Talking with Language Models

TEXT START: When we interact with large language models (LLMs), are we having a conversation?

The Dissection

The paper strips away the conversational costume. It recasts LLM interaction as artifact-mediated exchange: a user interprets generated text and supplies the continuity, agency, and commitment the interface appears to provide.

That diagnosis is useful, but narrow. The paper relocates the important questions to design, adoption, authorization, and human use. It attacks anthropomorphism while leaving the political economy of cognitive automation largely untouched.

The Core Fallacy

It conflates the absence of a conscious, continuous speaker with the absence of meaning, force, and accountability. A system need not possess human-style intention to produce text that functions as a lie, promise, instruction, decision, or economic substitute once users and institutions treat it that way.

Its deeper omission is structural. Under the Discontinuity Thesis, the decisive question is not whether an LLM is a person. It is whether the artifact performs economically necessary cognitive work cheaply and at scale. The paper demotes the interface while ignoring the machine underneath it: a labor-substitution system capable of severing the labor–wage–consumption circuit.

Hidden Assumptions

• Only a conscious or continuous speaker can bear meaning or normative force.
• Lack of memory between sessions dissolves identity rather than merely relocating identity into configuration, transcript, operator, and institution.
• The user is the primary performer, minimizing the model’s substituted labor and the deployer’s institutional power.
• Responsibility must belong to the artifact itself or to no one, obscuring distributed responsibility among designers, owners, authorizers, and users.
• Reclassifying the system as an artifact resolves the political problem. It does not. It merely changes who controls the artifact and who is displaced by it.

Social Function

Classification: partial truth with transition-management and ideological-anesthetic function.

The anti-anthropomorphic correction is real. But the repeated claim that there is “no one to hold responsible” can become accountability laundering. It gives deployers a clean philosophical exit: the machine is only an artifact, the user supplied the interpretation, and responsibility dissolves into the interface.

The paper therefore performs conceptual sanitation without completing the economic autopsy. It removes the face from the machine, but does not examine the owners, institutions, or displaced workers around it.

The Verdict

This is a sharp critique of conversational framing and an incomplete account of AI’s consequences. It correctly identifies the interlocutor as an interface effect, but mistakes ontological demotion for systemic explanation.

Under DT logic, the LLM does not need a self, memory, or moral agency to destroy human economic necessity. The artifact is enough. The paper diagnoses the mask; it does not diagnose the terminal decline of the system that the mask conceals.