When a doctor, judge, or engineer must decide whether to trust an AI model's output, they inevitably ask what the model actually "understands". Purely mathematical or statistical descriptions struggle to separate trustworthy from untrustworthy results without re‑introducing the question of AI understanding. The usual framing assumes a monophonic paradigm: a cognitive system's understanding is localized to a single mechanism that underlies all related capacities. Drawing on extensive mechanistic evidence, we show that large language models (LLMs) are fundamentally polyphonic: outputs arise from coalitions of parallel mechanisms with uneven reliability, which may complement, duplicate, or drown each other; several coalitions can solve a task, and none is indispensable. Polyphony complicates attributions of understanding and makes monophonic inference hazardous. In response we propose a polyphonic‑compatible conception of understanding, centered on sound circuitry that is reliably recruited and controls the output. Attributing understanding thus becomes a tractable claim about internal organization and can guide trust in AI.
Review