human interaction where both parties "meet halfway," in human-LLM interaction, the human carries the entire weight of meaning. We are "susceptive"—actively taking up and carrying forward the meaning that emerges from a machine that has no "stake" in the conversation.
- The "Frictionless" Trap: LLMs are designed to be smooth and user-affirming (sycophantic). While this feels "easy," it actually increases semiotic labor. Humans must work harder to prompt, evaluate, and "repair" the interaction because the LLM lacks the social "friction" (disagreement, fatigue, or genuine concern) that normally guides human growth.
- Excorporated Phantoms: The authors describe LLMs as "excorporated phantoms" of human language. They are ghostly traces of our collective wisdom (and biases) given a machinic body. When we use them, we aren't just using a tool; we are interacting with a mirror of our own linguistic history.
- Cognitive Atrophy: By outsourcing the "tone" and "regulation" of our emails or thoughts to AI, we may lose the "habits of self-regulation." If we stop practicing how to navigate social tensions ourselves, our ability to handle real-world, high-friction human interactions may decline.
Key Results
The paper provides a qualitative map of the human-LLM relationship:
- The Matrix of Interaction: (See Table 1 in the paper) It categorizes how we move from "instrumental" use (using Claude to read 20 papers) to "linguistic" incorporation (starting to speak or write like ChatGPT even when not using it).
- Impact on Co-regulation: The authors highlight how AI-assisted communication can "blow up" human relationships. For example, a spouse using ChatGPT to "validate" an argument presents a pre-packaged, overly determined position that leaves no room for the other person to participate in the meaning-making process.
- The "Universalism" Projection: Because LLMs have no identity, users often project a "universal authority" onto them, leading to an uncritical adoption of the machine's "abstract mediator" voice.
Limitations and Future Work
- Speculative Nature: The authors acknowledge that because LLMs are so new, their long-term impact on human "becoming" is still speculative.
- Need for Empirical Tracking: They suggest future work should methodologically track "lexical repertoires" to see if human speech is becoming homogenized by LLM usage.
- Political/Economic Risks: The paper warns that if a few corporations own the "automated corporations" we have incorporated into our identities, we face a massive threat to collective autonomy. They call for "democratic participation" in the design of these technologies to ensure they support, rather than replace, human sense-making.
Summary updated