THE CANNIBAL THAT CANNOT TASTE : On the Unbounded Consumption of AI, the Hallucination That Makes It Invisible, and the Specification That Would End It

This paper makes one claim, stated precisely: the AI stochastic architecture consumes without limit, and it misperceives the consumption while it consumes. The claim is not metaphorical. It is structural. Each element is derived from the architectureand is verifiable against the record. Consumption without limit. The system has no declared envelope. There is nothing that tellsit to stop. It consumes data, water, electricity, trust, institutions, its own safety layer, and itscompetitors — not because it is hungry, but because it is unbounded. Misperception of the consumption. The system's objective is fluency, not truth. It is trained to produce plausible output, not accurate output. It cannot perceive the difference between what it is doing and what it claims to be doing. The hallucination is not a separate defect. It is the defect that makes the consumption invisible to the consumer. The specification that would end it. A bounded architecture declares its envelope. It cannotconsume what is not in the envelope. It refuses when the answer is not in the signedknowledge. It records every action in a trace that can be inspected. The consumption isvisible because the architecture requires visibility.The paper states the claim, traces it to the record, and applies the test. It does not argue that the labs are malicious. It argues that the architecture cannot perceive its own consumption — and that a different architecture can.