AI as a High-Bandwidth Unreliable Source: Why It Is Neither Tool Nor Agent

The governance and philosophy of generative AI remain trapped between two inherited categories: the tool and the agent. Each frame captures something real; each fails on the properties that matter most once such systems are deployed into consequential work. Tools have a specifiable output space, are passive, and fail mechanically; large language models (LLMs) generate over an unspecifiable space, carry learned regularities that come apart from use, and fail by gradual drift. Agents can be held accountable, carry persistent intentions, and understand; LLMs satisfy none of these in the sense the agent frame requires. Existing correctives — "stochastic parrots," "cultural technologies" — rightly resist the agent reading, but each pulls the system back toward the tool pole rather than naming a category of its own. This paper argues that generative LLMs occupy a third category, which I call a High-Bandwidth Unreliable Source (HBUS): a system that produces large volumes of cross-domain content cheaply (high-bandwidth) while generating outputs that are lucky rather than apt (unreliable-source). The central claim is that these two features are not independent traits that happen to coincide but two faces of one training mechanism, read under two different demands — a thesis I defend on mechanical, not merely rhetorical, grounds. I derive three consequences of the HBUS framing — epistemic, normative, and for the division of labour — defend the proposal against four objections, and state explicit falsifiability conditions under which the category should be abandoned. The aim is conceptual engineering: naming AI correctly is prior to governing it well, and neither of the two dominant names is correct.