# Distributed reasoning account Reasoning can be understood as a property of the Human-AI transcript or interaction, not the machine itself. This allows for substantive claims about LLM-produced philosophy without requiring claims about inner mental states. ## The Problem The claim "if it survives brutal internal verification loops, reasoning has occurred" feels too strong. It seems to settle the metaphysics of reasoning by fiat. ## The Solution: Move Reasoning to the Process We don't have to say the *LLM* is a "Reasoner" (which implies intent and commitment). We can say that the **Human-AI Interaction** generates **Reasoning Events**. ## Three Formulations (Descending Boldness) ### 1. Artefact-level claim (safest) > If an LLM produces a novel argument that survives standard philosophical scrutiny, then a good piece of philosophy has been produced—regardless of the psychology (or lack thereof) behind it. ### 2. Process-level claim (moderate) > In such cases, the interaction realises a reasoning process (distributed across human constraint-setting and model generation). ### 3. As-if claim (useful bridge to Floridi) > In such cases, the system exhibits reasoning-grade performance even if its core mechanism remains stochastic. ## Why This Works All three formulations respect the "argument not author" instinct of analytic philosophy: we evaluate the logical object, not the inner phenomenology of the producer. This approach: - Doesn't require claiming the model literally reasons like humans - Accepts [[Floridi's critique of LLM abduction|Floridi's]] stochastic-core diagnosis - Still allows substantive claims about the quality of outputs The "Commitment" is supplied by the **Human-AI System**: the Human provides the "Aim" (the "Caring"), the AI provides the "Computational Search" through the [[Space of Reasons]]. Source: [[Web Clippings/chat with ChatGPT about Floridi, LLM reasoning, Move 37|ChatGPT conversation on Floridi et al.]]