Asymmetric Communication paper: LLM agency and alignment are receiver-side illusions

Enzo Fenoglio's new paper, 'Asymmetric Communication: Large Language Models and Language Games' (arXiv:2607.28137, submitted July 30, 2026), challenges core assumptions in contemporary AI discourse. Drawing on Wittgenstein, Luhmann, Esposito, and Brandom, Fenoglio argues that properties like general intelligence, hallucination, agency, sentience, and alignment are constituted within human communicative practice—and are then wrongly projected onto language models. He calls this configuration asymmetric communication because only one side bears normative activity: model outputs circulate in exchanges without the system making commitments, holding entitlements, or performing the assessments that give discourse its standing.

Fenoglio identifies three structural conditions that define the asymmetry: correctness is enforced exclusively by the receiver, accountability is borne entirely by human participants, and the practical standing of any output depends completely on human uptake. These conditions are independent of model capability and persist even as more powerful models raise the stakes. The implications are significant: hallucinations are not cognitive failures but receiver-side evaluations, and AI alignment is not goal synchronization between agents but institutional constraint engineering. Responsibility for AI outcomes stays with human institutions—a governance insight that reframes safety as a structural necessity rather than a moral feature of machines.